Most of the coverage around Claude Opus 4.7 reads like a press release wearing a byline. SWE-bench is up. Vision resolution tripled. Coding is best-in-class. All of that is true. None of it is the full picture.
Anthropic's own system card — 90-plus pages that almost no one actually reads — contains admissions that never make it into the headline benchmarks. There's a regression they kept in "for scientific honesty." There's an intentional capability suppression. There's an adaptive thinking mechanism that will decide, without asking you, that your task isn't worth thinking hard about. And there's a compute constraint that sits underneath all of it, quietly shaping every decision Anthropic is making right now.
This piece covers all of it. The wins are real. So are the caveats.
Table of Contents
- The Adaptive Thinking Problem Nobody Is Talking About
- What the Benchmarks Actually Say
- The Regressions Anthropic Admits To
- Vision: Genuinely Better, But Not Everywhere
- The Compute Problem Under Everything
- Claude Code Upgrades Worth Knowing
- The Mythos 4x Productivity Claim, Dissected
- The Dario-Brockman Rivalry and What It Means for the Models
- My Take
- Key Takeaways
- FAQ
The Adaptive Thinking Problem Nobody Is Talking About
Thesis: Adaptive thinking is framed as a feature. For certain workflows, it behaves like a liability.
Claude Opus 4.7 now thinks adaptively. The model assesses how complex it believes your task is, and scales its reasoning depth accordingly. Easy-looking task? Less thinking time. Complex task? More. Sounds sensible in theory.
The problem is the model's difficulty calibration isn't yours. Simple Bench — a benchmark built around trick questions designed to look easy but requiring genuine common sense to see through — exposed this directly. Because Opus 4.7 seems to think these questions are easier than they actually are, it scores worse than Opus 4.6. :antCitation[]{citations="82be3f2e-7f2b-4983-880b-945a87e7e227"} It's not that the model lacks the reasoning capacity. It just doesn't apply it because it has already decided the question is simple.
A practical example from a developer who works regularly with Claude: when Opus 4.7 added itself to a benchmarks leaderboard in a web app, it was the first model to skip attaching a specific tooltip that all previous Claude models had added automatically — without any instruction to do so. The model had decided, apparently, that the task was done well enough. It took an explicit follow-up instruction to complete the behavior every previous version had handled unprompted.
Anecdotal? Sure. But directionally consistent with what adaptive thinking actually is: the model allocates effort based on its own assessment. And that assessment will sometimes be wrong.
What makes this particularly relevant now: adaptive thinking is mandatory for extended thinking. You cannot force Opus 4.7 to always reason at maximum depth. You can encourage it to think longer. You cannot override its judgment. That's a meaningful shift in control from the user's side of the table.
Verdict: For tasks that look simple but aren't — ambiguous prompts, nuanced code changes, edge cases — verify outputs more carefully than you did with 4.6. The adaptive system can misread the difficulty of the very tasks where you need it most.
What the Benchmarks Actually Say
Thesis: The coding and agentic numbers are legitimate. The picture elsewhere is messier.
Start with what's real. On SWE-bench Pro — which tests a model's ability to resolve real-world software issues from open-source repositories — Opus 4.7 scores 64.3%, up from 53.4% on Opus 4.6, and ahead of GPT-5.4 at 57.7% and Gemini 3.1 Pro at 54.2%. :antCitation[]{citations="7de97761-f9f9-4db2-a694-5c2ceb4f6c64"} That's a 10.9-point jump in a single version. Meaningful. On CursorBench, which measures autonomous coding in the AI code editor, the score is 70%, up from 58%. :antCitation[]{citations="ba8df37e-224b-4662-8c8f-7419cd51a8f5"}
For agentic work — complex multi-step tool-calling workflows — Opus 4.7 delivers a 14% improvement over Opus 4.6 on complex multi-step workflows while using fewer tokens and producing a third of the tool errors. :antCitation[]{citations="547b08ed-05f5-4318-af86-9d265aeda221"} That last number matters more than the headline percentages if you're building agents that run unsupervised.
On graduate-level reasoning, the picture is different. Opus 4.7 scores 94.2% on GPQA Diamond, GPT-5.4 Pro scores 94.4%, and Gemini 3.1 Pro scores 94.3%. The differences are within noise. :antCitation[]{citations="e74d62e6-e104-416a-98af-9b8709189b44"} The frontier has saturated this benchmark. None of these models is meaningfully better than the others here anymore.
Then there's BrowseComp — Anthropic's benchmark for agentic web search, retrieving hard-to-find information from across the open web. Opus 4.7 regresses on this compared to Opus 4.6. Even Claude Mythos Preview underperforms GPT-5.4 there. Anthropic doesn't compare 4.7 to Gemini models on this specific benchmark in their official materials — which is worth noting, given that an external test found Opus 4.7 underperforms the dramatically cheaper Gemini 3 Flash on OCR tasks.
