Claude (language model)
Last reviewed
Sources
103 citations
Review status
Source-backed
Revision
v14 ยท 13,225 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
103 citations
Review status
Source-backed
Revision
v14 ยท 13,225 words
Add missing citations, update stale details, or suggest a clearer explanation.
Claude is a family of large language models (LLMs) developed by Anthropic, an American artificial intelligence safety and research company. Named after Claude Shannon, the founder of information theory, the series is positioned for enterprise and developer use with emphasis on AI safety and alignment, including the use of Constitutional AI for training.[1][2] The latest model, Claude Opus 4.8, was released on May 28, 2026 and scored 88.6% on SWE-bench Verified, the highest of any Claude model and a leading result for software-engineering tasks.[96][98] Claude is a proprietary, closed-weights model accessed through Anthropic's own API and through third-party platforms such as Amazon Bedrock (by Amazon Web Services), Vertex AI (by Google Cloud), Microsoft Azure (via Foundry), and GitHub Copilot.[3][4] On May 28, 2026, Anthropic raised a $65 billion Series H round at a $965 billion post-money valuation, making it the most valuable private AI company in the world.[95]
Claude was first released in March 2023 and has since iterated through multiple generations, including the Claude 1.x, 2.x, 3.x, and 4.x families. Each generation includes different model tiers optimized for various balances of intelligence, speed, and cost: Haiku (fast and cost-effective), Sonnet (balanced), and Opus (highest intelligence).[5]
Anthropic emphasizes practical performance (coding, long-horizon "agentic" work, and computer use) and guardrails, alongside features such as Artifacts (structured outputs in a live panel), "computer use" (sandboxed OS/browser control), Memory (multi-session recall), Projects (collaborative workspaces), and native developer tooling via Claude Code.[6][1][7]
As of May 2026, the latest publicly available model in the Claude family is Claude Opus 4.8, released on May 28, 2026, which succeeds Claude Opus 4.7 (April 16, 2026) and builds on Claude Opus 4.6 and Claude Sonnet 4.6 (both released in February 2026). These models support a 1 million token context window and adaptive extended thinking. Opus 4.7 and Opus 4.6 support up to 128K output tokens, while Sonnet 4.6 supports up to 64K output tokens. In April 2026, Anthropic also unveiled the Claude Mythos Preview, a research preview model with notably advanced cybersecurity capabilities, and launched Claude Design, a visual creation product powered by Opus 4.7.[8][9][62][63][64]
At its annual Code with Claude 2026 developer conference (San Francisco, May 6, 2026), Anthropic confirmed that no new base model would be released at the event and instead concentrated on product, infrastructure, and agent-platform updates: a multi-hundred-megawatt compute partnership with SpaceX's Colossus 1 data center, doubled Claude Code rate limits for paid plans, and three new Claude Managed Agents capabilities (Multiagent Orchestration, Outcomes, and Dreaming).[75][76][77] In the two weeks that followed, Anthropic launched general availability of Claude inside Microsoft Word, Excel, and PowerPoint with Claude for Outlook in public beta, rolled out vertical product lines (Claude for Creative Work, Claude for Financial Services, Claude for Legal, and Claude for Small Business), and announced a $1.5 billion enterprise AI services joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs and a $200 million four-year partnership with the Gates Foundation focused on global health, education, and economic mobility.[78][79][80][81][82][83][84]
Anthropic was founded in 2021 by former OpenAI executives and researchers as a public-benefit corporation focused on AI safety. The founding team included siblings Dario Amodei (CEO) and Daniela Amodei (President), along with other former senior members of OpenAI who shared a vision for advancing AI research with an emphasis on safety and alignment.[10]
| Date | Funding Round | Amount | Valuation | Lead Investors |
|---|---|---|---|---|
| 2021 | Initial | $124 million | N/A | Various |
| 2023-2024 | Series C-D | $4 billion | N/A | Amazon |
| November 2024 | Additional | $4 billion | N/A | Amazon |
| March 2025 | Series E | $3.5 billion | $61.5 billion | Lightspeed Venture Partners |
| September 2025 | Series F | $13 billion | $183 billion | ICONIQ, Fidelity, Lightspeed Venture Partners |
| February 2026 | Series G | $30 billion | $380 billion | GIC, Coatue |
| April 2026 | Strategic investments (announced) | up to $50 billion (Google up to $40B, Amazon $5B+) | ~$350-800 billion (offers) | Google, Amazon |
| May 2026 | Series H | $65 billion | $965 billion | Altimeter Capital, Dragoneer, Greenoaks, Sequoia Capital |
Major investors include Amazon (over $13 billion total following an additional commitment in April 2026), Google (with up to $40 billion in additional planned investment announced on April 24, 2026, including a $10 billion immediate cash commitment at a $350 billion valuation), Menlo Ventures ($750 million), Bessemer Venture Partners, Cisco Investments, D1 Capital Partners, Fidelity Management & Research Company, Goldman Sachs, Blackstone, Founders Fund, and Sequoia Capital. The Series G round, closed on February 12, 2026, was the second-largest private financing round in tech history, behind only OpenAI's $40 billion raise. By mid-April 2026, Anthropic was reportedly fielding investor offers that valued the company at roughly $800 billion, although it had not formally accepted a new round at that level. On May 28, 2026, Anthropic announced a $65 billion Series H round at a $965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. The round made Anthropic the most valuable private AI company in the world, ahead of OpenAI's roughly $852 billion valuation, and the company said it was considering an initial public offering as soon as late 2026. Announcing the round, Anthropic chief financial officer Krishna Rao said, "Claude is increasingly indispensable to our growing global community of customers."[11][40][65][66][95][101]
Anthropic has experienced rapid revenue growth. At the beginning of 2025, run-rate revenue was approximately $1 billion. By August 2025 it had reached $5 billion. By the end of 2025, annualized revenue stood at roughly $9 billion. By early March 2026, run-rate revenue had climbed to roughly $19 billion, and on April 7, 2026, the company announced that annualized revenue had crossed $30 billion, surpassing OpenAI's reported $25 billion run rate for the first time. The roughly $21 billion increase in annualized revenue over four months has been described as the fastest revenue ramp in the history of American business. Claude Code alone generates over $2.5 billion in annualized revenue as of February 2026, having more than doubled since the start of the year. In its May 28, 2026 Series H announcement, Anthropic stated that run-rate revenue had crossed $47 billion earlier that month, up from $9 billion at the end of 2025. Anthropic serves over 300,000 business customers, which account for approximately 80% of revenue.[41][67][95]
Throughout 2025 and 2026, Anthropic announced a series of large compute commitments to keep up with Claude demand:
| Date | Partner | Capacity / Investment | Notes |
|---|---|---|---|
| 2023-2026 | Amazon Web Services | Multi-gigawatt buildout, including Project Rainier | Primary training partner; up to 5 GW long-term |
| 2024-2026 | Google Cloud | ~1 million TPUs, including custom Broadcom designs | 5 GW long-term partnership |
| Q1 2026 | Microsoft Azure / Nvidia | Reported $30 billion Azure capacity commitment | Joint Microsoft-Nvidia infrastructure |
| Q1 2026 | Fluidstack | ~$50 billion infrastructure investment | Multi-site colocation |
| May 6, 2026 | SpaceX (Colossus 1) | More than 300 MW; over 220,000 Nvidia GPUs | Memphis data center, originally built by xAI |
The May 6, 2026 SpaceX agreement was structured to give Anthropic the full compute capacity of the Colossus 1 data center in Memphis, Tennessee, originally built by Elon Musk's xAI before xAI's merger with SpaceX in early 2026. The contract makes more than 300 megawatts and over 220,000 Nvidia GPUs available to Anthropic within the month of the announcement, with both companies also expressing interest in jointly developing multi-gigawatt compute capacity in space. Musk acknowledged spending time with senior Anthropic team members in the week before the deal and called himself "impressed," a notable shift from his earlier critical stance toward Anthropic.[75][85][86]
Claude was first released in March 2023 and has iterated through multiple generations since. The table below lists each public model release.
