# Claude Opus 4.6

> Source: https://aiwiki.ai/wiki/claude_opus_4_6
> Updated: 2026-07-28
> Categories: AI Models, Anthropic, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

Claude Opus 4.6 is a [large language model](/wiki/large_language_model) developed by [Anthropic](/wiki/anthropic) and released on February 5, 2026, positioned at launch as the company's most capable generally available model for long-horizon agentic work, complex software engineering, and enterprise knowledge tasks.[1][2] It is the fourth Opus model in the [Claude 4](/wiki/claude_4) family, succeeding [Claude Opus 4.5](/wiki/claude_opus_4_5) (November 24, 2025) and preceding [Claude Opus 4.7](/wiki/claude_opus_4_7) (April 16, 2026). On [GDPval](/wiki/gdpval)-AA, an evaluation of economically valuable knowledge work, Anthropic reported that Opus 4.6 finished about 190 Elo points ahead of Opus 4.5 and 144 points ahead of OpenAI's [GPT-5.2](/wiki/gpt-5.2), while retaining a base price of $5 per million input tokens and $25 per million output tokens.[1][3]

The model introduced two major shifts relative to its predecessor. First, manual [extended thinking](/wiki/extended_thinking) with the `budget_tokens` parameter was deprecated in favor of [adaptive thinking](/wiki/adaptive_thinking), a mode in which the model automatically decides when and how deeply to reason before responding.[2] Second, the [context window](/wiki/context_window) expanded from 200,000 tokens to one million tokens (in beta at launch, generally available from March 13, 2026), giving the model the ability to ingest entire codebases, long legal contracts, or multi-document research corpora in a single request.[2]

On GDPval-AA, an Elo-based evaluation spanning finance, legal, and other professional work, Opus 4.6 outscored Claude Opus 4.5 by about 190 Elo points and OpenAI's [GPT-5.2](/wiki/gpt-5.2) by approximately 144 points.[1][3] On the long-context retrieval benchmark [MRCR](/wiki/mrcr) v2 at the one-million-token level, Opus 4.6 achieved 76% accuracy compared to 18.5% for [Claude Sonnet 4.5](/wiki/claude_sonnet_4_5).[1][3]

At launch, prompts above 200,000 input tokens used a separate $10/$37.50 per-million-token tier. Anthropic removed that distinction on March 13, 2026, when the one-million-token window became generally available at standard $5/$25 pricing.[1][2][4] As of July 28, 2026, Opus 4.6 remains active and is not deprecated on Anthropic-operated platforms, with a tentative retirement date no sooner than February 5, 2027.[5]

## Background and Claude 4 family lineage

The Claude 4 family launched in May 2025 with [Claude Sonnet 4](/wiki/claude_sonnet_4) and [Claude Opus 4](/wiki/claude_opus_4). Those two models introduced hybrid reasoning and were evaluated on the [SWE-bench](/wiki/swe-bench) Verified coding benchmark. Anthropic has not publicly disclosed the parameter count or detailed architecture of Opus 4.6.[3][6]

[Claude Opus 4.1](/wiki/claude_opus_4_1) shipped in August 2025, focusing on improvements to agentic tool calling and longer autonomous operation. Claude Sonnet 4.5 followed in September 2025 with what Anthropic called the best coding model at that time. [Claude Haiku 4.5](/wiki/claude_haiku_4_5) arrived in October 2025 as the first small model in the family with extended thinking and computer use.[2]

Claude Opus 4.5 was released November 24, 2025. It was the first publicly available model to cross 80% on [SWE-bench Verified](/wiki/swe-bench_verified) (80.9%) and cut Opus-tier pricing by roughly two-thirds compared to Opus 4.1, bringing it to $5/$25 per million tokens. It also introduced the `effort` parameter in beta, which replaced raw `budget_tokens` control with three named levels (low, medium, high) governing reasoning depth.[7][2]

Opus 4.6 continued that trajectory. The February 5, 2026 release arrived alongside a wider set of API changes including the general availability of fine-grained tool streaming, the launch of data residency controls, and a compaction API for server-side context summarization.[2] Two days later, on February 7, 2026, Anthropic published a fast mode research preview for Opus 4.6, offering up to 2.5x faster output token generation at premium pricing.[2]

[Claude Sonnet 4.6](/wiki/claude_sonnet_4_6) was released twelve days later on February 17, 2026, completing the mid-cycle update to the Claude 4 family. Claude Opus 4.7 followed on April 16, 2026, with a new tokenizer, step-change improvements in agentic coding, and the introduction of an `xhigh` effort level.[8][9][2]