Verdict: If you're doing coding, agentic tool use, or financial analysis workflows — the upgrade is clearly justified. If your workflow relies on web retrieval or document OCR at scale, the picture is murkier and worth testing before committing.
The Regressions Anthropic Admits To
Thesis: Two regressions appear in the system card. One is intentional. One is not.
Regression 1 — Cybersecurity vulnerability reproduction. Opus 4.7 underperforms both Opus 4.6 and Mythos Preview on certain cybersecurity vulnerability benchmarks. Anthropic's explanation, from page 48 of the system card: this was deliberate. During training, they worked to reduce the model's ability to reproduce certain vulnerability classes. They don't want a publicly available model to be too capable in this specific area. Whether you agree with that tradeoff depends on your use case, but it's a capability suppression by design, not an accidental regression.
Regression 2 — Long-context reasoning in certain MRCR configurations. On one version of the long-context retrieval benchmark — specifically a test involving finding the fourth poem across 1 million tokens — Opus 4.7 regresses compared to Opus 4.6 even at the maximum setting. Anthropic's lead creator of Claude Code addressed this: the benchmark was kept in the system card for scientific honesty, but they're moving away from it because it's built around "stacking distractors to trick the model" rather than measuring genuine long-context ability. That's a reasonable critique of the benchmark. It doesn't change the regression.
Opus 4.7 comes with a tokenizer refactor that may increase token counts by 1.0 to 1.35x depending on content type. :antCitation[]{citations="d4bdc7e0-29cd-499f-be29-04039aeb7cc5"} Same price per token. More tokens used. Do the math before assuming costs stay flat.
Verdict: Neither regression is catastrophic. But the BrowseComp drop plus the long-context MRCR issue means anyone working on retrieval-heavy or document-intensive workflows should run parallel tests before switching.
Vision: Genuinely Better, But Not Everywhere
Thesis: The vision upgrade is real and significant for specific use cases. It's not uniformly dominant.
Opus 4.7 now accepts images up to 2,576 pixels on the long edge — approximately 3.75 megapixels — more than three times the resolution capacity of Opus 4.6. :antCitation[]{citations="3e323fba-def0-4777-9f1e-64cbb2b54999"} In practice this means: fine print in contracts becomes readable, dense UI screenshots become parseable, detailed architecture diagrams stop losing information in downscaling. For computer use agents navigating graphical interfaces, higher resolution input directly translates to fewer navigation errors.
Visual acuity jumps from 54.5% to 98.5% :antCitation[]{citations="bcf53bfc-4085-4bb6-8196-3a39be70680c"} on Anthropic's internal measure. That is not a marginal improvement. For any workflow that involves Claude looking at screenshots, diagrams, or scanned documents, this is a meaningful upgrade.
The caveat: when an external benchmark group ran a comprehensive OCR test across models — the kind that systematically tests reading text out of documents — Opus 4.7 underperformed Gemini 3 Flash. The model that costs more than ten times less. Opus 4.7 is better than Opus 4.6 on OCR, but not best-in-class. Anthropic's vision benchmark comparisons don't include this result in their official materials.
Verdict: Use Opus 4.7 for computer use, UI analysis, and dense diagram reading — that's where the resolution gain shows up in real work. For high-volume OCR on standard documents, benchmark against Gemini 3 Flash on cost-efficiency before committing.
The Compute Problem Under Everything
Thesis: Anthropic's infrastructure constraints are shaping product decisions more than the benchmarks suggest.
The runaway growth of Claude — both Claude and Gemini have roughly 4x'd their market share in the AI website traffic category over the past year, with OpenAI's share potentially falling below 50% — has a cost. Literally.
According to a leaked OpenAI internal memo, their assessment is that Anthropic has not acquired enough compute, and that this gap "is going to show up in product." The memo specifically points to throttling, weaker availability, and a less reliable experience as potential indicators.
Whether you trust OpenAI's competitive analysis of a rival is a legitimate question. But the circumstantial evidence from Anthropic's own product side points in the same direction. An AMD senior AI director posted, with documentation, that the number of characters used for Claude's thinking had dropped by three-quarters before 4.7 even launched — far more bailing out, far less actual reasoning at depth. The lead creator of Claude Code confirmed that medium effort had become the default, and that users needed to actively set effort to high or maximum.
Adaptive thinking being mandatory for extended thinking, rather than optional, fits this pattern. You cannot force the model to always think longer. You can budget tokens more carefully with the new Task Budgets feature — itself a beta tool designed to cap how many tokens an agent loop consumes. Useful feature. Also useful when you're managing compute scarcity.
One of the Codex leads at OpenAI made a pointed observation: their coding agent is compute-efficient, always available, never throttled. That comment landed before GPT-5.4 released. Reliability is a product feature. Right now it's one area where OpenAI still has a structural advantage over Anthropic.
Verdict: The benchmarks are real. The infrastructure constraints are also real. For teams building products on top of Claude at scale, availability and rate limits matter as much as benchmark scores. Factor both into the decision.
