Claude Opus 4.7 Review: Better Coding, 3x Vision, and Best-in-Class Tool Use

Anthropic continues its streak of powerful AI models that no competitor can come close too. Recently there was a lot of chatter about Mythos, Claude's most powerful unreleased model, which everyone was dubbing as a potential cybersecurity risk.

Claude Opus 4.7 looks like a controlled version of that.

Opus 4.7, released on April 16, 2026, is Anthropic's most capable generally available model to date. Where Opus 4.6 required close supervision on complex, long-running tasks, Opus 4.7 catches its own logical faults, verifies outputs before reporting back, and sustains coherent reasoning across hours-long agentic sessions.

This article covers what the model does, what the benchmarks mean in practice, how it compares to its predecessor, what it costs, and where it still falls short.

What Is Claude Opus 4.7?

Claude Opus 4.7 is Anthropic's most capable publicly available model, engineered for complex reasoning, long-horizon agentic work, vision-heavy workflows, and multi-session memory tasks.

Key specs at a glance:

• Model ID: claude-opus-4-7
• Context window: 1 million tokens
• Max output: 128k tokens
• Thinking: Adaptive (off by default, must be explicitly enabled)
• Available on: Claude.ai, Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, GitHub Copilot, Chatly

Within the Claude model family, Opus 4.7 sits above Sonnet and Haiku and is the highest-tier model available for general use. Claude Mythos Preview remains more broadly capable but is restricted to a limited set of enterprise partners.

For developers referencing the model in their stacks, Claude 4.7 Opus and claude-opus-4-7 both refer to the same generally available model. There are no Thinking, Pro, or Mini variants.

Key Capabilities and Benchmark Performance

As is the case with every Anthropic model ever, Opus 4.7 comes with significant improvements and newer features across multiple benchmarks.

1. Coding and Agentic Reliability

Coding is the headline story for Opus 4.7, and the benchmarks back it up across multiple dimensions.

Official benchmark scores vs. Opus 4.6:

• SWE-bench Verified: 87.6% vs. 80.8%
• SWE-bench Pro: 64.3% vs. 53.4% (ahead of GPT-5.4 at 57.7% and Gemini 3.1 Pro at 54.2%)
• CursorBench: 70% vs. 58%

The real-world partner results are just as significant:

• 13% resolution lift on a 93-task coding benchmark, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve
• 3x more production tasks resolved on Rakuten-SWE-Bench, with double-digit gains in code quality and test quality
• 14% lift over Opus 4.6 on complex multi-step workflows, at fewer tokens and a third of the tool errors
• First Claude model to pass implicit-need tests, meaning it handles ambiguous task specifications without asking for clarification at every turn

What separates these numbers from prior releases is the behavioral shift underneath them. Opus 4.7 catches its own logical faults during the planning phase and verifies outputs before reporting back. For teams comparing it against competing frontier models on coding and agentic tasks, GPT-5.4 narrows the gap on some benchmarks but trails on SWE-bench Pro and MCP-Atlas, the results that matter most for production agent workflows.

2. Vision and Multimodal Understanding

The vision upgrade in Opus 4.7 is a step-change that opens a category of use cases that were previously impractical.

Maximum image resolution increases from 1,568px at 1.15 megapixels to 2,576px at 3.75 megapixels, more than three times the pixel count of any prior Claude model. Beyond resolution, the model's coordinates now map 1:1 with actual pixels, eliminating the scale-factor conversion math that caused missed clicks and misread layouts in Opus 4.6 computer-use deployments.

The benchmark results explain this clearly:

• XBOW visual-acuity: 98.5% vs. 54.5% on Opus 4.6. Their single biggest Opus pain point, as they described it, effectively disappeared.
• OSWorld-Verified (computer use in a live OS): 78.0% vs. 72.7%, ahead of GPT-5.4 at 75.0%

In practice, dense screenshots, technical diagrams, chemical structures, complex UI mockups, and data-rich interfaces all come through at actual fidelity. Work that previously required downsampling workarounds now works cleanly.

3. Knowledge Work and Professional Output

Opus 4.7 is state-of-the-art on GDPval-AA, a third-party evaluation covering economically valuable knowledge work across finance, legal, and other professional domains. The capability improvements here are specific:

• Databricks' OfficeQA Pro reports 21% fewer errors than Opus 4.6 when working with source information
• BigLaw Bench score 90.9% at high effort, with notably better handling of ambiguous document editing tasks
• Better at producing and self-checking tracked changes in .docx files and slide layouts in .pptx files
• Improved programmatic chart and figure analysis, including pixel-level data transcription
• Correctly distinguishes document provisions that frontier models have historically confused, such as assignment provisions versus change-of-control provisions in legal contracts

For teams using Claude for financial analysis, slide creation, or document-heavy enterprise workflows, these are not marginal improvements. When the model can self-check its own output before returning it, the human review burden drops in a measurable way.

4. Memory and Long-Horizon Reasoning

Opus 4.7 is meaningfully better at using file system-based memory across long, multi-session work.