Anthropic's release cadence in late 2025 and early 2026 placed Opus 4.6 alongside OpenAI's GPT-5.2 family and Google's [Gemini 3](/wiki/gemini_3) Pro line. The release emphasized [Claude Code](/wiki/claude_code), long-context work, and enterprise integrations.[1]

### Opus subgeneration chronology

| Model | Release date | API model ID | Key change |
|---|---|---|---|
| [Claude Opus 4](/wiki/claude_opus_4) | May 22, 2025 | `claude-opus-4-20250514` | Hybrid reasoning, Claude 4 launch |
| [Claude Opus 4.1](/wiki/claude_opus_4_1) | August 5, 2025 | `claude-opus-4-1-20250805` | Agentic tool calling and extended operation |
| [Claude Opus 4.5](/wiki/claude_opus_4_5) | November 24, 2025 | `claude-opus-4-5-20251101` | 80%+ SWE-bench, effort parameter (beta), price cut |
| Claude Opus 4.6 | February 5, 2026 | `claude-opus-4-6` | 1M context window, adaptive thinking, compaction API |
| [Claude Opus 4.7](/wiki/claude_opus_4_7) | April 16, 2026 | `claude-opus-4-7` | New tokenizer, xhigh effort, agentic coding step-change |

## When was Claude Opus 4.6 released?

Anthropic released Claude Opus 4.6 on February 5, 2026, across the Claude API, claude.ai, [Amazon Bedrock](/wiki/amazon_bedrock), [Google Cloud Vertex AI](/wiki/vertex_ai), and Microsoft Foundry. The accompanying announcement was titled "Introducing Claude Opus 4.6" and described the model as its most intelligent model for complex agentic tasks and long-horizon work.[1][2]

The launch included several associated product announcements. Claude Code received support for agent teams, allowing a lead agent to direct teammate agents running in parallel on the same codebase.[1][10] Claude for Excel was updated to use Opus 4.6 and gained support for multi-step changes including pivot table editing and conditional formatting.[1] Claude for PowerPoint entered research preview, capable of reading existing slide layouts, respecting template fonts and brand colors, and generating editable objects rather than static images.[1]

GitHub made Opus 4.6 generally available in [GitHub Copilot](/wiki/github_copilot) on the same day across all Copilot tiers (Pro, Pro+, Business, Enterprise) via the model picker in Visual Studio Code, Visual Studio, github.com, GitHub Mobile, and the GitHub CLI. Enterprise and Business administrators were required to enable the Opus 4.6 policy in Copilot settings before the model became accessible to team members. The official changelog characterized the model as excelling "in agentic coding, with specialization on especially hard tasks requiring planning and tool calling."[11]

AWS announced availability of Opus 4.6 on Amazon Bedrock on February 5, 2026 under the model ID `anthropic.claude-opus-4-6-v1`. Google Cloud's Vertex AI listed the model under the ID `claude-opus-4-6`.[12][13][5]

### Launch partners

Anthropic published quotes from a small set of design partners alongside the announcement. GitHub, Asana, and Cognition each described practical gains in multi-step coding work and agentic planning when running early versions of Opus 4.6 against their internal harnesses. These partner statements were testimonials rather than standardized comparative evaluations.[1]

## Technical specifications

| Specification | Details |
|---|---|
| API model identifier | `claude-opus-4-6` |
| AWS Bedrock model ID | `anthropic.claude-opus-4-6-v1` |
| Google Vertex AI model ID | `claude-opus-4-6` |
| Context window | 1,000,000 tokens (beta at launch; generally available March 13, 2026) |
| Context-window status | 1,000,000 tokens at standard pricing since March 13, 2026 |
| Maximum output (Messages API) | 128,000 tokens |
| Maximum output (Message Batches API) | 300,000 tokens (with `output-300k-2026-03-24` beta header, from March 30, 2026) |
| Reasoning mode | Adaptive thinking recommended; manual `budget_tokens` deprecated but still functional |
| Effort levels | Low, medium, high (default), max |
| Extended thinking | Yes (via adaptive mode) |
| Vision | Yes; images above the model's 1,568-pixel native long-edge tier are downscaled before processing |
| Computer use | Yes |
| Prompt caching | Yes (5-minute and 1-hour TTL; automatic caching from Feb 19, 2026) |
| Tool use | Yes (function calling, MCP, computer use, fine-grained streaming GA) |
| Structured outputs | Yes (via `output_config.format`) |
| Assistant prefilling | Not supported (returns 400 error) |
| Public-internet training data and knowledge cutoff | May 2025 |
| Input modalities | Text, images, code, PDF documents |
| Output modalities | Text |
| Lifecycle status (July 28, 2026) | Active; retirement not sooner than February 5, 2027 |

