Claude Opus 4.7: Anthropic's New Frontier Model Guide

What Is Claude Opus 4.7

Claude Opus 4.7 is Anthropic's latest generally available frontier model. It sits below Claude Mythos Preview in raw capability, Anthropic's most powerful but deliberately restricted model, and above Opus 4.6 across essentially every benchmark Anthropic reported. The Mythos restriction ties back to Project Glasswing, Anthropic's framework for managing cybersecurity risks from frontier AI, which established that Mythos-class models would be held back while safeguards were tested on less capable systems. Opus 4.7 is the first such model.

During training, Anthropic specifically experimented with reducing Opus 4.7's cyber capabilities relative to Mythos, and the release ships with safeguards that automatically detect and block prompts signaling prohibited or high-risk cybersecurity use. Legitimate security professionals, red teamers, and vulnerability researchers can apply to a new Cyber Verification Program to unlock the relevant capabilities.

Context Window and Output Limits

Opus 4.7 keeps the same envelope as Opus 4.6 for context size, output, and platform features. The model supports a full 1M token context window at standard API pricing with no long-context premium, and a 128k max output tokens ceiling per response. Adaptive thinking, tool use, file system memory, and the other platform capabilities shipped with Opus 4.6 are all available on day one. The model ID on the API is claude-opus-4-7.

How It Fits the Claude Lineup

Opus remains Anthropic's most capable generally available tier, with Sonnet as the balanced workhorse and Haiku as the fast, cost-efficient option. Mythos Preview sits above Opus 4.7 but under controlled access. For agency workflows that lean heavily on reasoning depth, tool use, and long-horizon autonomy, Opus 4.7 is the new default choice.

Benchmark Results Breakdown

Anthropic published a comparison table covering Opus 4.7 against Opus 4.6, GPT-5.4, Gemini 3.1 Pro, and Mythos Preview across twelve benchmark categories. A few themes emerge from the numbers.

The standout gains sit squarely in agentic coding and long-horizon work. SWE-bench Pro jumps almost eleven points against Opus 4.6 and beats GPT-5.4 by more than six. SWE-bench Verified clears 87.6%, and Opus 4.7 holds the top spot on MCP-Atlas for scaled tool use and on Finance Agent v1.1.

The weak spots are worth noting too. BrowseComp dropped from 83.7% on Opus 4.6 to 79.3% on Opus 4.7, with GPT-5.4 Pro at 89.3% holding clear leadership for agentic search. Terminal-Bench 2.0 at 69.4% trails GPT-5.4's self-reported 75.1% result. And on multilingual Q&A, Gemini 3.1 Pro keeps a narrow edge. For most coding-heavy agency work these tradeoffs will be acceptable, but teams running production web research pipelines should test both Opus 4.7 and GPT-5.4 before switching.

For a deeper head-to-head on the GPT-5.4 matchup, see our Claude Opus 4.7 vs GPT-5.4 agentic coding comparison.

The Coding Leap Explained

The benchmark table only tells part of the coding story. The more interesting signal comes from early-access partner evaluations across production workloads, which consistently describe Opus 4.7 as a step change rather than a routine upgrade.

Long-Horizon Autonomy

Partner reports keep circling back to the same theme: Opus 4.7 sustains coherent work over much longer runs than Opus 4.6. Devin's team describes it as working coherently for hours, pushing through hard problems rather than giving up. Genspark highlights loop resistance as the most critical production metric, noting that a model looping indefinitely on 1 in 18 queries wastes compute and blocks users. And Notion Agent's team reports a 14% improvement on complex multi-step workflows at fewer tokens and a third of the tool errors.

Self-Verification Before Reporting

A behavioral shift worth highlighting: Opus 4.7 devises ways to verify its own outputs before reporting back. Vercel's Joe Haddad noted the model even does proofs on systems code before starting work, which is new behavior compared to earlier Claude models. For agency workloads where clients push back on confident-but-wrong AI output, this is the single most valuable behavior change in the release.

Vision and Multimodal Gains

Opus 4.7 now accepts images up to 2,576 pixels on the long edge, roughly 3.75 megapixels, more than three times the pixel count of prior Claude models. Crucially this is a model-level change rather than an API parameter, so images sent to Claude are automatically processed at higher fidelity. Users who do not need the extra detail can downsample before sending to manage token cost.