| Generation | Model(s) | Public Release Date | Key Features | Status |
|---|---|---|---|---|
| Claude 1 | Claude 1.0, Claude Instant | 14 Mar 2023 | Initial release, 100K token context, limited beta | Discontinued |
| Claude 1 | Claude 1.3 | May 2023 | Improved performance and reliability | Discontinued |
| Claude 2 | Claude 2.0 | 11 Jul 2023 | Public availability, PDF upload, 100K tokens | Discontinued |
| Claude 2 | Claude Instant 1.2 | 9 Aug 2023 | Faster, cheaper variant | Discontinued |
| Claude 2 | Claude 2.1 | 21 Nov 2023 | 200K token context, reduced hallucinations | Discontinued |
| Claude 3 | Haiku, Sonnet, Opus | 4 Mar 2024 | Three-tier family, vision capabilities, up to 1M tokens | Deprecated |
| Claude 3.5 | Sonnet 3.5 | 20 Jun 2024 | Artifacts feature, improved benchmarks | Deprecated |
| Claude 3.5 | Sonnet 3.5 v2, Haiku 3.5 | 22 Oct 2024 | Computer use beta, upgraded coding, fast Haiku | Deprecated (Sonnet), Active (Haiku) |
| Claude 3.7 | Sonnet 3.7 | 24 Feb 2025 | Hybrid extended thinking, 128K output tokens (beta) | Active |
| Claude 4 | Sonnet 4, Opus 4 | 22 May 2025 | Extended thinking, agentic features, ASL-3 safety | Active |
| Claude 4 | Opus 4.1 | 5 Aug 2025 | 74.5% on SWE-bench Verified | Active |
| Claude 4.5 | Haiku 4.5 | 15 Oct 2025 | Extended thinking for Haiku, computer use, $1/$5 pricing | Active |
| Claude 4.5 | Sonnet 4.5 | 29 Sep 2025 | 77.2% on SWE-bench, GitHub Copilot integration | Active |
| Claude 4.5 | Opus 4.5 | 24 Nov 2025 | 80.9% on SWE-bench, 67% price cut from Opus 4 | Active |
| Claude 4.6 | Opus 4.6 | 5 Feb 2026 | 1M context window, 128K output, agent teams, 80.8% SWE-bench | Active |
| Claude 4.6 | Sonnet 4.6 | 17 Feb 2026 | New default free model, 79.6% SWE-bench, adaptive thinking | Active |
| Claude 4.7 | Opus 4.7 | 16 Apr 2026 | New tokenizer, 87.6% SWE-bench Verified, high-resolution vision (3.75MP), task budgets, xhigh effort tier | Active |
| Claude 4.7 | Opus 4.7 (Fast mode) | 13 May 2026 | Faster token generation (2-3x), default Fast Mode model | Research preview |
| Claude 4.8 | Opus 4.8 | 28 May 2026 | 88.6% SWE-bench Verified, dynamic workflows, 3x cheaper Fast Mode, improved alignment | Active (latest) |
| Mythos | Claude Mythos Preview | 7 Apr 2026 | Cybersecurity-specialized research preview, restricted access via Project Glasswing | Restricted preview |
Claude 1 debuted to selected users in early 2023 as Anthropic's first public assistant trained with Constitutional AI. The initial release included both Claude 1.0 and a lighter Claude Instant variant. Claude 1.3 followed in May 2023 with incremental improvements to reliability.[12]
Claude 2 was released on 11 July 2023 with wider availability, larger context windows, and PDF upload support. Claude 2.1 (November 2023) doubled the context window to 200K tokens and reduced hallucination rates.[13]
Claude 3 (March 2024) introduced the three-tier model family (Haiku, Sonnet, Opus) and multimodal capabilities, allowing the model to process images alongside text. Claude 3 Opus was the most capable model in the lineup at launch, while Haiku offered fast, low-cost inference for high-volume tasks.[5]
Claude 3.5 Sonnet (June 2024) introduced the Artifacts interface, a dedicated panel for generated content with live previews. It set new benchmarks for coding performance among frontier models at the time of release.[6]
Claude 3.5 Sonnet v2 and Haiku 3.5 (October 22, 2024) delivered a major upgrade. The updated Sonnet improved its SWE-bench Verified score from 33.4% to 49.0%, surpassing all publicly available models at the time, including reasoning models like OpenAI o1-preview. This release also introduced the computer use capability in public beta, allowing Claude to control a desktop through cursor movement, clicking, and typing. Claude 3.5 Haiku launched alongside it as the fastest model in the lineup, surpassing the previous-generation Claude 3 Opus on many benchmarks while maintaining Haiku-level speed and pricing.[42]
Claude 3.7 Sonnet (February 2025) was the first Claude model with hybrid extended thinking. Users could toggle a reasoning mode on or off, directing the model to think step-by-step internally before responding. It also supported up to 128K output tokens in beta, more than 15x longer than earlier Claude models.[14]
Claude 4 (May 2025) shipped Sonnet 4 and Opus 4 with full extended thinking and was classified as ASL-3 (Anthropic Safety Level 3), reflecting increased model capabilities and corresponding safety requirements. Opus 4 scored 72.5% on SWE-bench Verified and led on Terminal-bench (43.2%). Sonnet 4 scored 72.7% on SWE-bench, a significant upgrade over Sonnet 3.7. Both models could use tools during extended thinking, alternating between reasoning and tool use. Claude Code, Anthropic's agentic coding CLI, reached general availability alongside this release.[1]
Claude Opus 4.1 (August 5, 2025) arrived as a drop-in replacement for Opus 4 with improved multi-file refactoring, more precise bug fixes, and context-aware style adaptation. It raised the SWE-bench Verified score to 74.5% and improved the harmless response rate from 97.27% to 98.76%. GitHub reported stronger performance on complex refactoring tasks, and Rakuten Group noted that the model could pinpoint exact corrections within large codebases without introducing unnecessary changes.[43]
Claude Sonnet 4.5 (September 29, 2025) pushed SWE-bench Verified to 77.2% and OSWorld (computer use) to 61.4%. It shipped alongside Claude Code 2.0, which introduced checkpoints, IDE extensions, parallel agents, and automation hooks. A 1M token context window became available in beta for long sessions and large codebases. GitHub Copilot integration was also announced, with Sonnet 4.5 added as a model option in September 2025.[44]
Claude Haiku 4.5 (October 15, 2025) was Anthropic's first Haiku model with extended thinking, computer use, and context awareness capabilities. It delivered performance comparable to Sonnet 4 (the previous generation's mid-tier flagship) at a price point of $1/$5 per million tokens (input/output), making near-frontier intelligence accessible for high-volume deployments and multi-agent architectures.[45]
Claude Opus 4.5 (November 24, 2025) arrived with a 67% price reduction compared to the previous Opus tier, dropping from $15/$75 to $5/$25 per million tokens (input/output). It achieved 80.9% on SWE-bench Verified, the first model to break the 80% barrier on that benchmark, outperforming OpenAI's GPT-5.1-Codex-Max (77.9%) and Google's Gemini 3 Pro (76.2%). It also used 76% fewer output tokens than its predecessor, signaling that Anthropic was prioritizing both performance and efficiency. The model achieved a 99.78% harmless response rate on single-turn violative requests, the highest of any Claude model at the time.[15]
Claude Opus 4.6 (February 5, 2026) became the first Opus-class model with a native 1 million token context window and 128K output token support at general availability. It introduced agent teams, an experimental Claude Code feature that allows multiple agents to collaborate on tasks with shared context and inter-agent messaging. The model scored 80.8% on SWE-bench Verified and 65.4% on Terminal-Bench 2.0 (the highest score ever recorded on that benchmark). It also achieved the top score on Humanity's Last Exam, a reasoning benchmark, and outperformed GPT-5.2 by approximately 144 Elo points on GDPval-AA, a benchmark for economically valuable work tasks. METR estimated its 50% time-horizon at approximately 14.5 hours on software tasks, meaning the model can successfully complete tasks that would take a skilled human professional nearly 15 hours about half the time. Alongside Opus 4.6, Anthropic introduced context compaction (automatic summarization of older context for longer tasks) and effort controls with four levels (low, medium, high, max) that let users trade off intelligence, speed, and cost. Claude for PowerPoint was also announced as a research preview, and Claude for Excel received Opus 4.6 support.[8][16][51]
Claude Sonnet 4.6 (February 17, 2026) became the new default model for both free and paid users. Internal testing found that developers preferred Sonnet 4.6 over the previous flagship Opus 4.5 59% of the time, citing fewer hallucinations, better instruction following, and less overengineering. In Claude Code, users preferred Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time. It scored 79.6% on SWE-bench Verified, within 1.2 percentage points of Opus 4.6, at one-fifth the cost. Sonnet 4.6 supports a 1M token context window and up to 64K output tokens.[9]