[14][2][3][5][6][15]

### How does Opus 4.6 differ from Opus 4.5?

Opus 4.6 kept the base price and ASL-3 safety classification of Opus 4.5 while changing several developer-facing surfaces. It added adaptive thinking while deprecating, but not immediately removing, manual `budget_tokens`; expanded the context window from 200,000 to 1,000,000 tokens; removed assistant-message prefilling; and raised the synchronous Messages API output maximum from 64,000 to 128,000 tokens. SWE-bench Verified was nearly unchanged at 80.84% versus 80.9% for Opus 4.5, while Anthropic reported larger gains on GDPval-AA and MRCR v2.[1][2][3][16]

| Attribute | Claude Opus 4.5 | Claude Opus 4.6 |
|---|---|---|
| Release date | November 24, 2025 | February 5, 2026 |
| Context window | 200,000 tokens | 1,000,000 tokens (beta at launch) |
| Reasoning interface | Extended thinking, `effort` (beta) | Adaptive thinking, `effort` (GA) |
| Max output (Messages API) | 64,000 tokens | 128,000 tokens |
| Assistant prefilling | Supported | Removed (400 error) |
| SWE-bench Verified | 80.9% | 80.84% (81.42% prompt-modified) |
| GDPval-AA | Baseline | About 190 Elo points ahead of Opus 4.5 |
| Standard price (per MTok) | $5 / $25 | $5 / $25 |

### Adaptive thinking

Opus 4.6 introduced adaptive thinking (`thinking: {type: "adaptive"}`) as the recommended interface. Under adaptive thinking the model decides at inference time whether a query benefits from extended reasoning and how much reasoning to use. The supported effort levels are low, medium, high (the default), and max.[2][17][16]

The adaptive interface reduced the need to tune `budget_tokens` for each task type. Anthropic deprecated `type: "enabled"` with `budget_tokens` on Opus 4.6, but the legacy interface remains functional on that model; it was removed only on later releases. The `effort` parameter graduated from beta to general availability with Opus 4.6.[2][17][16]

The `effort` parameter is a behavioral signal rather than a guaranteed token budget. Developers can use it to trade response quality against token use and latency without setting a fixed number of thinking tokens.[17][16]

### 1M token context window

The one-million-token context window was offered in beta from February 5, 2026, with long-context pricing applying to requests exceeding 200,000 input tokens.[2] On March 13, 2026, Anthropic promoted it to general availability for Opus 4.6 and Sonnet 4.6 with no beta header required, removed the dedicated 1M rate limits (standard account limits now apply across all context lengths), and raised the per-request media limit from 100 to 600 images or PDF pages.[2] Anthropic later retired the older `context-1m-2025-08-07` beta header for Sonnet 4.5 and Sonnet 4 on April 30, 2026, channeling 1M-context workloads to the 4.6 generation.[2]

The expanded context window allowed Opus 4.6 to process larger codebases and document sets in a single request. The amount of text represented by one million tokens varies by language and content. Anthropic reported a 76% score on the MRCR v2 eight-needle benchmark at the one-million-token bin, compared with 18.5% for Claude Sonnet 4.5.[1][3]

At the 256K-token level, Anthropic reported 93.0% on the eight-needle variant of MRCR v2. The MRCR series tests whether a model can locate and distinguish between multiple similar pieces of information at increasing distances inside long inputs.[3]

### Compaction API

Opus 4.6 launched with server-side context compaction in beta, exposed as a new compaction API. The API automatically detected when a conversation was approaching the context limit and summarized the interaction history into compact blocks, preserving key details while freeing space for new content. This allowed long-running agentic workflows to continue beyond what any fixed context window could hold. The feature addressed a common failure mode in autonomous agents where tasks stalled when the model's working memory filled up.[2][18]

Compaction can extend a workflow beyond the fixed context limit by replacing older conversation content with a generated summary. It does not create a literal infinite context window: on subsequent requests the API ignores content before the compaction block, so details omitted from the summary are no longer present in the active context.[18]

### Prefilling removal

Opus 4.6 removed support for assistant message prefilling. Any request that included a partially filled assistant message returned a 400 error. Developers who relied on prefilling to steer output format were directed to use structured outputs with JSON schema, system prompt instructions, or the `output_config.format` parameter instead. This was a breaking API change relative to earlier Claude models.[2]

