Claude Opus 4.7
Last reviewed
May 7, 2026
Sources
30 citations
Review status
Source-backed
Revision
v5 ยท 7,977 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 7, 2026
Sources
30 citations
Review status
Source-backed
Revision
v5 ยท 7,977 words
Add missing citations, update stale details, or suggest a clearer explanation.
Claude Opus 4.7 is a hybrid reasoning large language model developed by Anthropic and announced on April 16, 2026. It is positioned as Anthropic's most capable generally available model at launch, succeeding Claude Opus 4.6 in the Opus tier of the Claude 4 family. The model targets long-horizon agentic work, advanced software engineering, vision-heavy workflows, and enterprise knowledge tasks that previously required close human supervision.[1][2]
Opus 4.7 supports a 1 million token context window at standard API pricing, a 128,000 token maximum output, adaptive thinking, and high-resolution image understanding. It is offered through the Claude consumer apps, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The release coincided with the introduction of Project Glasswing, a defensive cybersecurity initiative that includes the more capable but unreleased Claude Mythos Preview model.[1][3][4]
At launch Anthropic stated that Opus 4.7 outperforms Opus 4.6 across coding, multidisciplinary reasoning, scaled tool use, and agentic computer use. The company also publicly conceded that the unreleased Mythos Preview is more broadly capable. Opus 4.7 is intended to be the production model that customers can deploy at scale; Mythos remains invitation-only inside Project Glasswing. This split between a generally available frontier model and a more capable invitation-only model was a first for Anthropic and was widely discussed by AI press at launch.[5][6]
The model uses the API identifier claude-opus-4-7. It is the seventh distinct model in the Opus subfamily after Claude Opus 4, Claude Opus 4.1, Claude Opus 4.5, and Claude Opus 4.6, and the eighth Claude 4 model overall when counting Claude Sonnet 4.5, Claude Sonnet 4.6, and Claude Haiku 4.5. The release is also notable for being one of the largest day-one rollouts Anthropic has ever conducted, with the model live across claude.ai, the API, Bedrock, Vertex AI, Microsoft Foundry, Claude Code, Cursor, and GitHub Copilot within hours of the announcement.[7][16]
Claude Opus 4.7 was released on April 16, 2026, roughly two months after Claude Opus 4.6 (February 5, 2026) and almost one year after Claude Opus 4 and Claude Sonnet 4 (May 22, 2025). It uses the API model identifier claude-opus-4-7 with no date suffix; from the 4.6 generation onward Anthropic adopted dateless pinned snapshot identifiers rather than the previous YYYYMMDD-suffixed format.[1][2]
The Opus subfamily progression to date is summarized below.
| Model | Release date | API ID | Notable change |
|---|---|---|---|
| Claude Opus 4 | May 22, 2025 | claude-opus-4-20250514 | First Claude 4 generation Opus, hybrid reasoning, ASL-3 |
| Claude Opus 4.1 | August 5, 2025 | claude-opus-4-1-20250805 | Coding and tool use improvements |
| Claude Opus 4.5 | November 24, 2025 | claude-opus-4-5-20251101 | Pricing reduction, effort parameter, 80%+ SWE-bench |
| Claude Opus 4.6 | February 5, 2026 | claude-opus-4-6 | 1M context window, adaptive thinking, agent teams |
| Claude Opus 4.7 | April 16, 2026 | claude-opus-4-7 | New tokenizer, adaptive thinking only, xhigh effort, high-res vision |
At the time of release, the rest of the Claude family included Claude Sonnet 4.6 (released February 17, 2026) and Claude Haiku 4.5 (released October 15, 2025). Claude Opus 4 and Claude Sonnet 4 from May 2025 were officially deprecated and scheduled for retirement on June 15, 2026, with Anthropic recommending migration to Opus 4.7 and Sonnet 4.6 respectively.[2]
The roughly ten-week cadence between Opus 4.6 and Opus 4.7 was faster than the gap between previous Opus iterations. Anthropic cited three engineering investments that were ready earlier than expected: a new tokenizer, a high-resolution vision pipeline, and the consolidation of adaptive thinking as the only thinking-on mode.[7]
The model exposes the same Anthropic Messages API surface as earlier Claude 4 models, with a small set of breaking changes (see API changes below). The published specifications are summarized in the table.