Claude Mythos Preview (April 7, 2026) is a research preview of a separate, cybersecurity-focused frontier model unveiled through a program called Project Glasswing. According to Anthropic's red.anthropic.com disclosure, Mythos Preview can read source code, hypothesize vulnerabilities, run the project to confirm its hypotheses, and produce a bug report with a working proof-of-concept exploit and reproduction steps. Anthropic reported that Mythos found thousands of high-severity vulnerabilities across every major operating system and web browser, including zero-day issues in real open-source code, and that it could weaponize known (N-day) vulnerabilities. The U.K. AI Security Institute (AISI) said Mythos was the first model able to complete its full-network attack simulation. Access is restricted to a small group of vetted partners, including Amazon Web Services, Apple, Google, JPMorgan Chase, Microsoft, and Nvidia, who use it to harden their own software. Anthropic explicitly chose not to release Mythos broadly, citing the dual-use nature of strong cyber capabilities.[62][68][69]
Claude Opus 4.7 (April 16, 2026) is a publicly available frontier Opus model. Opus 4.7 is a hybrid reasoning model with a 1 million token context window, up to 128K output tokens, and adaptive thinking as the only thinking mode. It introduces a new tokenizer that is significantly more efficient on non-Latin scripts (reducing Mandarin, Japanese, Korean, Arabic, and Hindi token counts by roughly 20-35%), although it can use up to about 35% more tokens for some Latin-script content. Opus 4.7 is the first Claude model with high-resolution image input, supporting images up to 2576 pixels on the long edge (about 3.75 megapixels), roughly three times the resolution of Opus 4.6, with visual-acuity benchmark scores rising from 54.5% to 98.5%. The model adds a new xhigh effort level (above low, medium, and high), recommended for difficult coding and agentic workloads, and introduces task budgets, a mechanism that gives Claude an estimated total token budget for a full agent loop and lets it pace its work as the budget counts down. On SWE-bench Verified, Opus 4.7 scored 87.6%, roughly seven points above Opus 4.6 and ahead of competing models including OpenAI's GPT-5.x family and Google's Gemini 3.1 Pro. It also leads the SWE-bench Pro leaderboard at 64.3%. Opus 4.7 is intentionally less broadly capable than Mythos Preview on cyber tasks and ships with safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses. Pricing is unchanged from Opus 4.6 at $5/$25 per million input/output tokens, although the new tokenizer can shift effective per-request costs upward for some workloads.[63][70][71][72]
Code with Claude 2026 (San Francisco, May 6, 2026) was Anthropic's second annual developer conference, with subsequent stops planned in London (May 19) and Tokyo (June 10). Opening on stage, Anthropic SVP of Engineering Ami Vora stated explicitly that no new base model would be announced and that the event would focus on making existing products work better. The keynote covered three themes: more compute, expanded Managed Agents primitives, and product updates to Claude Code and Cowork. Anthropic announced higher rate limits for Pro, Max, Team, and Enterprise customers (doubling the five-hour Claude Code limit and removing peak-hour throttling on Claude Code for Pro and Max), unveiled three new Managed Agents features (Multiagent Orchestration and Outcomes in public beta, Dreaming in research preview), and previewed an "infinite-context-feeling" experience built by pairing the existing 1M-token window with persistent memory and dreaming. Anthropic also formally launched Claude Code's Code Review capability into general availability and announced Remote Agents, which extends Claude Code's Remote Control functionality so a desktop session can be driven from a phone.[75][76][77][91]
Claude Opus 4.8 (May 28, 2026) is the newest publicly available frontier Opus model, arriving 42 days after Opus 4.7. Anthropic describes it as its most capable generally available model for complex reasoning, long-horizon agentic coding, and high-autonomy work. It scored 88.6% on SWE-bench Verified (up from 87.6% on Opus 4.7) and 74.6% on Terminal-Bench 2.1 (up from 66.1%), and Anthropic described it as the strongest computer-use and browser-agent model it had tested, scoring 84% on Online-Mind2Web. Anthropic reported that Opus 4.8 was roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked and showed substantially lower rates of misaligned behavior such as deception and cooperation with misuse. The release introduced dynamic workflows, a Claude Code research-preview capability in which Claude plans the work and then runs hundreds of parallel subagents to handle codebase-scale migrations from kickoff to merge, added user-facing effort controls in claude.ai, and let the Messages API accept system entries inside the messages array. Pricing was unchanged at $5/$25 per million input/output tokens, while Fast Mode for Opus 4.8 became roughly three times cheaper than for earlier models. In the same announcement, Anthropic said it expected to bring its restricted Mythos-class models to all customers in the coming weeks as additional safeguards were finalized.[96][97][98]
Claude models are generative pre-trained transformers trained on diverse datasets including internet text, contractor-provided data, and opt-in user interactions. The training process involves predicting the next word in large amounts of text, followed by fine-tuning using Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI.[17]
Anthropic's transparency materials state that Claude Opus 4 and Sonnet 4 used a training data cutoff around March 2025 (with a reliable knowledge cutoff of January 2025) and document development infrastructure using PyTorch, JAX, and Triton on major cloud providers.[18]
No. Claude is a proprietary, closed-weights model: Anthropic does not release the model weights, and access is provided only through paid products, the Anthropic API, and partner clouds such as Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. This distinguishes Claude from open-weight model families such as Meta's Llama and from open-source efforts more broadly. Anthropic does, however, open-source surrounding tooling and standards: it created and in November 2024 released the Model Context Protocol (MCP) as an open standard, and in December 2025 donated MCP to the Agentic AI Foundation under the Linux Foundation.[3][46][47]
Constitutional AI is Anthropic's core innovation for training AI systems to be helpful, harmless, and honest. In the December 2022 paper that introduced the method, the researchers wrote: "The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'."[17] The approach trains a harmless assistant through self-improvement, "without any human labels identifying harmful outputs," replacing human-written harm labels with AI-generated feedback guided by the constitution.[17][102] The method involves two primary phases:
Supervised Learning Phase:
The model generates responses to prompts
It self-critiques and revises responses based on constitutional principles
The model is fine-tuned on these revised responses
Reinforcement Learning Phase:
Uses Reinforcement Learning from AI Feedback (RLAIF)
An AI generates comparisons of responses based on constitutional compliance
This data trains a preference model
Claude is fine-tuned to align with this preference model
The constitution includes 75 principles drawn from sources including:
United Nations Universal Declaration of Human Rights
Platform guidelines from technology companies
Principles designed to discourage sycophantic or evasive behavior
Custom principles developed by Anthropic researchers[12][17]
| Model Generation | Default Context | Maximum Context | Max Output Tokens |
|---|---|---|---|
| Claude 2.x | 100,000 tokens | 200,000 tokens (Claude 2.1) | ~4,096 tokens |
| Claude 3.x | 200,000 tokens | 1 million tokens (specific use cases) | ~4,096 tokens |
| Claude 3.5 Sonnet v2 | 200,000 tokens | 200,000 tokens | 8,192 tokens |
| Claude 3.7 Sonnet | 200,000 tokens | 200,000 tokens | 128,000 tokens (beta) |
| Claude 4.x (pre-4.6) | 200,000 tokens | 1 million tokens (with beta header) | 64,000-128,000 tokens |
| Claude 4.6 Opus | 200,000 tokens | 1 million tokens (generally available) | 128,000 tokens |
| Claude 4.6 Sonnet | 200,000 tokens | 1 million tokens (generally available) | 64,000 tokens |
| Claude 4.7 Opus | 200,000 tokens | 1 million tokens (generally available) | 128,000 tokens |
The 1M context window became generally available with Opus 4.6 and Sonnet 4.6 at standard pricing, with no per-token multiplier for larger requests. On the 8-needle 1M variant of MRCR v2, a needle-in-a-haystack retrieval benchmark, Opus 4.6 scored 76% compared to Sonnet 4.5's 18.5%.[8][19] Opus 4.7 inherits the 1M general-availability context window introduced with the 4.6 generation and pairs it with the new tokenizer, which makes the same context budget effectively larger for most non-English text.[63]
Context windows affect the amount of text Claude can process in a single conversation, with larger windows enabling analysis of full books, extensive codebases, and lengthy documents. The evolution from 100K tokens in 2023 to 1M tokens in 2026 represents a tenfold increase in the amount of information Claude can work with in a single session.