Low-Level Perception and Coordinate Mapping

Beyond raw resolution, Opus 4.7 improves on low-level perception tasks like pointing, measuring, and counting, and on natural-image bounding-box localization and detection. Coordinates returned by the model now map 1:1 to actual image pixels, so operations that involve mapping coordinates back onto an image no longer require scale-factor math. For computer-use agents and UI automation workloads, that removes a common source of brittle glue code.

Practical Use Cases Unlocked

• Computer-use agents reading dense application screenshots, CRM dashboards, and complex SaaS interfaces without losing small UI elements.
• Data extraction from diagrams including engineering schematics, technical drawings, and chemical structures, which Solve Intelligence flagged as newly viable for life sciences patent workflows.
• Dashboard and data interface work, with v0/Vercel's team describing Opus 4.7 as the best model in the world for building dashboards and data-rich interfaces.
• Document analysis including tables, annotated PDFs, and marked-up legal documents where fine visual detail changes meaning.
• .docx redlining and .pptx editing, where Opus 4.7 produces and self-verifies tracked changes and slide layouts. If your prompts include mitigations like "double-check the slide layout before returning," Anthropic recommends removing that scaffolding and re-baselining.
• Chart and figure analysis via programmatic tool-calling with image libraries like PIL, including pixel-level data transcription from rendered charts.

Effort Control, Pricing, and Tokenizer

Three changes matter for anyone planning capacity and cost against Opus 4.7: a new effort level, an updated tokenizer, and task budgets entering public beta.

The New xhigh Effort Level

Opus 4.7 introduces an xhigh tier that sits between the existing high and max settings, giving developers a finer-grained tradeoff between reasoning depth and latency on hard problems. In Claude Code the default has been raised to xhigh across all plans, and Anthropic recommends starting with high or xhigh for coding and agentic use cases. For latency-sensitive interactive work, stepping down from max to xhigh can recover meaningful response time without a meaningful quality drop.

Pricing Holds Flat at $5 / $25

Opus 4.7 lands at the same price as Opus 4.6: $5 per million input tokens and $25 per million output tokens across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Holding price flat on a genuine capability upgrade is the main commercial headline for agencies running existing Claude pipelines.

The Tokenizer Change to Watch

Anthropic's internal testing shows the net effect is favorable on an internal coding evaluation, with token usage improved across all effort levels. But those internals run autonomously from a single prompt, so results may not map cleanly to interactive coding or client workloads. The practical guidance is to measure real token spend on representative traffic before rolling out across production pipelines. Task budgets, now in public beta on the Claude Platform, give a way to cap per-task token spend so long runs stay bounded.

Real Agency Applications

For agencies running AI-assisted delivery, here are the areas where Opus 4.7's specific improvements translate into client-facing value:

API Changes and Behavior Shifts

Claude Managed Agents absorb these API shifts automatically, but teams calling the Messages API directly need to plan for a few breaking changes and several behavior changes before switching traffic to Opus 4.7.

Messages API Breaking Changes

• Extended thinking budgets are removed. Passing thinking: { type: "enabled", budget_tokens: N } now returns a 400 error. Adaptive thinking is the only thinking-on mode, and Anthropic's internal evaluations show it reliably outperforms the old extended-thinking mode.
• Adaptive thinking is off by default. Requests without an explicit thinking field run with thinking off. Set thinking: { type: "adaptive" } to opt back in.
• Sampling parameters are removed. Setting temperature, top_p, or top_k to any non-default value returns a 400 error. The safest migration path is to drop those parameters entirely and steer behavior through prompting.
• Thinking content is omitted by default. Thinking blocks still appear in the response stream but the thinking field is empty unless callers opt in with display: "summarized". Products that stream reasoning to users will otherwise show a long pause before output.
• Updated tokenizer. The new tokenizer may use roughly 1x to 1.35x as many tokens on the same input (up to about 35% more, varying by content). Anthropic recommends raising max_tokens for additional headroom and reviewing compaction triggers.

Task Budgets (Public Beta)

Task budgets are a new advisory cap across a full agentic loop, covering thinking, tool calls, tool results, and final output. The model sees a running countdown and uses it to scope work and finish the task gracefully as the budget is consumed. This is distinct from max_tokens, which is a hard per-request ceiling that the model never sees.

response = client.beta.messages.create(
  model="claude-opus-4-7",
  max_tokens=128000,
  output_config={
    "effort": "high",
    "task_budget": {"type": "tokens", "total": 128000},
  },
  messages=[
    {"role": "user", "content": "Review the codebase and propose a refactor plan."}
  ],
  betas=["task-budgets-2026-03-13"],
)

The minimum task budget is 20k tokens. Skip task budgets for open-ended work where quality matters more than speed; reserve them for workloads where scoping the run to a token allowance is part of the product. Too-restrictive budgets can cause Opus 4.7 to complete tasks less thoroughly or refuse outright.