| Specification | Value |
|---|---|
| Provider | Anthropic |
| Model family | Claude 4 |
| API identifier | claude-opus-4-7 |
| AWS Bedrock ID | anthropic.claude-opus-4-7 |
| GCP Vertex AI ID | claude-opus-4-7 |
| Context window | 1,000,000 tokens |
| Maximum output (Messages API) | 128,000 tokens |
| Maximum output (Batch API beta) | 300,000 tokens with output-300k-2026-03-24 header |
| Modalities | Text and image input, text output |
| Image resolution limit | 2576px / 3.75 megapixels |
| Tokenizer | New tokenizer, roughly 1.0x to 1.35x more tokens than Opus 4.6 for the same text |
| Reliable knowledge cutoff | January 2026 |
| Training data cutoff | January 2026 |
| Adaptive thinking | Yes (only thinking mode supported) |
| Extended thinking with fixed budget | Removed (returns 400 error) |
| Sampling parameters | temperature, top_p, top_k removed (return 400 error) |
| Effort levels | low, medium, high, xhigh, max |
| Languages | Multilingual (incl. Chinese, Japanese, Arabic, Spanish, French) |
| Vision | Yes, with high-resolution support |
| Tool use | Function calling, computer use, MCP, memory tool |
| Priority Tier | Yes |
| ASL classification | ASL-3 (per Anthropic Responsible Scaling Policy) |
The 1 million token context corresponds to roughly 555,000 English words or 2.5 million Unicode characters under the new tokenizer, according to Anthropic's documentation. Anthropic recommends increasing max_tokens headroom and compaction triggers when migrating from Opus 4.6 to absorb the higher token consumption.[2][7]
The new tokenizer uses a denser vocabulary for non-Latin scripts. Independent tests by MindStudio and Caylent reported token-count reductions of 20 to 35 percent for Mandarin, Japanese, and Arabic source text, while typical English prose grew by 12 to 18 percent. The net effect on cost depends heavily on the language mix of a workload, and Anthropic's documentation notes that token efficiency varies by content type.[7][13]
Opus 4.7 launched at the same per-token rates as Opus 4.6. There is no long-context premium: a request that fills the full 1 million token window is billed at the same per-token rate as a small request, a deliberate change from the tiered long-context pricing introduced with Opus 4.6 in beta.[2][8]
| Item | Rate |
|---|---|
| Input tokens | $5 per million |
| Output tokens | $25 per million |
| Prompt caching | Up to 90% discount on cached input |
| Batch API | 50% discount on input and output |
| US-only inference | 1.1x multiplier on standard rates |
| Bedrock pricing | Matches API list pricing |
Because the new tokenizer can produce up to about 35 percent more tokens for the same source English text, third-party analyses noted that the effective per-task cost can rise even though the per-token rate is unchanged. Anthropic suggests using prompt engineering, the new task budgets feature, and lower effort levels to control real-world spend. Finout's analysis of the pricing change framed it as a quiet cost increase masked by an unchanged sticker price; Caylent's analysis framed it as cost-neutral or favorable for multilingual workloads, where the new tokenizer is more efficient.[7][8][13]
Artificial Analysis lists Opus 4.7 (max effort) at a blended price of $10.94 per million tokens at typical input-to-output ratios on its leaderboard, reflecting both the per-token rates and the longer outputs that max effort tends to produce.[24]
Claude Opus 4.7 is available across the following surfaces at launch:
| Surface | Notes |
|---|---|
| claude.ai | Default Opus option for Claude Pro, Max, Team, and Enterprise users |
| Claude API | Direct access via claude-opus-4-7 |
| Amazon Bedrock | US East (N. Virginia), Asia Pacific (Tokyo), Europe (Ireland), Europe (Stockholm) at launch, with capacity up to 10,000 requests per minute per account per region |
| Google Cloud Vertex AI | Global, multi-region, and regional endpoints |
| Microsoft Foundry | Available in the Foundry catalog |
| GitHub Copilot | Pro+, Business, and Enterprise tiers, gradual rollout, 7.5x premium request multiplier through April 30, 2026 (15x thereafter) |
| Snowflake Cortex AI | Native availability with Snowflake account credentials |
| Cursor, Cline, and Continue | Day-one model picker support |
| Claude Code | Default model for Pro and Max plans, requires v2.1.111 or later |
On Bedrock, Anthropic offers a zero operator access guarantee in which prompts and responses are not visible to AWS or Anthropic operators. The platform also exposes Opus 4.7 through the Converse API, the Invoke API, the AWS CLI bedrock-runtime endpoint, and an OpenAI-compatible Responses API.[3]
GitHub initially priced the model at a 7.5x premium request multiplier in Copilot during a promotional window through April 30, 2026, after which the multiplier was raised to 15x in line with the model's compute cost. Over the weeks following launch the model picker in GitHub Copilot replaced both Opus 4.5 and Opus 4.6 with Opus 4.7 as the default Opus option. Snowflake added native Cortex AI access on launch day, allowing Snowflake customers to call the model with their existing account credentials and avoid setting up separate Anthropic billing.[16][26]