When an agent maintains a scratchpad or notes file between sessions, Opus 4.7 writes to it more reliably and uses those notes to reduce up-front context overhead on subsequent tasks. This reduces cold-start friction on multi-day projects, where earlier models effectively forgot what they had already done.

On a research-agent benchmark spanning six modules, Opus 4.7 tied for the top overall score of 0.715 and delivered the most consistent long-context performance of any model tested. Specific results worth noting:

• General Finance module (the largest): 0.813 vs. 0.767 on Opus 4.6
• Deductive logic: Solid performance in an area where Opus 4.6 had notable gaps
• Overall consistency: Top-ranked for sustained performance across long agentic traces, with fewer mid-task failures and better recovery when context grows long

Claude Opus 4.7 vs. Opus 4.6

The differences between the two models go beyond benchmark scores. There are behavioral shifts and API-level changes that affect how you work with the model day to day.

The three biggest capability deltas are:

• Instruction following
• Vision resolution
• Agentic loop reliability

Where Opus 4.6 interpreted instructions loosely and sometimes skipped steps, Opus 4.7 takes them literally. The 3x vision resolution upgrade is a model-level change, not a tuning adjustment. And the loop resistance improvements mean Opus 4.7 pushes through tool failures that stopped Opus 4.6 cold.

What's new in Opus 4.7 that did not exist in 4.6:

• xhigh effort level, sitting between high and max for finer reasoning-cost control
• Task budgets in public beta for managing token spend across agentic loops
• /ultrareview command in Claude Code, which functions like a skeptical senior engineer reviewing your changes

What was removed (breaking changes):

• Extended thinking budgets (budget_tokens) are gone. Adaptive thinking is now the only supported thinking-on mode.
• Sampling parameters temperature, top_p, and top_k have been removed. Setting any of them returns a 400 error.

The tone has shifted too. Opus 4.7 is more direct and opinionated, with fewer emoji and response length that calibrates to task complexity rather than defaulting to a fixed verbosity. These are not API breaking changes, but prompts written for Opus 4.6 that relied on its interpretive behavior may produce different results and are worth revisiting before deploying in production.

Migrating to Claude Opus 4.7

Opus 4.7 is not a drop-in replacement for Opus 4.6. Three changes require deliberate action before switching in production, and several behavioral shifts will affect outputs even without code changes.

1. Breaking Changes That Require Code Updates

Sampling parameters removed. temperature, top_p, and top_k return a 400 error if set to any non-default value. Any pipeline using these parameters needs code changes before migrating.

• Fix: Remove the parameters entirely. Use prompting to guide output style and variability instead.

Extended thinking budgets removed. thinking: {"type": "enabled", "budget_tokens": N} no longer works. Adaptive thinking is the only supported thinking-on mode.

• Fix: Switch to thinking: {"type": "adaptive"}. Adaptive thinking must be explicitly enabled — it is off by default.

Thinking content omitted by default. Thinking blocks appear in the response stream but the thinking field is empty unless explicitly configured.

• Fix: Add display: "summarized" if your product surfaces reasoning to users.

2. Tokenizer Change

Opus 4.7 uses a new tokenizer that processes the same text as up to 35% more tokens than Opus 4.6, varying by content type.

• Static token budgets built around 4.6 need re-measuring on 4.7
• Measure on your specific content type before assuming a fixed cost increase
• Use the xhigh effort level and task budgets to control spend on agentic runs

3. Behavioral Shifts That Affect Prompts Without Code Changes

These do not cause errors but will change outputs for any prompt relying on Opus 4.6's interpretive behavior:

• Literal instruction following: Soft language like "consider" or "you might" is now treated as optional. Replace with explicit directives.
• Response length calibration: Opus 4.7 scopes response length to perceived task complexity. Add explicit length or depth instructions if you need verbose outputs.
• Fewer subagents by default: Opus 4.7 spawns subagents less aggressively than Opus 4.6. If your workflow depends on automatic delegation, add explicit instructions for when to use subagents.
• Reduced emoji and warmer phrasing: Opus 4.7 defaults to a more direct tone. Specify tone explicitly if warmth or informality is required.

So you will need to design new system prompts for Opus 4.7 made with token consumption in mind and for efficient outputs.

What You Do Not Need to Change

Anthropic reports that Opus 4.7 performs well out of the box on most existing Opus 4.6 prompts. The migration effort concentrates on the breaking changes above. If your pipeline does not use sampling parameters or extended thinking budgets, the primary work is prompt review for behavioral alignment, not code changes.

All these factors make migration a major decision. But there are ways to use Opus 4.7 for free to test its limits before you fully commit.

Pricing

Pricing is unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens across the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

The practical consideration is the new tokenizer.

Opus 4.7 uses an updated tokenizer that can use roughly 1x to 1.35x as many tokens when processing text compared to previous models, with the variance depending on content type. The same prompt that cost a specific amount on Opus 4.6 may cost up to 35% more on Opus 4.7 at the token level, before any intelligence or efficiency gains are factored in.