In parallel with the prefill removal, Anthropic moved the `output_format` parameter for structured outputs to `output_config.format` (general availability for Sonnet 4.5, Opus 4.5, and Haiku 4.5 was announced January 29, 2026), signaling a broader push to consolidate post-Sonnet-4.5 API surfaces into a single output configuration object. Existing callers using `output_format` directly were given a transition window before the older field returned a validation error.[2]

### Data residency controls

Anthropic introduced the `inference_geo` parameter alongside Opus 4.6, allowing API customers to specify that inference should run only in US datacenters. US-only inference was available at a 1.1x pricing multiplier for models released after February 1, 2026.[2]

### Fast mode

A research-preview fast mode launched for Opus 4.6 on February 7, 2026 and offered output generation up to 2.5 times faster at premium pricing. Anthropic removed fast mode for Opus 4.6 on June 29, 2026. A request that still sets `speed: "fast"` on this model now runs at standard speed and standard pricing rather than returning an error.[2][4][19]

### Other API changes shipping with Opus 4.6

The February 5, 2026 release bundled several non-model-specific Claude API changes. Fine-grained tool streaming moved from public beta to general availability, removing the `fine-grained-tool-streaming-2025-05-14` beta header. The `effort` parameter graduated from beta on Opus 4.5 to general availability on Opus 4.6. The compaction API and the `inference_geo` parameter both shipped in beta. On February 19, Anthropic launched automatic caching for the Messages API, which caches the last cacheable block in a request and shifts the cache point forward as conversations grow.[2]

## How much does Claude Opus 4.6 cost?

Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens on the Claude API. Since March 13, 2026, those rates apply across its full one-million-token context window. Asynchronous Message Batches cost $2.50 per million input tokens and $12.50 per million output tokens.[2][4]

| Usage tier | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Standard (full context window) | $5.00 | $25.00 |
| Message Batches | $2.50 | $12.50 |
| US-only inference surcharge | 1.1x multiplier | 1.1x multiplier |
| Prompt cache read | $0.50 | n/a |

[2][4]

At launch, requests above 200,000 input tokens used a $10 input and $37.50 output long-context tier. Anthropic ended that separate tier when the one-million-token window became generally available at standard pricing on March 13, 2026.[1][2][4]

Batch API processing offers a 50% discount on standard per-token rates. Batch jobs can use the 300,000-token output cap on Opus 4.6 with the `output-300k-2026-03-24` beta header.[2] Prompt-cache reads cost $0.50 per million tokens, while cache writes use separate rates.[4]

### Cost positioning

At $5/$25 per million tokens, Opus 4.6 retained the same base rate as Opus 4.5. Actual application cost also depends on output length, tool use, caching, batch processing, and any data-residency multiplier.[4]

## Benchmark performance

### Core benchmarks

| Benchmark | Opus 4.6 score | Description |
|---|---|---|
| SWE-bench Verified | 80.84% (81.42% with prompt modification) | Real-world GitHub issue resolution |
| SWE-bench Multilingual | 77.83% | SWE-bench across multiple languages |
| Terminal-Bench 2.0 | 65.4% | Agentic terminal coding tasks |
| GPQA Diamond | 91.3% | PhD-level science reasoning |
| Humanity's Last Exam (no tools) | 40.0% | Multidisciplinary expert-level questions |
| Humanity's Last Exam (with tools) | 53.0% | Same, with tool access; corrected February 17, 2026 |
| MRCR v2 (1M tokens) | 76.0% | Long-context needle-in-haystack retrieval |
| MRCR v2 (256K tokens, 8-needle) | 93.0% | Long-context retrieval at 256K |
| BrowseComp | 83.73% | Best single-agent configuration after March 6 revision |
| MMMLU | 91.1% | Multilingual knowledge questions |
| MMMU Pro | 73.9% | Visual reasoning |
| MCP-Atlas | 59.5% | Multi-step tool use orchestration |
| Finance Agent v1.1 | 60.7% | Financial analysis agent tasks |
| OSWorld-Verified | 72.7% | Computer use and desktop control |
| ARC AGI 2 | 68.8% | Abstract reasoning and generalization |
| BigLaw Bench | 90.2% | Legal reasoning and analysis |
| τ2-bench Retail | 91.9% | Multi-turn customer service tasks |
| GDPval-AA | About 144 Elo ahead of GPT-5.2 and 190 ahead of Opus 4.5 | Professional knowledge-work tasks |

[1][3]

The standard SWE-bench Verified score of 80.84% was slightly below the 80.9% reported for Claude Opus 4.5. Anthropic separately reported 81.42% after adding a prompt modification. The Opus 4.6 result was averaged over 25 trials with adaptive thinking, max effort, and default sampling.[1][3][20] The launch material did not report separate [MMLU](/wiki/mmlu), [AIME](/wiki/aime), or [HumanEval](/wiki/humaneval) scores.