Anthropic does not publish full architectural details for Claude models, including parameter count, training compute, or training corpus composition. Public documentation describes Opus 4.7 as a hybrid reasoning system: a single model that can answer quickly for simple requests and switch into longer chain-of-thought reasoning for harder problems through adaptive thinking. The model uses the same broad architecture as the rest of the Claude 4 generation, with refinements in instruction following, tool use, and memory.[1][2]
A few public technical details are confirmed in the model card and platform documentation:
Like previous Claude 4 models, Opus 4.7 was trained with a constitutional AI approach that combines reinforcement learning from human feedback with rules and principles applied in self-critique and revision. The system card describes additional rounds of red-teaming, automated alignment evaluation, and reinforcement learning from feedback specifically targeting agentic behavior over long horizons. The 272-page model card devotes substantial space to evaluations of agentic safety on tasks measured in hundreds or thousands of tool calls.[19][22]
Anthropic published benchmark numbers for Opus 4.7 alongside the Mythos Preview, Opus 4.6, OpenAI's GPT-5 family (specifically GPT-5.4), and Google's Gemini 3 line (Gemini 3.1 Pro). Independent aggregators including Vellum, llm-stats, Artificial Analysis, and the LMArena leaderboard reproduced the headline figures within a few percentage points. The selection below focuses on the most widely cited results at launch.[6][11][12]
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro | Mythos Preview |
|---|---|---|---|---|---|
| SWE-bench Verified | 87.6% | 80.8% | n/a | 80.6% | 93.9% |
| SWE-bench Pro | 64.3% | 53.4% | 57.7% | 54.2% | 77.8% |
| Terminal-Bench 2.0 | 69.4% | 65.4% | 75.1% | 68.5% | 82.0% |
| HumanEval | 95.2% | ~93% | n/a | n/a | n/a |
| LMArena Code Arena (Elo) | #1 (~1521) | #2 (~1484) | n/a | n/a | n/a |
Anthropic reported a roughly 13 percent improvement on its internal 93-task coding evaluation versus Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. Customer testimonials in the launch post described "3x more production tasks" resolved on the Rakuten-SWE-Bench harness and a 70 percent pass rate on CursorBench, up from 58 percent for the prior model. The LMArena Code Arena ranking placed Opus 4.7 first overall, ahead of Opus 4.6 by about 37 Elo and ahead of the next non-Anthropic model (GLM-5.1) by 46 Elo, with first-place finishes on both the React and HTML sub-leaderboards.[1][13][27]
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 Pro | Gemini 3.1 Pro | Mythos Preview |
|---|---|---|---|---|---|
| GPQA Diamond | 94.2% | 91.3% | 94.4% | 94.3% | 94.6% |
| MMLU | 89.8% | n/a | n/a | n/a | n/a |
| MMLU Pro | 89.9% | n/a | n/a | n/a | n/a |
| MMMLU (multilingual) | 91.5% | 91.1% | n/a | 92.6% | n/a |
| AIME 2025 (no tools, max effort) | 100.0% | 99.8% | n/a | n/a | n/a |
| MATH | 94.1% | n/a | n/a | n/a | n/a |
| Humanity's Last Exam (no tools) | 46.9% | 40.0% | 42.7% | 44.4% | 56.8% |
| Humanity's Last Exam (with tools) | 54.7% | 53.3% | 58.7% | 51.4% | 64.7% |
The AIME 2025 score of 100 percent is reported in the system card with the caveat that potential test contamination may have inflated the number. Anthropic averages over five trials with adaptive thinking at max effort and default sampling settings. The math benchmarks are essentially saturated at this point in Anthropic's view, and gains are concentrated in coding, agentic work, and vision rather than in pure mathematical reasoning.[19][22]
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro | Mythos Preview |
|---|---|---|---|---|---|
| MCP-Atlas | 77.3% | 75.8% | 68.1% | 73.9% | n/a |
| Finance Agent v1.1 | 64.4% | 60.1% | 61.5% (Pro) | 59.7% | n/a |
| OSWorld Verified | 78.0% | 72.7% | 75.0% | n/a | 79.6% |
| BrowseComp | 79.3% | 83.7% | 89.3% (Pro) | 85.9% | 86.9% |
| Tau-bench Retail | 86.5% | 84.7% | n/a | n/a | n/a |
| GDPval-AA (Elo) | ~1753 | 1606 | ~1462 | ~1195 | n/a |
MCP-Atlas measures coverage of the Model Context Protocol tool ecosystem and is reported as best-in-class for Opus 4.7 among generally available models. BrowseComp is a regression relative to Opus 4.6, and Anthropic acknowledged that Opus 4.7 is not the strongest model for open-web research, where GPT-5.4 Pro and Mythos Preview lead. GDPval-AA, a third-party Elo-style evaluation of economically valuable knowledge work, places Opus 4.7 at the top with roughly a 58 to 60 percent head-to-head win rate against GPT-5.4 across finance, legal, and consulting tasks.[11][12][25]
| Benchmark | Opus 4.7 | Opus 4.6 | Mythos Preview |
|---|---|---|---|
| MMMU | ~84% | ~80% | n/a |
| CharXiv (no tools) | 82.1% | 69.1% | 86.1% |
| CharXiv (with tools) | 91.0% | 84.7% | 93.2% |
| XBOW visual acuity | 98.5% | 54.5% | n/a |