For teams managing costs at scale, the effort parameter and task budgets are the primary controls. The xhigh effort level provides a meaningful sweet spot between the depth of max and the speed of high for most production workloads. The principles covered in Claude's token cost efficiency still apply to how you think about optimizing spend at this tier.

Strengths

Opus 4.7 earns its position as the most capable generally available Claude model across several dimensions. These are capabilities that translate directly into fewer engineering hours, more reliable production deployments, and workflows that did not work before.

Coding Autonomy on the Hardest Tasks

The SWE-bench Pro score of 64.3% does not tell the whole story. What matters operationally is that Opus 4.7 solves problems Opus 4.6 could not, including race conditions, concurrency bugs, and multi-file refactors that require sustained reasoning across a full codebase.

Multiple early-access partners explicitly noted solving tasks that no prior Claude model had passed.

Vision that Eliminates an Entire Class of Workarounds

Going from 54.5% to 98.5% on visual-acuity is not an improvement in the normal sense. It removes a ceiling. Computer-use agents, screenshot readers, technical diagram parsers, and life sciences document tools that were unreliable or impractical with Opus 4.6 are now viable at production scale.

Best-in-class Tool Use

Opus 4.7 leads MCP-Atlas at 77.3%, ahead of GPT-5.4 at 68.1% and Gemini 3.1 Pro at 73.9%. For orchestration agents that route across multiple tools in a single workflow, this is the benchmark that best reflects real production reliability.

Fewer tool errors, better loop resistance, and graceful error recovery reduce the maintenance overhead that makes long-running agents expensive to operate.

Instruction Precision

For enterprise teams running standardized workflows, a model that follows instructions literally and does not silently generalize from one item to another is more predictable and auditable. The tradeoff is that loose prompts need to be tightened, but the upside is workflows that behave consistently.

No Regressions across 20+ early-access partners

Multiple organizations explicitly called out Opus 4.7 as a clean upgrade. Deploying a model that improves one area while quietly degrading another costs more in engineering time than the capability gain is worth. Opus 4.7 does not create that problem.

Competitive Position Among Frontier Models

When comparing GLM 4.7 vs Claude Opus on coding and agentic benchmarks, Opus 4.7 holds clear advantages on SWE-bench Pro and MCP-Atlas tool use. The broader frontier model comparison shows Opus 4.7 leading on the benchmarks most relevant to production agent workflows, with the exception of BrowseComp, where GPT-5.4 holds the advantage for web-connected research tasks.

Limitations

Opus 4.7 is a strong upgrade, but is not without a few hurdles. Teams migrating from Opus 4.6 will hit real obstacles, and a few of them require deliberate engineering work before going to production.

Tokenizer Increases Effective Cost

Up to 35% more tokens on the same input content is a real cost increase at scale, even with pricing held flat. Teams processing high volumes of text need to measure this against their specific content types before committing.

Opus 4.6 Prompts May Behave Unexpectedly

Literal instruction following is a strength for purpose-built workflows. It is a friction point for anything that relied on Opus 4.6's interpretive behavior. Prompts that assumed the model would fill in gaps or generalize instructions may produce incomplete or unexpected outputs without modification.

Sampling Parameters are Gone

Any pipeline using temperature, top_p, or top_k now returns a 400 error. The migration path is to remove these parameters entirely and use prompting to guide model behavior, but this requires deliberate code changes in any affected integration.

Extended Thinking Budgets Require a Rebuild

Teams that are built around thinking: {"type": "enabled", "budget_tokens": N} need to migrate to adaptive thinking. The new approach outperforms the old one in internal evaluations, but it is a breaking change that requires migration work rather than a drop-in swap.

Cybersecurity Safeguards May Interrupt Legitimate Security Work

Opus 4.7 ships with automated detection and blocking of prohibited cybersecurity uses as part of Anthropic's Project Glasswing framework. Security researchers, penetration testers, and red teamers need to apply to the Cyber Verification Program before using the model for legitimate security work. Requests that trigger the safeguards mid-workflow will be blocked without notice otherwise.

Adaptive Thinking is off by Default

Setting thinking: {"type": "adaptive"} must be explicit. Teams that assume adaptive thinking is active without enabling it will not get the reasoning depth the model is capable of.

Try Claude Opus 4.7 on Chatly Today

Claude Opus 4.7 is a meaningful upgrade for any team doing complex coding, agentic work, or vision-heavy workflows. The benchmark scores are real, and the early-access feedback across more than twenty production companies confirms that the gains show up in actual work, not just controlled evaluations.

Opus 4.7 is the most opinionated Claude model yet. It follows instructions literally, verifies its own outputs, and carries long-running tasks through to completion with less hand-holding. That makes it more powerful for teams with well-defined workflows.

The tokenizer change and the removal of sampling parameters are the friction points worth planning around before upgrading. Everything else is an improvement.

Chatly gives you access to Claude Opus 4.7 alongside every other frontier model in one place, so you can test it against your actual workflows without committing to a full API migration first._