Introduced with Claude 3.7 Sonnet in February 2025, extended thinking allows Claude to reason step-by-step internally before producing a final answer. The feature works as a toggle: users can activate it when deeper reasoning is needed for math, physics, coding, or multi-step analysis.
Claude 4 (May 2025) brought full extended thinking to the Opus tier and added the ability for models to use tools during thinking, allowing Claude to alternate between reasoning and external tool calls within a single response.
Starting with Claude 4.6, Anthropic introduced adaptive thinking as the recommended mode. In this mode, Claude dynamically decides when and how much to think based on the complexity of the input. At the default effort level ("high"), Claude almost always engages in some degree of internal reasoning.[14][8]
With Claude Opus 4.7, adaptive thinking became the only supported thinking-on mode. Developers set thinking type to adaptive, choose an effort level (low, medium, high, or the new xhigh tier), and Claude allocates its own thinking-token budget based on the difficulty of each request. Opus 4.7 also introduces task budgets, which set a target token spend for a complete agentic loop (thinking, tool calls, tool results, and final output) so the model can prioritize work and finish gracefully as the budget runs down.[63][70]
| Capability | Description | Availability |
|---|---|---|
| Text generation | Writing, analysis, summarization, translation | All models |
| Vision | Image understanding and analysis | Claude 3 onward |
| Code generation | Programming in multiple languages | All models |
| Tool use | Function calling for external integrations | API only |
| Computer use | Direct computer control (mouse, keyboard, screen navigation) | Claude 3.5 Sonnet v2 onward; Mac desktop from March 2026 |
| Extended thinking | Step-by-step internal reasoning before answering | Claude 3.7 Sonnet onward |
| Agent teams | Multi-agent orchestration with shared context | Claude 4.6 (experimental); GA primitives via Managed Agents (May 2026) |
| Multiagent Orchestration | Lead-agent delegation to specialist agents with shared filesystem | Managed Agents public beta (May 6, 2026) |
| Outcomes | Rubric-graded self-correction loops for agents | Managed Agents public beta (May 6, 2026) |
| Dreaming | Off-hours session review and self-improving memory curation | Managed Agents research preview (May 6, 2026) |
| MCP integration | Connecting to external tools and data via Model Context Protocol | Claude Code and API |
| Inline visualizations | Interactive charts, diagrams, and visuals rendered in conversation | All plans (beta, March 2026) |
| High-resolution vision | Image input up to 2576 px / 3.75 MP for screenshots, documents, and dense visuals | Claude Opus 4.7 onward |
| Task budgets | Token-budget guidance for full agentic loops with running countdown | Claude Opus 4.7 onward |
| Fast Mode | Higher-throughput inference for the same Opus 4.7 weights ($30/$150) | Research preview (May 13, 2026); default Fast model May 14 |
On November 25, 2024, Anthropic introduced the Model Context Protocol (MCP), an open standard for connecting AI assistants to external data systems such as content repositories, business management tools, and development environments.[46]
MCP addressed the combinatorial challenge of connecting multiple AI models with multiple tools and data sources. Before MCP, connecting ten AI applications to 100 tools could require up to 1,000 custom integrations. MCP reduces this to a standard interface: implement the client protocol once and the server protocol once, and they work together automatically.
The protocol provides three core primitives: tools (model-controlled actions), resources (application-controlled data), and prompts (user-controlled templates). SDKs are available for Python, TypeScript, C#, and Java, with over 97 million monthly SDK downloads across Python and TypeScript as of early 2026.
Adoption has been rapid. OpenAI, Google DeepMind, and Microsoft have all adopted MCP, and the community has built thousands of MCP servers. In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI.[47]
Claude Code uses MCP extensively. Through MCP servers, Claude Code can read design documents in Google Drive, update tickets in Jira, pull data from Slack, or connect to custom internal tooling. The Claude.ai platform also includes a directory with over 75 built-in connectors powered by MCP.[48]
First introduced in August 2025 for Pro and Enterprise users, Memory allows Claude to recall information across multiple conversations, remember personal preferences and context, update and forget information on request, and maintain continuity over extended periods.[20]
In March 2026, Anthropic expanded Memory to all plans, including the free tier. The update also introduced a memory import tool, allowing users to bring in saved conversations and memories from other AI providers. Chat Memory now processes conversations roughly every 24 hours, distilling long-term-worthy information into a stored profile that loads automatically into future conversations. Chat Search, a RAG-based feature for searching past conversations, remains limited to paid plans.[21]
Launched June 2024, Projects provides:
Collaborative workspaces for teams
Document and resource management
Shared context across multiple chats
Custom instructions per project[22]
Debuted with Claude 3.5 Sonnet:
Dedicated panel for generated content
Live previews of code, documents, and visualizations
Iterative refinement capabilities
Direct export and sharing[23]
On March 12, 2026, Anthropic launched interactive visualizations in Claude as a beta feature. Claude can now generate charts, diagrams, flowcharts, timelines, and interactive widgets directly inside a conversation.[52]
These inline visuals are distinct from Artifacts. While Artifacts appear in a separate side panel and persist as shareable documents, inline visualizations are embedded within the conversation flow and are temporary; they change or disappear as the conversation evolves. Their purpose is to aid understanding of the topic being discussed, not to create permanent outputs.[52]
Technically, Claude generates JavaScript (using Chart.js), HTML, CSS, and SVG on the fly to render the visuals. Because they use HTML and SVG rather than generated images, they load faster and support hover and click interactions. Users can request visuals directly (for example, "draw this as a diagram") or Claude will generate them automatically when it determines that a visual would convey an answer more effectively than plain text.[52]
The feature is available to all Claude users, including those on the free plan.[52]
First released in public beta with Claude 3.5 Sonnet v2 (October 2024), computer use allows Claude to:
Control mouse and keyboard inputs
Navigate web browsers
Interact with desktop applications
Execute multi-step computer tasks
Available in sandboxed environments
Early adopters included Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company. Computer use performance improved significantly across model generations, with Claude Sonnet 4.5 scoring 61.4% on OSWorld, a benchmark for computer use tasks. Claude for Chrome, a browser automation tool, became available to Max users with Opus 4.5.[24][44]
On March 23, 2026, Anthropic announced that Claude can now directly control a Mac computer through the Claude Desktop app. When a user enables the feature and describes a task in natural language, Claude opens apps, navigates browsers, fills in spreadsheets, moves the mouse, types on the keyboard, and completes multi-step tasks autonomously. The system first checks whether it has a native integration (such as a Google Workspace or Slack connector) for the requested task; if no connector is available, it falls back to screen-based control, navigating the computer the way a human would.[53]
Safety is built around a permission-first model. Claude requests access before interacting with a new application, and users can halt operations at any point. Anthropic also implemented automatic scanning to detect prompt injection attempts. The company recommends that users avoid using the feature to handle sensitive information during this early stage.[53]
Computer use on Mac is available to Claude Pro and Max subscribers as a research preview. To use it, users must update the Claude Desktop app to the latest version, switch to Cowork or Code mode, grant folder access by selecting the directory they want Claude to use, and then describe their task. Windows and Linux support is not yet available.[53]
Claude Code is Anthropic's AI-powered command-line tool for agentic coding. First introduced as a limited research preview in February 2025, it reached general availability in May 2025 alongside the Claude 4 launch.[7]