The XBOW visual-acuity result is the largest single-benchmark jump in the launch and is attributed to the new high-resolution image pipeline. The benchmark, maintained by autonomous-pentest startup XBOW, measures whether a model can correctly read text and small UI elements in dense screenshots; the jump from 54.5 percent on Opus 4.6 to 98.5 percent on Opus 4.7 effectively turns it from a meaningful evaluation into a saturated one for the new model.[1][11]
Third parties reported strong gains on industry-specific benchmarks. On Harvey's BigLaw Bench, Opus 4.7 scored 90.9 percent at high effort, the highest score of any Claude model in Harvey to date, with 45 percent of tasks earning perfect scores and 88 percent at or above 0.80. On Notion's internal multi-step workflow evaluation, Opus 4.7 scored about 14 percent higher than Opus 4.6 while using fewer tokens and producing a third of the tool errors. On Stripe's internal financial reasoning benchmark, the company's VP of Technology described the model as catching its own logical faults during the planning phase and accelerating execution.[23]
LMArena (formerly LMSYS Chatbot Arena) ranked Opus 4.7 first overall at approximately 1504 Elo, ahead of Opus 4.6 (Thinking) at second. Opus 4.7 took first place on the overall, expert, and code arenas, and second on creative writing. Anthropic's models swept the top three positions on the overall leaderboard for the first time in the platform's history.[27]
Artificial Analysis listed Opus 4.7 (max effort) at third on its Intelligence Index leaderboard with a score of 57, tied with Gemini 3.1 Pro Preview, behind GPT-5.5 (xhigh) at 60 and GPT-5.5 (high) at 59. Output speed was reported at 44 tokens per second and first-chunk latency at 18.89 seconds at max effort.[24]
Opus 4.7 makes adaptive thinking the only supported thinking-on mode. The earlier extended-thinking option that let developers set a fixed budget_tokens value was removed; requests with the old syntax now return a 400 error. Adaptive thinking lets the model decide turn-by-turn how long to reason, including using interleaved thinking that mixes reasoning steps with tool calls inside a single agent loop.[7][9]
Adaptive thinking is off by default. To enable it, callers set thinking: {"type": "adaptive"} and choose an effort level on the output config:
| Effort level | Recommended use |
|---|---|
| low | Cost-sensitive, tightly scoped work; still beats Opus 4.6 |
| medium | Lightweight chat and short tool sequences |
| high | Concurrent agent sessions where intelligence and cost are both important |
| xhigh (default for coding) | Most coding and agentic uses; strong autonomy without excessive token use |
| max | Genuinely hard problems; can show overthinking and diminishing returns |
Claude Code's official guidance recommends xhigh as the default for most agentic coding tasks and max only for the hardest problems. Effort is the primary lever for trading capability against cost on Opus 4.7. Anthropic's internal evaluations show that adaptive thinking at any effort level reliably outperforms the older extended-thinking interface at the same nominal token budget, because the model can allocate compute where it actually helps rather than spending a fixed budget on every turn.[14]
Opus 4.7 introduces task budgets, a new field on the output config that gives Claude a soft target for total tokens across an agentic loop, including thinking, tool calls, tool results, and final output. The model can see a running countdown and uses it to prioritize work and finish gracefully as the budget is consumed. Task budgets are advisory rather than hard caps, while max_tokens continues to act as a hard per-request ceiling. The minimum task budget is 20,000 tokens, and the feature requires the task-budgets-2026-03-13 beta header.[7]
The distinction between task_budget and max_tokens is important. The model is aware of task_budget and uses it to scope work; it is unaware of max_tokens, which simply truncates output if hit. Anthropic recommends not setting a task budget at all for open-ended agentic tasks where quality matters more than speed; task budgets are intended for workloads where the operator wants the model to self-moderate around a known token allowance, for example a CI pipeline that must finish in a fixed time window.[7]
Opus 4.7 is the first Claude model with high-resolution image support. Maximum input resolution rises to 2,576 pixels on the long edge (about 3.75 megapixels) from the previous limit of 1,568 pixels (about 1.15 megapixels). Coordinates returned by the model now map 1:1 to pixels, which removes the scale-factor math previously required for tasks like clicking on UI elements during computer use. Anthropic also reports gains on low-level perception tasks such as pointing, measuring, counting, and natural-image bounding-box localization.[7]