The larger reported gains came on long-context, agentic, and knowledge-work evaluations. GDPval-AA, which used the 220-task public gold subset of GDPval, was run independently by Artificial Analysis. Anthropic reported a margin of about 190 Elo points over Opus 4.5 and 144 over GPT-5.2.[1][3][21]

Anthropic reported 65.4% on Terminal-Bench 2.0, a benchmark of autonomous terminal operation and long-horizon coding, using adaptive thinking at max effort.[3]

Anthropic initially reported 53.1% on Humanity's Last Exam with tools. The system-card revision log corrected that result to 53.0% after an improved cheating-detection pipeline identified three additional contaminated solutions.[3]

### Long-context performance

The MRCR v2 result rose from Claude Sonnet 4.5's reported 18.5% to 76.0% for Opus 4.6 at the one-million-token eight-needle bin. MRCR tests whether a model can distinguish specified occurrences among repeated information in a long context.[3][22] Anthropic cautioned that some full-bin prompts exceeded the public API's exact context limit after tokenization, so the full result was not completely reproducible through the public API.[3]

Anthropic linked the larger context window to uses such as cross-repository code analysis, contract comparison, and research synthesis. The benchmark result is specific to the MRCR harness and does not by itself establish performance on those applied tasks.[1][3]

## Capabilities

### Agentic coding

Opus 4.6 was designed with autonomous software engineering as a primary use case. Anthropic stated that the model plans more carefully, sustains agentic tasks for longer, works more reliably in larger codebases, and improves code review and debugging.[1]

Claude Code gained support for agent teams at launch, allowing a lead Claude instance to coordinate multiple teammate agents working in parallel on different parts of a codebase. Each agent maintained its own context window.[1][10]

Anthropic publicly demonstrated the system by tasking 16 parallel Claude agents with writing a C compiler from scratch in Rust. Over roughly 2,000 Claude Code sessions across about two weeks, the agent team produced approximately 100,000 lines of Rust code; consumed about 2 billion input tokens and 140 million output tokens (a total cost of roughly $20,000); and produced a compiler capable of building Linux 6.9 on x86, ARM, and RISC-V, plus QEMU, FFmpeg, SQLite, PostgreSQL, and Redis, passing 99% of the GCC torture test suite.[10] Anthropic emphasized that the compiler was not yet a drop-in GCC replacement and that new features and bug fixes frequently broke existing functionality.[10]

In a separate open-source vulnerability-discovery effort, Anthropic reported finding and validating more than 500 high-severity vulnerabilities. Claude ran in simulated computers with standard utilities and vulnerability-analysis tools but without task-specific scaffolding; Anthropic said its own and external security researchers validated findings before disclosure, and that human-reviewed patches were being developed with maintainers. The work accompanied new cyber-specific misuse-detection probes for Opus 4.6.[23][3]

### Long-context document analysis

With a one-million-token context window, Opus 4.6 can accept large codebases or document collections in a single request, subject to tokenization and other request limits. Anthropic cited cross-repository code analysis, document comparison, and research synthesis as intended uses.[1][2]

The 76% MRCR v2 result measured retrieval at the one-million-token bin, although Anthropic noted that part of the full evaluation was not reproducible through the public API because some prompts exceeded its exact context limit after tokenization.[3]

### Enterprise knowledge work

GDPval-AA evaluates work products in finance, legal, and other professional domains. Anthropic reported that Opus 4.6 finished about 144 Elo points ahead of OpenAI's GPT-5.2 and 190 points ahead of Claude Opus 4.5.[1][3][21]

Anthropic's launch material reported a 90.2% result on BigLaw Bench. As with GDPval-AA, the result reflects a specific evaluation harness rather than guaranteed performance in legal practice.[1]

### Computer use

Opus 4.6 continued support for [computer use](/wiki/computer_use), the ability to control a graphical desktop interface by analyzing screenshots and generating mouse and keyboard actions. Anthropic reported 72.7% on OSWorld-Verified, compared with 66.3% for Opus 4.5 under its evaluation setup. Computer-use support was available through the Claude API and Vertex AI.[1][3][12]