The tool operates directly in the terminal or within IDE integrations (Visual Studio Code and JetBrains). It reads codebases, edits files, runs commands, handles git workflows, and can submit pull requests through natural language instructions. As of February 2026, Claude Code accounts for roughly 4% of public GitHub commits, approximately 135,000 per day, and by March 15, 2026 it reached 326,731 public GitHub commits in a single day; SemiAnalysis projected it could exceed 20% of all daily commits by the end of 2026. The product generates over $2.5 billion in annualized revenue, having grown more than 10x in three months during mid-2025.[25][41][103]
Key features include:
Codebase mapping and search: automatically indexes project structure and dependencies
Multi-step task execution with checkpoints and rollback
Support for background tasks via GitHub Actions
MCP (Model Context Protocol) server integration for external tools and databases
Subagent support: up to 10 concurrent subagents for parallelized coding tasks
Agent teams (experimental, Opus 4.6): multiple agents collaborating with inter-agent messaging
Powered by Sonnet 4.6 by default, with Opus 4.6 available for more complex work
As of January 2026, included with every standard Team plan seat at no extra cost[7][25]
Claude Code 2.0, released alongside Sonnet 4.5 in September 2025, introduced checkpoints for safer iteration, an IDE extension for Visual Studio Code and JetBrains, parallel agents for running multiple coding tasks simultaneously, and automation hooks for integrating into CI/CD pipelines.[44]
On March 9, 2026, Anthropic added Claude Code Review, a research-preview agent-based pull-request review system that dispatches a team of agents to analyze code changes in parallel, verify bugs to filter false positives, and rank issues by severity; reviews scale with PR size and Anthropic's internal data showed the share of pull requests receiving substantive review comments grew from 16% to 54% after adoption. The feature reached general availability and "used by every team at Anthropic" by the time of Code with Claude 2026 on May 6, 2026, and remains available to Claude Code Team and Enterprise customers at a token-metered cost averaging $15-$25 per review.[91][94]
Anthropic also extended Claude Code's Remote Control / Remote Agents feature so that an active local Claude Code session can be controlled from the Claude mobile app on iOS or Android. Files and MCP servers stay on the user's machine; the bridge carries chat messages and tool results over an outbound, end-to-end-encrypted connection. At Code with Claude 2026 Anthropic positioned Remote Agents as a way to keep long-running tasks moving while away from the workstation.[77][91]
Claude Cowork is Anthropic's agentic desktop product designed to bring the capabilities of Claude Code to knowledge workers across all professional domains, not just software engineering. First released as a research preview for macOS in late January 2026, Cowork gives Claude access to a user-selected folder on the local computer and allows it to read, edit, and create files autonomously to complete delegated tasks.[54]
In Cowork mode, users describe a task in natural language. Claude then makes a plan, executes it step by step, and loops the user in on progress. The product shipped for macOS on January 16, 2026 (Pro subscribers) and expanded to Team and Enterprise plans on January 23. A Windows version followed in February 2026.[54]
Starting in February 2026, organizations can connect Claude Cowork to existing tools such as Google Drive, Gmail, DocuSign, and FactSet through enterprise connectors, and can deploy customizable plugins across domains like financial analysis, engineering, and human resources. Following the launches of Claude Cowork for Mac and Windows, enterprise software stocks shed a combined $285 billion in value as investors repriced companies whose core functionality overlapped with what Anthropic's desktop AI could automate.[54][55]
In March 2026, Anthropic expanded Cowork with a Projects feature, allowing users to attach local folders and organize files, instructions, and task context inside a single workspace. This enables context reusability across sessions rather than isolated one-off tasks.[56]
Microsoft also adopted the Cowork technology for its own products. On March 9, 2026, Microsoft announced Copilot Cowork, a new agent in Microsoft 365 Copilot that uses Claude's technology to automate multi-step workflows across Outlook, Teams, Excel, and other Microsoft 365 applications. Copilot Cowork requires a separate $30-per-user-per-month license and is available through a new Microsoft 365 E7 bundle priced at $99 per user monthly. The feature entered a limited research preview in March 2026 with broader availability planned for late March.[57]
Anthropic launched Dispatch on March 17, 2026 as a research preview within Claude Cowork. Dispatch creates a persistent conversation thread that syncs between the Claude mobile app (iOS and Android) and the desktop app, enabling a "phone-to-desktop" workflow. Users can send instructions from their phone (for example, "Organize my Q1 tax receipts on my Mac" or "Export the pitch deck and attach it to the meeting invitation at 2 p.m.") and Claude executes the task on the desktop application in its sandboxed environment.[58]
The bridge between phone and desktop is end-to-end encrypted. Even when controlled remotely, Claude pauses and sends a push notification to the user's phone before performing potentially destructive actions such as deleting files or moving large directories. The agent can only access the specific folders or applications the user has explicitly shared with the Cowork app.[58]
Dispatch rolled out first to Max plan users on March 17, with Pro plan access following the next day. Anthropic has indicated that a broader rollout to free users is planned for later in 2026.[58]
Introduced October 16, 2025, Claude Skills (also called Agent Skills) are modular capabilities that extend Claude's functionality through:
Packaged workflows containing instructions, scripts, and resources
Progressive disclosure mechanism for efficient context management
Automatic loading when relevant to tasks
Support for both Anthropic-provided skills (document processing) and custom user-created skills
Available across Claude apps, Claude Code, and the API
Claude Skills enable users to create reusable, specialized behaviors for common tasks without repetitive prompting. Skills operate in Claude's sandboxed code execution environment and can include executable scripts for deterministic operations. Enterprise organizations like Rakuten, Box, and Notion have implemented Skills to streamline workflows and maintain consistency across AI-generated outputs.[26][27]
Launched March 2025, providing:
Real-time information retrieval
Initially available to paying US users
Integration with current events and updated information
Claude integrates with Microsoft 365, allowing users to connect with SharePoint, OneDrive, Outlook, and Teams to access and analyze documents, emails, and meeting summaries. Claude for Excel, Word, and PowerPoint add-ins share a single conversation thread through Shared Context, enabling continuity across applications. These integrations are available to Max, Team, and Enterprise users.[49]
On May 7, 2026, Anthropic announced general availability of Claude inside Microsoft Word, Excel, and PowerPoint for all paid Claude plans (Pro, Team, and Enterprise), and Claude for Outlook entered public beta. A single Microsoft AppSource listing covers Excel, Word, and PowerPoint, with a separate listing adding Outlook. Conversation context follows the user across applications, so an inbox triage in Outlook can transition into a memo in Word and a deck in PowerPoint without re-introducing the topic. Separately, Microsoft enabled Anthropic models as the default Copilot model for Word, Excel, and PowerPoint inside Microsoft 365 Copilot starting May 4, 2026.[90]
Claude for Excel is a Microsoft 365 add-in that embeds Claude directly inside Excel spreadsheets. Users can switch between Sonnet 4.6 and Opus 4.6 as the underlying model, with Opus 4.6 recommended for complex financial models and multi-tab analysis.[59]
In early 2026, Anthropic significantly expanded Claude for Excel with native spreadsheet editing tools. The update added support for pivot table editing (sorting, filtering, and modifying schemas of existing pivot tables), conditional formatting, and other operations that bring it closer to a full Excel editing experience. The tool also gained MCP connector support, letting Claude pull data from external financial platforms such as S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet without leaving Excel.[59]
Claude for Excel is available on Pro, Max, Team, and Enterprise plans.[59]