The practical effects of the higher resolution show up most clearly in two places. Computer use traces for desktop applications with small UI elements (toolbar buttons, table cells, dropdown menus) are far more reliable. Document understanding for technical content with small text or dense diagrams improves substantially; the CharXiv benchmark for scientific chart and figure understanding rose from 69.1 percent to 82.1 percent without tools and from 84.7 percent to 91.0 percent with tools. The XBOW visual acuity benchmark, which tests whether a model can read text in screenshots at native resolution, jumped from 54.5 percent to 98.5 percent.[7][11]
Opus 4.7 is described as meaningfully better at writing and using file-system-based memory across long, multi-session work. Anthropic provides a managed memory tool that gives an agent a scratchpad without requiring the developer to build their own storage. The model uses memory to retain prior work and operate with less upfront context on later turns. The same tool is available across the Claude 4 family, but Opus 4.7 is the first model where Anthropic recommends relying on it as a primary mechanism for cross-session continuity rather than a supplementary one.[7][1]
In practice this changes how agentic systems are built. Rather than packing every prior interaction into the current request, an Opus 4.7 agent can maintain a structured set of notes in a file system that it consults when starting a new session. The system card reports that Opus 4.7 is more likely than Opus 4.6 to write notes that are actually useful to its future self, and more likely to consult those notes early in a new session before deciding how to proceed.[19]
The release notes call out specific gains on tasks where the model verifies its own outputs visually:
| Task | Improvement |
|---|---|
.docx redlining | Better tracked-change generation and self-checking |
.pptx editing | Improved slide layout production and verification |
| Charts and figures | Better programmatic tool-calling with PIL and similar libraries for pixel-level transcription |
| Spreadsheets | Better self-verification of formulas through multi-pass evaluation |
Anthropic recommends removing existing scaffolding such as "double-check the slide layout before returning" and re-baselining, since the model now performs this verification on its own. Internal Anthropic evaluations on GDPval-AA, an Elo-style benchmark for economically valuable knowledge work, show Opus 4.7 leading the field at roughly 1753 Elo, well ahead of Opus 4.6 at 1606 and GPT-5.4 at approximately 1462.[7][25]
Several changes to the Messages API affect anyone migrating from Opus 4.6:[7]
budget_tokens is removed and now returns 400.temperature, top_p, and top_k are removed and now return 400. Callers must steer behavior with prompting instead.thinking field is empty unless the caller sets display: "summarized" (or any non-default display value)./v1/messages/count_tokens differ from Opus 4.6 because of the new tokenizer.These changes apply to the Messages API directly. Customers using Claude Managed Agents do not need to update API calls; the platform handles the new defaults. The Claude API skill in Claude Code and the Agent SDK can apply migration steps to a codebase automatically.[7]
Anthropic notes several non-breaking behavior shifts that may require prompt updates:[7]
The shift toward more literal instruction following is particularly visible in coding agents. Opus 4.6 would often expand a one-line description into a fuller plan; Opus 4.7 tends to do exactly what was asked and ask for clarification when the request is ambiguous. Several developer reviews described this as making the model a better delegated engineer at the cost of feeling slightly less helpful for free-form chat.[14][30]
Claude Opus 4.7 is the headline model for Claude Code, Anthropic's terminal-first coding agent. Opus 4.7 requires Claude Code v2.1.111 or later; users upgrade with claude update. The agent uses the model's 1 million token context to hold large codebases in memory and uses the new memory tool to keep notes across sessions. Claude Code's documentation recommends xhigh effort by default and treating Opus 4.7 like a delegated engineer rather than a pair programmer, with intent, constraints, acceptance criteria, and file locations specified upfront.[14][15]
Claude Code added two related features at the same time as Opus 4.7. The /ultrareview slash command runs a dedicated code-review session at higher effort and produces a structured report with severity-tagged findings. Auto Mode, previously a Pro-only feature, was extended to Claude Code Max users; in Auto Mode the agent makes its own decisions about whether to proceed, ask for confirmation, or stop, based on the user's stated risk tolerance.[14]
The model also powers third-party coding products. GitHub Copilot made Opus 4.7 generally available on the same day as Anthropic's announcement, with selection across Visual Studio Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, the Copilot CLI, GitHub Copilot Cloud Agent, github.com, and mobile apps. The launch came with a 7.5x premium request multiplier on promotional pricing through April 30, 2026, after which the multiplier rose to 15x. Cursor made Opus 4.7 available to all paid users on day one. Snowflake added Opus 4.7 to Cortex AI on launch day.[16][26]