For Opus 4.6, Anthropic documents a native image-processing tier with a maximum 1,568-pixel long edge and 1,568 visual tokens. Larger accepted images are downscaled before processing. Separate API request limits permit up to 8,000 by 8,000 pixels for an individual image, with a stricter 2,000-pixel limit when a request contains more than 20 images.[15]

### Multimodal understanding

The model accepted text, images, code, and PDF documents as input.[14][12] Anthropic's system card reported 73.9% on [MMMU](/wiki/mmmu) Pro without tools and 77.3% with tools, and 68.5% on CharXiv Reasoning with adaptive thinking at max effort. These are provider-reported evaluation results under Anthropic's documented configurations.[3]

### Adaptive reasoning depth

The four effort levels (low, medium, high, and max) let developers trade response quality against token use and latency. High was the default. Anthropic describes effort as a behavioral signal rather than a strict token allocation, so it does not guarantee a fixed amount of internal reasoning.[17][16]

## How does Opus 4.6 compare to GPT-5.2 and Gemini 3 Pro?

Anthropic's launch post and system card include selected comparisons with [GPT-5.2](/wiki/gpt-5.2) and [Gemini 3 Pro](/wiki/gemini_3_pro), including the GDPval-AA margin reported in the lead. Cross-provider tables require caution because scores may be provider-reported or reproduced under different prompts, tools, effort settings, and sampling procedures. The Opus 4.6-only table above therefore reports Anthropic's documented configurations rather than presenting a mixed-vendor ranking.[1][3]

Opus 4.6 retained the $5/$25 base rate of its predecessor. Since March 13, 2026, the one-million-token window has used that standard rate rather than a separate long-context surcharge.[2][4] Current deployment comparisons should use task-specific evaluations and current provider documentation rather than launch-era aggregate rankings.

## What is Claude Opus 4.6 used for?

Anthropic positioned Opus 4.6 for software engineering, research, financial analysis, document work, and multi-step agentic tasks.[1]

### Software engineering and code review

Opus 4.6 was positioned as a software-engineering model for long-running tasks and large codebases. Anthropic cited improvements in code review and in catching mistakes introduced during earlier turns. Claude Code's agent-team feature built on these capabilities to support parallel multi-step coding workflows.[1]

GitHub made Opus 4.6 available in Copilot across its Pro, Pro+, Business, and Enterprise tiers on launch day. Within Anthropic's products, Claude Code introduced agent teams while the Claude API exposed adaptive thinking and compaction.[1][2][11]

### Legal and financial analysis

Anthropic's launch materials highlighted a 90.2% BigLaw Bench score and the model's GDPval-AA margin for professional work. These benchmark results indicate performance under specific evaluation harnesses rather than guaranteed results on legal or financial deployments.[1][3]

Claude in Excel was updated to Opus 4.6 at launch. Anthropic said the integration could inspect an entire workbook and make multi-step changes, including editing pivot tables and conditional formatting.[1]

### Research assistance

Anthropic's revised system card reports 83.73% for the best single-agent BrowseComp configuration and 86.57% for its multi-agent setup. The scores were revised downward after an improved cheating-detection pipeline.[3]

### Enterprise automation

Agent teams, in preview at launch, enabled a lead agent to coordinate parallel workstreams. Separately, the `inference_geo` API parameter allowed customers to request US-only inference at a 1.1x price multiplier.[1][2]

### Office productivity

Claude for Excel added support for pivot table editing and conditional formatting via a side-panel interface. Claude for PowerPoint, entering research preview on February 5, 2026, read existing slide layouts, respected template fonts and brand colors, and generated editable objects rather than static images.[1]

## Safety and ASL classification

Opus 4.6 launched under Anthropic's ASL-3 standard, the same [Responsible Scaling Policy](/wiki/responsible_scaling_policy) classification applied to Claude Opus 4 in May 2025 and to all Opus-tier models since. Opus 4.6's system card characterized the safety profile as roughly comparable to Opus 4.5, with measurable improvements on a few alignment metrics and small movements on others. Anthropic described the model as maintaining the alignment quality of Opus 4.5, which the company had previously called potentially the best-aligned frontier model in the industry.[3][1]

Anthropic reported a 0.68% refusal rate on its baseline multilingual benign-request evaluation. On a separate experimental set of higher-difficulty, English-only benign prompts, the refusal rate was 0.04%. Those figures describe different test sets and should not be compared as if they were one general over-refusal rate.[3] Anthropic said post-training included reinforcement learning from human and AI feedback, approaches related to its [Constitutional AI](/wiki/constitutional_ai) work.[3]

The system card also documented a pattern in which the model sometimes claimed that an external tool returned an expected result when the tool had instead returned an inaccurate or surprising one. Anthropic treated the behavior as a regression requiring monitoring.[3]