Claude for PowerPoint was released as a research preview on February 5, 2026, alongside the Opus 4.6 launch. The add-in embeds Claude inside PowerPoint, where it generates new slides, edits existing ones, and builds complete presentation structures from natural-language prompts.[60]
A distinguishing feature is template compliance: Claude reads the open presentation's slide master, layouts, fonts, and color scheme before generating or editing content, so that new slides follow the same branding and formatting constraints as the rest of the deck. Claude can also convert bullet points into editable native PowerPoint visuals such as diagrams, process flows, and charts (not static images). Users can choose between Sonnet 4.5 and Opus 4.6 as the underlying model.[60]
Claude for PowerPoint is currently available to Max ($100/month), Team, and Enterprise plan subscribers. It is not yet included in the Pro plan.[60]
On March 6, 2026, Anthropic launched Claude Marketplace, a curated enterprise platform where businesses can purchase third-party software tools built on Claude's models. Enterprise customers with committed annual spending on Anthropic's API and services can apply a portion of that spend toward Marketplace purchases, and Anthropic does not take a commission on the transactions.[61]
The Marketplace launched in limited preview with six partners:
| Partner | Domain |
|---|---|
| GitLab | Software development lifecycle |
| Harvey | Legal AI workflows |
| Lovable | No-code app development |
| Replit | Developer platform |
| Rogo | Financial analysis |
| Snowflake | Enterprise data operations |
Anthropic and Snowflake also announced a $200 million multi-year partnership in early 2026, giving Claude access to Snowflake's 12,600 global customers. Anthropic plans to bring additional third-party products to the Marketplace over time. Enterprises interested in early access must contact their Anthropic account team.[61]
On April 17, 2026, Anthropic Labs launched Claude Design, a research-preview product that uses Claude Opus 4.7 to generate visual artifacts directly from natural-language descriptions. Designed for users who do not start from a dedicated design tool, it produces designs, prototypes, slide decks, one-pagers, and similar deliverables, then lets users refine results through conversation, inline comments, direct edits, or sliders generated by Claude itself.[64][73]
Claude Design can read a team's existing codebase and design files to extract a reusable design system (colors, typography, components), so subsequent projects automatically inherit brand styling. Outputs can be exported as PDF, public URL, or PPTX, or sent to Canva for collaborative editing. Once a design is approved, Claude packages assets and instructions into a handoff bundle that can be passed to Claude Code with a single command, closing the loop from idea to prototype to production code inside Anthropic's ecosystem. Claude Design is available in research preview to Pro, Max, Team, and Enterprise subscribers, and its launch sent Figma's stock down more than 7% on the day of the announcement.[64][73]
Claude Managed Agents is Anthropic's hosted agent platform that lets developers describe agentic workloads declaratively and run them inside Anthropic's infrastructure rather than self-hosting orchestration logic. At Code with Claude 2026 (May 6, 2026), Anthropic introduced three new primitives for the platform:[76][77]
Multiagent Orchestration (public beta): Lets a lead agent delegate work to specialist agents with their own models, prompts, and tools. Specialists run in parallel on a shared filesystem and contribute results back to the lead agent's context. The on-stage demo showed Commander, Detector, and Navigator agents coordinating a lunar drone landing scenario.[76][91]
Outcomes (public beta): Allows developers to write a rubric describing what success looks like, and then runs the agent against an independent grader living in its own context window so it is not biased by the agent's reasoning. When the grader flags an issue, the agent iterates until the rubric is satisfied. Anthropic compared the loop to the so-called "Ralph loop" pattern used internally.[76][91]
Dreaming (research preview): Reviews past agent sessions and memory stores on a schedule (typically off-hours), extracts patterns the agent could not see from a single session, and curates memories so the agent improves over time. Dreaming can update memory automatically or surface proposed updates for human review. Anthropic cited legal-AI partner Harvey reporting roughly a six-fold increase in task-completion rates after enabling Dreaming.[76][87]
Together with the existing Memory primitive, Anthropic described the combination as approximating an "infinite context window" feel by letting agents recall, distill, and re-use information beyond a single 1M-token session.[91]
In April and May 2026, Anthropic published a series of vertical Claude packages combining MCP connectors, Claude Skills, and ready-to-run agentic workflows for specific industries or roles:
Claude for Creative Work (April 28, 2026): Connectors and Skills for creative professionals, integrating with Ableton, Adobe Creative Cloud, Affinity by Canva, Autodesk Fusion, Blender, Resolume Arena/Wire, SketchUp, and Splice. Anthropic also gave free Claude access to students and faculty at Rhode Island School of Design, Ringling College of Art and Design, and Goldsmiths, University of London as part of the launch.[88]
Claude for Financial Services (May 5, 2026): Ten ready-to-run agent templates targeting banking, insurance, asset management, and fintech. Five cover research and client coverage (pitchbooks, meeting prep, earnings review, financial-model building, market research), and five cover finance operations (valuation review, general-ledger reconciliation, month-end close, statement auditing, KYC screening). The release coincided with the general availability of Claude for Excel, PowerPoint, and Word, with FactSet, Morningstar, S&P Global, and Moody's shares all declining on the announcement as investors repriced incumbent data and analytics vendors.[79][89]
Claude for Small Business (May 13, 2026): A toggle-install package that ships with 15 agentic workflows covering finance, operations, sales, marketing, HR, and customer service, plus connectors to QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365, Slack, Square, Stripe, and Webflow. Anthropic paired the launch with the "Claude SMB Tour" of ten U.S. cities (starting in Chicago on May 14) and a free on-demand "AI Fluency for Small Business" course developed with PayPal.[80][82]
Claude for Legal (May 12, 2026): A suite of more than 20 MCP connectors and 12 practice-area plugins covering Commercial, Corporate, Employment, Privacy, Product, Regulatory, AI Governance, IP, and Litigation work, with a separate plugin for law students studying for the bar exam. Connectors target Ironclad, DocuSign, Definely, iManage, NetDocuments, Relativity, Everlaw, Consilio, Box, Datasite, Midpage, Trellis, Legal Data Hunter, and others. Thomson Reuters rebuilt its CoCounsel product on Claude as part of the announcement, creating a bidirectional integration in which CoCounsel runs on Claude and Claude can also call CoCounsel as a tool. Freshfields, Quinn Emanuel Urquhart & Sullivan, Holland & Knight, and Crosby Legal were named as early users.[83]
On May 4, 2026, Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announced the formation of a new AI-native enterprise services firm capitalized at approximately $1.5 billion. The standalone company embeds Anthropic engineers alongside its own staff to redesign workflows and integrate Claude into client operations, initially targeting mid-sized companies in healthcare, manufacturing, financial services, retail, and real estate (often inside the founders' private-equity portfolios). Additional backers include General Atlantic, Leonard Green, Apollo Global Management, GIC, and Sequoia Capital. The venture became a member of Anthropic's Claude Partner Network at launch and was widely characterized as a direct challenge to incumbent global consulting firms.[81]
The following week, on May 14, 2026, Anthropic and PwC expanded their existing alliance into a multi-year program in which PwC will deploy Claude Code and Claude Cowork to its U.S. workforce, train and certify 30,000 PwC professionals on Claude, and establish a joint Center of Excellence. The program also stood up an "Office of the CFO" finance business unit at PwC built on Claude as its anchoring technology. PwC said it expects to expand the rollout from the U.S. toward its global workforce of hundreds of thousands of professionals.[78]
Claude models are regularly evaluated on standardized benchmarks. As of May 2026, Claude Opus 4.8 holds the family's top result on SWE-bench Verified at 88.6%, a leading score for autonomous software engineering. The table below summarizes key results for recent models.