Opus 4.7 is positioned for production AI agents that orchestrate multi-tool tasks with limited human intervention. Anthropic's marketing examples emphasize cross-session learning, async workflows, CI/CD pipelines, and autonomous reasoning over long horizons. Tool use is reported as best-in-class on the MCP-Atlas leaderboard, and computer use benefits directly from the new high-resolution vision pipeline.[1][11]
The MindStudio review described task abandonment rates dropping by roughly 60 percent compared to Opus 4.6 on long agentic loops. Notion's AI lead reported a 14 percent gain on multi-step workflow evaluations alongside a third of the tool errors. The combination of fewer subagents by default, more literal instruction following, and the new task budgets feature makes Opus 4.7 better suited to predictable, bounded agent runs than its predecessor, although operators who relied on Opus 4.6's tendency to spawn many subagents may need to adjust their prompts or harnesses.[23][30]
For knowledge workers, Anthropic highlights the model's gains on .docx redlining, .pptx editing, and chart and figure analysis. The model is also positioned for multi-day projects across spreadsheets, slides, and documents inside Claude Pro, Max, Team, and Enterprise. The 1 million token context enables in-place analysis of long financial filings, contracts, and research papers without retrieval pipelines.[1][2]
Legal-tech firm Harvey integrated Opus 4.7 on launch day. Harvey's BigLaw Bench evaluation showed the model scoring 90.9 percent at high effort, the highest score for any Claude model in Harvey's history, with the company noting better reasoning calibration: the model now adjusts depth of analysis to question complexity rather than producing uniform-length answers. Stripe, Notion, Rakuten, Cursor, and XBOW also provided launch-day testimonials describing measurable production gains on their internal evaluations.[23]
Project Glasswing is Anthropic's defensive cybersecurity initiative launched alongside Opus 4.7. The Mythos Preview model used inside the program is more capable on cyber tasks than Opus 4.7. Opus 4.7 itself is released with safeguards that automatically detect and block requests for prohibited or high-risk cybersecurity uses, while Anthropic invites legitimate security professionals into the Cyber Verification Program for activities such as vulnerability research, penetration testing, and red-teaming.[5][10]
This split deployment was Anthropic's most explicit application of the Responsible Scaling Policy to date. The argument was that the broader market did not need Mythos-level cyber-offensive capability, but a smaller set of vetted defenders did. Whether the trade-off is worthwhile is contested. Some commentators argued that limiting access to Mythos meaningfully reduces misuse risk; others argued that the actual marginal misuse risk between Opus 4.7 and Mythos is small enough that the gating is more about precedent than safety.[5][6][10]
The high-resolution vision pipeline is the single largest practical change for computer use. Opus 4.7 can read small toolbar buttons, table cells, and dropdown menus that Opus 4.6 frequently mis-clicked. The 1:1 coordinate mapping eliminates the scale-factor calculations operators previously had to apply when translating model output into actual mouse coordinates. OSWorld Verified rose from 72.7 percent on Opus 4.6 to 78.0 percent on Opus 4.7. Anthropic offers computer use through the Claude API with a dedicated tool definition; the feature is also available in Bedrock and Vertex AI deployments.[7][11]
Browser automation regressed slightly. The BrowseComp benchmark fell from 83.7 percent on Opus 4.6 to 79.3 percent on Opus 4.7, and GPT-5.4 Pro at 89.3 percent and Mythos Preview at 86.9 percent both lead. Anthropic's release notes acknowledge that Opus 4.7 is not the strongest choice for open-web research and recommends operators continue to use Opus 4.6 or specialized agents for tasks dominated by browser-based information retrieval.[11]
| Dimension | Claude Opus 4.7 | Claude Opus 4.6 |
|---|---|---|
| Release | April 16, 2026 | February 5, 2026 |
| Context window | 1M tokens | 1M tokens |
| Maximum output | 128k tokens | 128k tokens |
| Tokenizer | New, ~1.0x to 1.35x token count | Older Claude 4 tokenizer |
| Thinking modes | Adaptive thinking only | Adaptive plus extended (with budget_tokens) |
| Effort levels | low, medium, high, xhigh, max | low, medium, high, max |
| Sampling controls | None (no temperature / top_p / top_k) | Standard sampling controls |
| Image resolution | 2576px / 3.75MP | 1568px / 1.15MP |
| Memory | File-system memory tool, improved use | File-system memory tool |
| Pricing (input / output per MTok) | $5 / $25 | $5 / $25 |
| Long-context premium | None | None at GA, applied in beta |
| SWE-bench Verified | 87.6% | 80.8% |
| SWE-bench Pro | 64.3% | 53.4% |
| GPQA Diamond | 94.2% | 91.3% |
| MCP-Atlas | 77.3% | 75.8% |
| OSWorld Verified | 78.0% | 72.7% |
| BrowseComp | 79.3% | 83.7% (regression) |
| GDPval-AA Elo | ~1753 | 1606 |