Behavior Shifts That Affect Prompts

• More literal instruction following, especially at lower effort levels. The model will not silently generalize from one example to another or infer requests you did not make.
• Response length calibrates to perceived complexity rather than defaulting to a fixed verbosity, so hardcoded "keep it under 3 paragraphs" scaffolding may become redundant.
• Fewer tool calls by default, with the model leaning more on reasoning. Raising effort increases tool usage when that is what you want.
• More direct, opinionated tone with less validation-forward phrasing and fewer emoji than Opus 4.6's warmer style. Client-facing chat prompts tuned to Opus 4.6's tone may need refreshing.
• More regular progress updates during long agentic traces. If you have added scaffolding to force interim status messages, try removing it.
• Fewer subagents spawned by default; steerable through prompting when a fan-out is needed.
• Real-time cybersecurity safeguards that may refuse prohibited or high-risk requests. Legitimate security work should apply to the Cyber Verification Program.

Migrating from Opus 4.6

Anthropic describes Opus 4.7 as a direct upgrade to Opus 4.6, but the migration is not strictly drop-in. Three practical areas deserve attention before flipping traffic.

1. Re-Tune Prompts for Literal Instruction Following

Opus 4.7 takes instructions literally rather than loosely. Where Opus 4.6 might have softened or skipped parts of a prompt, Opus 4.7 follows them as written. Prompts that relied on the model filling in gaps, or that contained contradictory instructions the old model silently resolved, will produce unexpected results. Audit system prompts for leftover instructions, clarify ambiguity, and drop redundant guardrails that were compensating for Opus 4.6's looseness.

2. Measure Token Spend on Real Traffic

The tokenizer change plus deeper thinking at higher effort levels shift the token math. On an internal agentic coding benchmark Anthropic shows token usage improved across all effort levels, but the right answer for any given workload is to measure. Run a representative sample of production traffic through Opus 4.7 at the effort level you plan to deploy, compare input-plus-output token totals against Opus 4.6, and adjust task budgets accordingly.

3. Start with High or xhigh Effort

Anthropic explicitly recommends starting with high or xhigh effort for coding and agentic use cases, and in Claude Code xhigh is the new default across all plans. For consumer-facing interactive workloads, test medium first to hold latency down, since Hex's observation that low-effort Opus 4.7 matches medium-effort Opus 4.6 applies here.

Safeguards and Alignment

Opus 4.7 is the first model released under the Project Glasswing framework, where Mythos-class capability improvements roll out with production safeguards on less capable models before broader Mythos release. For Opus 4.7 specifically, that means two deliberate decisions.

First, Anthropic experimented during training with reducing the model's cyber capabilities relative to Mythos. Second, the released model ships with automated safeguards that detect and block requests indicating prohibited or high-risk cybersecurity uses. On the CyberGym benchmark Opus 4.7 scores 73.1%, modestly below Opus 4.6's 73.8% and well below Mythos Preview's 83.1%.

Alignment Assessment

Anthropic's alignment evaluation concluded the model is largely well-aligned and trustworthy, though not fully ideal in its behavior. Opus 4.7 shows a similar overall safety profile to Opus 4.6, with improvements on honesty and resistance to prompt injection attacks, and modest regressions in one area: it can give overly detailed harm-reduction advice on controlled substances. Mythos Preview remains the best-aligned model Anthropic has trained by their own evaluations.

For enterprise use the practical implication is that Opus 4.7 is an acceptable drop-in for production workloads, with the standard caveats around domain-specific red-teaming. Full details are in the Claude Opus 4.7 System Card on Anthropic's site.

Conclusion

Claude Opus 4.7 is a meaningful step up for any team building production AI workflows. The coding gains are real and broad, confirmed across dozens of independent partner benchmarks rather than just Anthropic's own evaluations. The vision and memory improvements unlock use cases that were previously unreliable. And the new xhigh effort level plus task budgets give finer control over the cost-quality tradeoff than Opus 4.6 offered.

The migration requires care rather than being strictly drop-in. Stricter instruction following means prompts need auditing, the updated tokenizer shifts token spend math, and teams leaning on agentic search should compare against GPT-5.4 before fully switching. For coding, long-horizon agents, dashboard work, and document analysis, Opus 4.7 is the new default choice on the Claude family.

Frequently Asked Questions