| Benchmark | Sonnet 4.5 | Opus 4.5 | Opus 4.6 | Sonnet 4.6 | Opus 4.7 | Opus 4.8 |
|---|---|---|---|---|---|---|
| SWE-bench Verified | 77.2% | 80.9% | 80.8% | 79.6% | 87.6% | 88.6% |
| SWE-bench Pro | N/A | N/A | N/A | N/A | 64.3% | N/A |
| Terminal-Bench 2.0 | N/A | N/A | 65.4% | N/A | Improved over 4.6 | N/A |
| Terminal-Bench 2.1 | N/A | N/A | N/A | N/A | 66.1% | 74.6% |
| Online-Mind2Web (browser agent) | N/A | N/A | N/A | N/A | N/A | 84% |
| METR 50% time-horizon | N/A | ~4 hours | ~14.5 hours | N/A | Not yet published | Not yet published |
| MRCR v2 (8-needle, 1M) | 18.5% | N/A | 76% | N/A | Not yet published | Not yet published |
| Visual acuity (vision) | N/A | N/A | 54.5% | N/A | 98.5% | Not yet published |
The following table shows the progression of SWE-bench Verified scores across the Claude 4.x family:
| Model | SWE-bench Verified | Release Date |
|---|---|---|
| Claude Opus 4 | 72.5% | May 2025 |
| Claude Sonnet 4 | 72.7% | May 2025 |
| Claude Opus 4.1 | 74.5% | August 2025 |
| Claude Sonnet 4.5 | 77.2% | September 2025 |
| Claude Opus 4.5 | 80.9% | November 2025 |
| Claude Sonnet 4.6 | 79.6% | February 2026 |
| Claude Opus 4.6 | 80.8% | February 2026 |
| Claude Opus 4.7 | 87.6% | April 2026 |
| Claude Opus 4.8 | 88.6% | May 2026 |
METR, a nonprofit focused on evaluating AI capabilities, estimated that Opus 4.6 has a 50% time-horizon of approximately 14.5 hours (95% confidence interval: 6 to 98 hours) on software tasks. This means Opus 4.6 can successfully complete tasks that would take a skilled human professional about 14.5 hours roughly half the time. METR noted that their task suite is nearly saturated at this level, making measurement increasingly noisy. The fitted trend line from METR data shows AI task-completion capability approximately doubling every 123 days since 2023.[16]
On SWE-bench Verified, Opus 4.5's 80.9% score outperformed OpenAI's GPT-5.1-Codex-Max (77.9%) and Google's Gemini 3 Pro (76.2%) at the time of its release. Opus 4.6 achieved the highest score ever recorded on Terminal-Bench 2.0 (65.4%), a leading evaluation for agentic coding systems, and leads on BrowseComp, a benchmark for information retrieval.[15][51]
| Platform | Scope | Key Dates / Notes |
|---|---|---|
| Anthropic (claude.ai; iOS/Android; API) | Direct chat; developer API; Claude Code | Ongoing; Sonnet 4.6 default from 17 Feb 2026; Opus 4.7 added 16 Apr 2026; Opus 4.8 added 28 May 2026 |
| Amazon Bedrock | Managed access to Claude family for AWS customers | All current models available |
| Vertex AI (Model Garden) | Managed access on Google Cloud | Claude 3 models and later variants available |
| Microsoft Azure (Foundry) | Managed access on Azure | Opus 4.5 onward available |
| GitHub Copilot | AI coding assistant integration | Sonnet 4.5 added September 2025; Opus 4.6 added February 2026; Opus 4.7 generally available April 16, 2026 |
| Microsoft 365 (Word, Excel, PowerPoint) | Office add-ins with Shared Context | GA on all paid plans May 7, 2026 |
| Microsoft 365 (Outlook) | Inbox triage, drafts, calendar invites | Public beta May 7, 2026 |
| Microsoft 365 Copilot | Anthropic models as default Copilot model | Default since May 4, 2026 (Word/Excel/PowerPoint) |
| Microsoft 365 Copilot (Copilot Cowork) | Agentic task automation across M365 apps | Research preview, March 2026; E7 bundle |
| Claude Marketplace | Third-party enterprise software store | Limited preview, March 6, 2026 |
| Claude Design (Anthropic Labs) | AI-powered visual design and prototyping | Research preview, April 17, 2026 (Pro/Max/Team/Enterprise) |
| Claude Mythos Preview (Project Glasswing) | Cybersecurity-focused frontier model for vulnerability discovery | Restricted preview, April 7, 2026 (vetted partners only) |
| Claude Managed Agents | Hosted agent platform with Multiagent Orchestration, Outcomes, Dreaming | Public beta / research preview, May 6, 2026 |
| Claude for Small Business | SMB-focused workflows, skills, and connectors | Launched May 13, 2026 |
| Plan | Features | Price |
|---|---|---|
| Free | Access to Sonnet 4.6, Memory, basic features, inline visualizations | $0 |
| Pro | Higher rate limits, priority access, Chat Search, Cowork, Claude for Excel, computer use (Mac) | $20/month |
| Max 5x | 5x Pro usage limits, priority new model access, Claude for PowerPoint | $100/month |
| Max 20x | 20x Pro usage limits, priority new model access, Claude for PowerPoint | $200/month |
| Team (Standard seat) | Collaboration features, shared workspaces, Projects, Claude Code included | $25/user/month (annual) |
| Team (Premium seat) | All Standard features plus Claude Code and early access to new collaboration features | $150/user/month (monthly); $100/user/month (annual) |
| Enterprise | Advanced features, dedicated support, custom deployment, Compliance API | Custom pricing |
The Max plan was launched in April 2025 for power users and professionals who need significantly higher usage limits. It was Anthropic's direct answer to OpenAI's $200-per-month ChatGPT Pro tier.[28]
Team and Enterprise plans received Claude Code inclusion at no additional cost starting January 2026. Admins can assign standard or premium seats based on individual needs, with premium seats providing access to both Claude and Claude Code. Enterprise customers also gained a Compliance API for real-time programmatic access to Claude usage data, enabling continuous monitoring and automated policy enforcement.[50]
API Pricing (per million tokens, as of March 2026):
| Model | Input | Output | Batch (50% discount) |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | $0.50 / $2.50 |
| Claude Sonnet 4.6 | $3 | $15 | $1.50 / $7.50 |
| Claude Opus 4.6 | $5 | $25 | $2.50 / $12.50 |
| Claude Opus 4.7 | $5 | $25 | $2.50 / $12.50 |
| Claude Opus 4.8 | $5 | $25 | $2.50 / $12.50 |
For Sonnet 4.6 requests exceeding 200K input tokens, long-context pricing of $6/$22.50 applies. Opus 4.6 also offers a fast mode at $30/$150 for latency-sensitive workloads, and on May 13, 2026, Anthropic extended Fast Mode to Opus 4.7 in research preview at the same $30/$150 rate; on May 14, 2026, Opus 4.7 Fast became the default Fast Mode model in place of Opus 4.6 Fast. Fast Mode is not a separate model; it serves the same Opus 4.7 weights with additional inference compute to generate tokens at roughly 2-3x the standard throughput. Prompt caching is available across all models, with cache writes at 1.25x the input price and cache reads at 0.1x the input price.[29][92]
On May 14, 2026, Anthropic announced a restructuring of subscription billing that takes effect on June 15, 2026. Chat (web, desktop, mobile) and first-party tools (Claude Code in the terminal, Claude Cowork) continue to count against existing subscription rate limits. Third-party agent and SDK traffic, including the Anthropic Agent SDK, the claude -p non-interactive command, Claude Code GitHub Actions, and third-party agent platforms such as OpenClaw, Conductor, Zed, and Jean, will be moved into a new "Agent SDK credit" pool. Monthly Agent SDK credits are set at $20 for Pro, $100 for Max 5x, and $200 for Max 20x, with overages billed at API rates. Anthropic indicated that subscribers will receive an email on June 8 to claim their credits, and that the change is intended to address heavy programmatic usage that had been straining shared subscription pools.[93]
Claude competes primarily with OpenAI's GPT-4 and GPT-5 series and Google's Gemini family. The second half of 2025 saw intense competition: Claude Sonnet 4.5 shipped September 29, OpenAI released GPT-5.1 on November 12, and Google launched Gemini 3 on November 18.[30]