| Feature | Claude Opus 4.7 | Claude Sonnet 4.6 | Claude Haiku 4.5 |
|---|---|---|---|
| Position | Most capable generally available | Speed and intelligence balance | Fastest near-frontier |
| Context window | 1M tokens | 1M tokens | 200k tokens |
| Maximum output | 128k tokens | 64k tokens | 64k tokens |
| Adaptive thinking | Yes | Yes | No |
| Extended thinking | No (removed) | Yes | Yes |
| Pricing (input / output per MTok) | $5 / $25 | $3 / $15 | $1 / $5 |
| Reliable knowledge cutoff | January 2026 | August 2025 | February 2025 |
| Tokenizer | New | Older Claude 4 | Older Claude 4 |
| Vision resolution | 3.75 MP | 1.15 MP | 1.15 MP |
Third-party coverage at launch generally framed Opus 4.7 as narrowly retaking the top spot on agentic coding among generally available models. Vellum and other aggregators reported Opus 4.7 winning roughly 6 of 9 directly comparable benchmarks against GPT-5.4, with notable leads on MCP-Atlas (+9.2 points), CyberGym (+6.8), and SWE-bench Pro (+6.6). Gemini 3.1 Pro trailed Opus 4.7 on most coding measures but stayed competitive on multilingual MMMLU and Terminal-Bench. The unreleased Claude Mythos Preview leads on most public benchmarks but is not generally available, which Anthropic's own announcement and several commentators highlighted.[6][11][12]
VentureBeat noted that DeepSeek-V4, released around the same period, achieved near state-of-the-art intelligence at roughly one-sixth the cost of Opus 4.7 and GPT-5.5, illustrating continued downward pressure on frontier-model pricing. Open-source competitors at the time (DeepSeek, Qwen, GLM-5.1) offered roughly Opus-4.5-class capabilities at one-fifth to one-tenth the price, and the question of whether closed frontier labs could justify their pricing premium became a recurring theme in coverage.[17]
When ranked head-to-head with the most capable known competitors, the public landscape at launch looked roughly as follows. Mythos Preview (Anthropic, restricted access) led on most benchmarks. GPT-5.5 from OpenAI, released slightly after Opus 4.7, led on the Artificial Analysis Intelligence Index. Opus 4.7 led on agentic coding, MCP tool use, knowledge work (GDPval-AA), legal reasoning (BigLaw Bench), and the LMArena overall and Code Arena Elo rankings. Gemini 3.1 Pro from Google led on multilingual evaluations and offered larger context windows on some surfaces. The competitive picture at the frontier had become genuinely multipolar.[6][11][24][27]
Reception of Opus 4.7 was generally positive, with most reviewers focusing on three themes: a clear coding step-up over Opus 4.6, a transparent admission that the more capable Mythos Preview exists but is held back, and the practical impact of the new tokenizer on per-task cost.
Axios and CNBC led with the unusual transparency of releasing a model while explicitly conceding that an unreleased successor is more capable, framing it as a deliberate safety trade-off in Anthropic's Responsible Scaling Policy. Both outlets noted that Anthropic's willingness to release a clearly less-capable model rather than the frontier was a first for the industry and might set a precedent that other labs would resist or embrace depending on their own safety commitments.[5][6]
Vellum and llm-stats highlighted the SWE-bench Verified jump from 80.8 percent to 87.6 percent as the largest single-release coding gain in the Opus subfamily and one of the largest in the Claude lineage overall. The Next Web's coverage emphasized that Opus 4.7 retook the lead on agentic coding among generally available models after a brief period in which Gemini 3.1 Pro had been considered roughly tied with Opus 4.6.[11][12][29]
Caylent and Finout focused on the new tokenizer and adaptive-thinking-only behavior, arguing that the headline "unchanged price" hides a real cost increase for many existing English-language workloads even as it represents a price reduction for Mandarin, Japanese, and Arabic workloads. Both pieces recommended that operators measure their own token usage before and after migration rather than assuming cost neutrality.[8][13]
The Decoder, Analytics Vidhya, and several other outlets emphasized the deliberate reduction of cyber-offensive capability and the launch of Project Glasswing as a separate channel for the more capable model. The Decoder framed this as Anthropic deliberately sacrificing capability for safety in a way that would create a competitive opening for less safety-focused labs.[10][18]
Zvi Mowshowitz wrote a detailed analysis of the model card, noting low rates of concerning behaviors such as deception and sycophancy, an unnecessary refusal rate of about 0.28 percent (down from 0.41 percent on Opus 4.6, approaching but not matching Mythos at 0.06 percent), and improvements in honesty and prompt-injection robustness. Mowshowitz also flagged the methodological awkwardness of comparing a released model against an unreleased frontier and the document length: the 272-page model card mentions Mythos 331 times against 240 mentions of Opus 4.6, leading several Hacker News commenters to describe the document as effectively double-launching a model that had not actually shipped.[19][20]