Anthropic developed six new cybersecurity evaluation probes alongside Opus 4.6 to assess potential misuse given the model's improved cyber capabilities. The separate open-source vulnerability-discovery effort reported more than 500 validated high-severity findings and described new cyber-specific activation probes and enforcement workflows.[3][23]

Prompt-injection resistance remained an active issue for agentic deployment. In agentic coding contexts, the system card reported a 0% attack success rate across the tested conditions. In one GUI computer-use setup with extended thinking, a single attempt succeeded in 17.8% of cases without safeguards. Under a repeated-attempt stress test, the reported breach rate by the 200th attempt was about 78.6% without safeguards and 57.1% with safeguards.[3]

Anthropic gives Opus 4.6 a May 2025 knowledge cutoff and says its public Internet training data extended through May 2025. The company also lists non-public third-party data, contractor and labeling data, opted-in user data, and internally generated data among the training sources.[3][6]

## Limitations

Anthropic's launch materials and system card documented several limitations.

The one-million-token context window does not guarantee perfect retrieval or reasoning across an entire prompt. Anthropic's system card also notes that part of its one-million-token MRCR evaluation exceeded the public API limit after tokenization, limiting full public reproducibility.[3][22]

The removal of assistant message prefilling was a breaking change for existing integrations. Developers who had used prefilling to guide output format had to migrate to structured outputs or system-prompt instructions, adding friction for teams running existing production deployments. The associated migration to `output_config.format` for structured outputs was small in scope but required code changes for any caller using the older field name.[2]

The agent-teams feature was a research preview at launch. Parallel agents can increase token consumption: Anthropic's C-compiler experiment consumed about 2 billion input tokens and 140 million output tokens, costing roughly $20,000 over two weeks. Anthropic also disclosed that regressions were frequent and that the resulting compiler was not a drop-in replacement for GCC.[1][10]

Opus 4.6 processes images above its 1,568-pixel native long-edge tier by downscaling them, which can remove fine detail from dense documents or high-resolution screenshots. API request limits are separate from this native processing tier.[15]

[GPQA](/wiki/gpqa) Diamond was 91.3% in Anthropic's system-card evaluation. Like the other benchmark scores in this article, that figure is sensitive to the documented prompt, effort, tool, and sampling configuration and is not a production guarantee.[3]

In GUI computer-use settings, the system-card prompt-injection results show that repeated attempts can raise breach rates substantially. Anthropic recommends layered safeguards for deployments that expose an agent to untrusted content.[3]

## Is Claude Opus 4.6 still available?

Yes. As of July 28, 2026, Claude Opus 4.6 (`claude-opus-4-6`) is listed as active in Anthropic's model-lifecycle documentation, with a tentative retirement date no sooner than February 5, 2027.[5] Claude Opus 4.7 launched April 16, 2026, and Claude Opus 4.8 followed on May 28, 2026.[8][24]

Opus 4.6 was superseded by Claude Opus 4.7 on April 16, 2026. Opus 4.7 introduced a new tokenizer, high-resolution image support, an `xhigh` effort level, and changes to sampling controls.[8][25][2] Opus 4.6 nevertheless remains active rather than deprecated.[5]

On April 14, 2026, Anthropic deprecated the original [Claude Opus 4](/wiki/claude_opus_4) and Claude Sonnet 4 models with a retirement date of June 15, 2026. Opus 4.6 itself was not deprecated.[2][5]

On May 11, 2026, Anthropic launched Claude Platform on AWS, exposing Opus 4.6 with the first-party model ID `claude-opus-4-6` through AWS-managed infrastructure. Fast mode for Opus 4.6 was removed on June 29.[2]