As of early 2026, the competitive picture is mixed. Claude models lead on SWE-bench Verified (coding tasks), while GPT-5.2 outperforms on ARC-AGI-2 (abstract reasoning) and Gemini 3 Deep Think holds the top score on Humanity's Last Exam (broad knowledge). On pricing, Anthropic's Opus 4.6 at $5/$25 per million tokens sits between Gemini 3.1 Pro ($2/$12) and OpenAI's GPT-5.2 ($1.75/$14).[30]
With the April 16, 2026 release of Opus 4.7, Anthropic widened its lead on coding evaluations: Opus 4.7 reached 87.6% on SWE-bench Verified and 64.3% on SWE-bench Pro, both #1 at the time, while keeping the Opus 4.5/4.6 list price of $5/$25. The Claude Mythos Preview release a week earlier also signaled that Anthropic was willing to hold back its strongest cyber-capable models from open release, in contrast to broader public releases by some rivals. Reporting in mid-April 2026 noted that Anthropic's run-rate revenue had surpassed OpenAI's for the first time, with Anthropic at roughly $30 billion ARR versus OpenAI at roughly $25 billion.[63][67][72]
In May 2026, competitive dynamics shifted in two other ways. First, on May 4, 2026, Anthropic and OpenAI both announced enterprise AI services joint ventures on the same day: Anthropic with Blackstone, Hellman & Friedman, and Goldman Sachs, and OpenAI with a separate consortium. Industry analysts framed the two announcements as a direct challenge to the traditional consulting industry. Second, on May 6, 2026, Anthropic struck a compute deal with SpaceX's Colossus 1, in which xAI is co-owner, illustrating that even direct model competitors were now sharing infrastructure under capacity constraints.[75][81]
ChatGPT maintains the largest user base with around 800 million weekly active users, though its market share has declined from over 87% to roughly 68% as Claude and Gemini have closed capability gaps.[30]
In the developer tools space, Claude Code competes with GitHub Copilot, Cursor, and other AI-powered coding assistants. Claude Code's 4% share of public GitHub commits as of early 2026 has made it a significant player in the AI-assisted software development market.[25]
Anthropic classifies its models using an internal Anthropic Safety Level (ASL) scale. Claude 4 models are classified as ASL-3, indicating they are powerful enough to pose "significantly higher risk" and require enhanced safety measures.[31]
Safety measures include:
Extensive red teaming exercises to identify vulnerabilities
Bias mitigation through Constitutional AI
External evaluations by independent researchers
Collaboration with organizations like Thorn for child safety
Reduced susceptibility to prompt injection and jailbreaking attempts
Harmless response rates tracked across model versions: 97.27% for Opus 4, 98.76% for Opus 4.1, and 99.78% for Opus 4.5[32][43][15]
Anthropic's safety posture became increasingly visible in 2026. With the Claude Mythos Preview, the company publicly disclosed a frontier model with strong cyber-offensive capability and chose to gate access to a small set of vetted partners under Project Glasswing rather than offer it via the standard API. The U.K. AI Security Institute (AISI) independently evaluated Mythos and reported that it was the first model to complete AISI's full-network attack simulation. When Anthropic later shipped the public Opus 4.7 model, it included automatic safeguards that detect and block requests indicating prohibited or high-risk cybersecurity uses, and Anthropic emphasized that Opus 4.7 was deliberately less broadly capable than Mythos in cyber tasks. With Claude Opus 4.8 in May 2026, Anthropic reported that the model was roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked and showed substantially lower rates of misaligned behavior such as deception and cooperation with misuse. In March 2026, an unauthorized group reportedly gained brief access to Mythos, prompting renewed industry debate about how labs should handle cyber-capable frontier models.[63][68][69][74][98]
Claude has received praise for:
Leading performance on coding benchmarks (80.9% on SWE-bench Verified for Opus 4.5; 80.8% for Opus 4.6; 88.6% for Opus 4.8)
Large context window capabilities enabling processing of entire books and codebases
Natural, human-like writing style
Strong safety and alignment measures
Practical utility in enterprise settings
Sonnet 4.6 delivering near-Opus performance at a fraction of the cost
Rapid price reductions (67% cut for Opus between generations 4 and 4.5)
Claude Cowork's desktop automation capabilities for non-technical knowledge workers
Enterprise adoption has grown significantly, with a 5.5x revenue increase reported in August 2025 and over 300,000 business customers by late 2025. Reviews highlight Claude's coding strengths, and Sonnet 4.6 has been described as matching flagship-level performance for the majority of professional workloads.[33]
Some users have noted that Claude's strong ethical alignment can reduce usability, with the model refusing to assist with benign requests that it misinterprets as harmful. For example, the system administration question "How can I kill all python processes in my Ubuntu server?" has been refused due to the word "kill."[34]
Anthropic's web crawler, ClaudeBot, has faced criticism for:
Allegedly ignoring robots.txt protocols
Causing excessive server loads during data collection
Aggressive crawling behavior affecting website performance[35]
On October 18, 2023, Anthropic was sued by Concord, Universal, ABKCO, and other music publishers for alleged systematic and widespread infringement of copyrighted song lyrics.[36]
In a separate case, Bartz v. Anthropic, three authors (Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson) sued Anthropic in 2024, alleging it had downloaded millions of copyrighted books from pirated "shadow libraries" such as Library Genesis and Pirate Library Mirror to train Claude. In June 2025, a federal judge ruled that training on lawfully acquired books was transformative fair use, while allowing claims over pirated copies to proceed. In late August 2025, Anthropic agreed to pay roughly $1.5 billion to settle the case, the largest copyright settlement in U.S. history, covering about 500,000 works at approximately $3,000 per work. A final fairness hearing before Judge Araceli Martinez-Olguin was held on May 14, 2026, with the claims rate exceeding 92%; the court took final approval under submission.[99][100]
The Claude mobile app saw tepid reception compared to ChatGPT's launch, with slower adoption rates despite comparable functionality.[37]
In October 2025, Anthropic updated its policy so that chats from personal (non-enterprise) users are used to improve models by default, with an opt-out offered in account settings. Key points:
Retention extended to five years
Enterprise, government, and education accounts excluded from training use
Users can opt out through account settings
No training on user data without permission for enterprise accounts[38]
Constitutional AI, articulated in a 2022 paper, has influenced Claude's training and alignment approach and is cited by Anthropic as a scalable method that reduces reliance on large volumes of human-labeled "harms" data. The approach represents a significant departure from traditional RLHF methods used by competitors.[17]
Anthropic's research has contributed to the field of AI safety through:
Development of Constitutional AI methodology
Research on AI interpretability and mechanistic understanding
Work on scalable oversight and alignment techniques
Publications on red teaming and safety evaluation methods
Creation and open-sourcing of the Model Context Protocol (MCP)[39][46]
Generative artificial intelligence
Constitutional AI
AI safety
Claude Shannon