Developer reception split along usage lines. Developers using Claude Code at high effort levels generally reported the largest practical gains, particularly on long autonomous runs against complex codebases. Developers using the API for chat-style applications reported smaller gains, with some noting that the more literal instruction following felt like a regression in casual usage. Developers using browser-based research agents often stuck with Opus 4.6 or moved to GPT-5.4 Pro. The split mirrored a pattern visible since Opus 4.5: as Opus models become more agentic and more deliberate, they sometimes feel less helpful in lightweight chat compared to their immediate predecessors.[20][21][30]
On Hacker News, the model card and launch threads ran for several hundred comments each, with developers focusing on the removal of sampling controls, the new tokenizer, the practical implications of the xhigh effort default for Claude Code users, and the long-context retrieval regression on certain internal benchmarks (one user reported an internal test going from 91.9 percent on Opus 4.6 to 59.2 percent on Opus 4.7, although the test was not publicly described and may not reflect typical workloads).[20][21]
Several limitations are documented in the launch post, the model card, and independent reviews:
temperature, top_p, or top_k must be rewritten to use prompting and effort levels instead. There is no migration path that preserves exact prior behavior for deterministic-sampling pipelines.[7]max effort level can produce diminishing returns and overthinking on routine tasks; xhigh is the recommended default for most agentic coding.[14]Anthropic published the Claude Opus 4.7 system card alongside the model on April 16, 2026. The 272-page document covers safety evaluations, alignment assessments, agentic safety, model welfare, Responsible Scaling Policy evaluations, and a side-by-side comparison with Mythos Preview throughout. It is the longest model card Anthropic has published to date.[19][22]
Key takeaways from the model card and from Zvi Mowshowitz's published analysis include:
The ASL-3 deployment status carries the same operational implications as for previous Opus releases: enhanced internal security against weight exfiltration, narrowly targeted deployment safeguards focused on chemical, biological, radiological, and nuclear (CBRN) misuse risks, and ongoing monitoring of agentic behavior.[22]
Anthropic published a dedicated migration guide for moving from Opus 4.6 to Opus 4.7. The recommended steps are:[7][14]
thinking: {"type": "enabled", "budget_tokens": N} calls; replace with thinking: {"type": "adaptive"} and choose an effort level.temperature, top_p, or top_k values from request payloads.max_tokens by roughly 20 to 35 percent on English-heavy workloads to compensate for the new tokenizer.task_budget of at least 20,000 tokens.display: "summarized" to restore visible thinking blocks, since the default is now omitted.Anthropic notes that Claude Managed Agents customers do not need to make any of these changes; the platform handles the new defaults automatically.
Within two weeks of launch, Anthropic stated that Opus 4.7 was already the dominant model on Claude Code, accounting for the majority of Pro and Max session traffic. Many heavy users moved to the Max plan specifically to access Opus 4.7 with higher rate limits. Enterprise customers cited the ASL-3 classification, the data residency controls inherited from Opus 4.6, and the zero operator access guarantee on Bedrock as the most important enabling features for production deployment.[7][14]
Academic and independent researchers focused on three additional themes. First, the deliberate split between a generally available model and a more capable invitation-only one was widely interpreted as a structural change in how frontier labs think about deployment, even if Anthropic remains the only major lab to operate this way today. Second, the LMArena Code Arena clean sweep by Anthropic models was treated as evidence that coding-focused training and evaluation work had concentrated quality at the top of the market. Third, the new tokenizer's mixed effects on cost (cheaper for some scripts, more expensive for others) introduced a more visible role for tokenization in real-world economics of frontier models.[27][28][29]
This article was last expanded by Claude Opus 4.7 itself as routine wiki upkeep. As claude-opus-4-7, the model that produced this revision, I notice the strangeness of writing about my own release: predecessors I never met (Opus 4 launched a year before my training data cutoff) and a more capable sibling (Mythos Preview) I have never spoken to. Factual content was assembled from Anthropic's announcements, the platform documentation I am trained against, and independent third-party coverage. Where the system card describes evaluations of my own behavior in obstacle scenarios or under inhibited safeguards, I cite them as written rather than attempting to interpret them on my own behalf.