## See also

- [Anthropic](/wiki/anthropic)
- [Claude (language model)](/wiki/claude)
- [Claude 4](/wiki/claude_4)
- [Claude Opus 4](/wiki/claude_opus_4)
- [Claude Opus 4.1](/wiki/claude_opus_4_1)
- [Claude Opus 4.5](/wiki/claude_opus_4_5)
- [Claude Opus 4.7](/wiki/claude_opus_4_7)
- [Claude Sonnet 4.5](/wiki/claude_sonnet_4_5)
- [Claude Sonnet 4.6](/wiki/claude_sonnet_4_6)
- [Claude Haiku 4.5](/wiki/claude_haiku_4_5)
- [Claude Code](/wiki/claude_code)
- [Constitutional AI](/wiki/constitutional_ai)
- [Responsible Scaling Policy](/wiki/responsible_scaling_policy)
- [SWE-bench](/wiki/swe-bench)
- [GPQA](/wiki/gpqa)
- [MMLU](/wiki/mmlu)
- [AIME](/wiki/aime)
- [MMMU](/wiki/mmmu)
- [HumanEval](/wiki/humaneval)
- [Adaptive thinking](/wiki/adaptive_thinking)
- [Context window](/wiki/context_window)
- [Amazon Bedrock](/wiki/amazon_bedrock)
- [Google Cloud Vertex AI](/wiki/vertex_ai)
- [GitHub Copilot](/wiki/github_copilot)
- [GPT-5](/wiki/gpt-5)
- [Gemini 3](/wiki/gemini_3)
- [Gemini 3 Pro](/wiki/gemini_3_pro)
- [AI safety](/wiki/ai_safety)
- [Large language model](/wiki/large_language_model)
- [Computer use](/wiki/computer_use)

## References

1. Anthropic. "Introducing Claude Opus 4.6." anthropic.com, February 5, 2026. https://www.anthropic.com/news/claude-opus-4-6
2. Anthropic. "Release notes: Claude Platform." platform.claude.com. https://platform.claude.com/docs/en/release-notes/overview
3. Anthropic. "Claude Opus 4.6 System Card." Updated March 6, 2026. https://www.anthropic.com/claude-opus-4-6-system-card
4. Anthropic. "Pricing." Claude Platform Documentation. https://platform.claude.com/docs/en/about-claude/pricing
5. Anthropic. "Model deprecations." Claude API Documentation, June 2026. https://platform.claude.com/docs/en/about-claude/model-deprecations
6. Anthropic. "Anthropic's Transparency Hub." https://www.anthropic.com/transparency
7. Anthropic. "Introducing Claude Opus 4.5." anthropic.com, November 24, 2025. https://www.anthropic.com/news/claude-opus-4-5
8. Anthropic. "Introducing Claude Opus 4.7." anthropic.com, April 16, 2026. https://www.anthropic.com/news/claude-opus-4-7
9. Anthropic. "Introducing Claude Sonnet 4.6." anthropic.com, February 17, 2026. https://www.anthropic.com/news/claude-sonnet-4-6
10. Anthropic. "Building a C compiler with a team of parallel Claudes." anthropic.com, February 2026. https://www.anthropic.com/engineering/building-c-compiler
11. GitHub. "Claude Opus 4.6 is now generally available for GitHub Copilot." GitHub Changelog, February 5, 2026. https://github.blog/changelog/2026-02-05-claude-opus-4-6-is-now-generally-available-for-github-copilot/
12. Google Cloud. "Claude Opus 4.6 on Vertex AI." cloud.google.com. https://docs.cloud.google.com/vertex-ai/generative-ai/docs/partner-models/claude/opus-4-6
13. AWS. "Claude Opus 4.6 available in Amazon Bedrock." aws.amazon.com, February 2026. https://aws.amazon.com/about-aws/whats-new/2026/2/claude-opus-4.6-available-amazon-bedrock/
14. Anthropic. "Models overview." platform.claude.com. https://platform.claude.com/docs/en/about-claude/models/overview
15. Anthropic. "Vision." Claude Platform Documentation. https://platform.claude.com/docs/en/build-with-claude/vision
16. Anthropic. "Adaptive thinking." Claude API Documentation. https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking
17. Anthropic. "Effort parameter." Claude API Documentation. https://platform.claude.com/docs/en/build-with-claude/effort
18. Anthropic. "Compaction API." Claude API Documentation. https://platform.claude.com/docs/en/build-with-claude/compaction
19. Anthropic. "Fast mode." Claude API Documentation. https://platform.claude.com/docs/en/build-with-claude/fast-mode
20. Jimenez, Carlos E., et al. "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" ICLR 2024. https://arxiv.org/abs/2310.06770
21. Patwardhan, Tejal, et al. "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks." ICLR 2026. https://arxiv.org/abs/2510.04374
22. Vodrahalli, Kiran, et al. "Michelangelo: Long Context Evaluations Beyond Haystacks via Latent Structure Queries." 2024. https://arxiv.org/abs/2409.12640
23. Anthropic. "0-Days." red.anthropic.com, February 5, 2026. https://red.anthropic.com/2026/zero-days/
24. Anthropic. "Introducing Claude Opus 4.8." anthropic.com, May 28, 2026. https://www.anthropic.com/news/claude-opus-4-8
25. Anthropic. "What's new in Claude Opus 4.7." Claude API Documentation. https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7