Claude Code Upgrades Worth Knowing
A few genuine innovations shipped alongside Opus 4.7 in Claude Code, separate from the model itself.
Routines (Research Preview): You can trigger prompts on a schedule. The laptop doesn't need to be open when the task runs. :antCitation[]{citations="eb8b035b-789c-4174-81ec-17b1aabbf454"} Useful for background automation — nightly analysis, scheduled reports, anything you'd currently set up a cron job to handle.
/ultrareview command: A new deep code review that targets bugs standard review passes miss. It found a bug that GPT-5.4 had missed — though GPT-5.4 found a bug that Claude missed in the same codebase. :antCitation[]{citations="186998bf-6ef0-4590-a351-c43f5839e6be"} Neither is the perfect reviewer. Both are catching things the other doesn't.
Dispatch: Assign a task from your phone. Claude Code runs it on your local machine via the desktop app. The machine needs to be on; you don't need to be at it. Practical for longer-running tasks you want to kick off and walk away from.
Task Budgets (Public Beta): Set a global token budget before starting an agent loop. The model sees a budget countdown in each response and adjusts reasoning depth and tool calls based on remaining budget. :antCitation[]{citations="c2374c2d-fb8e-468e-8b0b-e759736a159f"} Before the budget runs out, it wraps up the core task. Direct control over cost and compute for agents — genuinely useful, not just a marketing feature.
xhigh effort level: A new setting between high and max, giving finer control over the reasoning-latency tradeoff. Claude Code defaults to xhigh for all plans. :antCitation[]{citations="dc9b194b-a51c-41b9-a9b7-8151ae9db4bf"} This is the effort tier you want active for complex coding and agentic tasks.
The Mythos 4x Productivity Claim, Dissected
Thesis: The internal 4x figure from Anthropic's Mythos survey is methodologically weak. That doesn't mean Mythos isn't exceptional — it means the specific number shouldn't be taken at face value.
Anthropic's system card for Mythos Preview included a striking claim: an internal survey found Mythos was speeding up Anthropic's own engineers by 4x. The AI world took notice. If true, recursive self-improvement was being raised as a serious near-term possibility.
The Opus 4.7 system card, on page 29, gives more detail about the survey. The question asked was: "How much more output did you produce over the past week compared to if you had no model access?" Not time saved. Not quality improvement. Just volume of output. And the survey was opt-in — meaning the engineers most likely to respond were those who had used Mythos most and found it most useful. A self-selected sample reporting on an output volume question is about as far from a controlled study as you can get while still calling it data.
The mismatch is real. Anthropic's CEO regularly discusses 50% white-collar unemployment risk, imminent disease cures, and the transformative impact of AI systems. Those are serious claims that the public is being asked to take seriously. The evidence base used to support them includes this survey. That's a tension worth naming.
On the cybersecurity side: external security researchers attempted to replicate several of the flagship vulnerabilities Mythos found that Anthropic publicized. In almost every case, models like Opus 4.6 or GPT-5.4 with appropriate scaffolding reached the same core vulnerability or got close. The more accurate frame, according to researchers, isn't "one lab has a magical model." It's that the economics of finding vulnerability signals are getting cheaper across the board. Mythos is probably exceptional. It's not uniquely magical.
The Dario-Brockman Rivalry and Why It Shapes the Models
The Wall Street Journal published details of a nine-year personal rivalry between Dario Amodei and Greg Brockman that provides useful context for understanding why Claude and Codex have developed differently.
Amodei joined OpenAI around mid-2016. He worked closely with Brockman in the early years, staying late to train agents on video games. By 2020, Amodei was saying he couldn't work with Brockman at all. The specific breaking points, according to reporting drawing likely on Amodei's own account: a proposed staff reduction process that Amodei found needlessly cruel, and a fundraising concept involving selling AGI access to nuclear-power UN Security Council members — including, potentially, Russia and China — which Amodei viewed as a fundamental ethical line. Brockman's framing of these events differs. But the personal animus was real and persistent.
Why does this matter for the models? Brockman is now leading Codex — OpenAI's primary coding agent, Claude Code's direct rival. And Brockman recently disclosed exactly where he believes OpenAI fell behind in the coding wars. In his words: OpenAI optimized for abstract programming competitions, pristine benchmarks, first-principles reasoning. Anthropic grounded their training data in messy real-world codebases. That one data strategy difference explains a significant portion of why Claude Code pulled ahead in developer preference over the past year — and why Brockman says catching up required a fundamental reorientation around "the last mile usability" that OpenAI only got serious about around mid-2025.
Two companies, shaped significantly by two men who couldn't work together, now competing on the most consequential software tools most developers will ever use. The personal history isn't gossip. It's strategy context.
My Take
The coding numbers are legitimate. SWE-bench Pro jumping nearly 11 points in a single release cycle, with a third of the tool errors at the same price point — that's real engineering progress, not benchmark theater. If your work is primarily in agentic coding, tool use, or complex multi-step workflows, Opus 4.7 is a clear step forward from 4.6.
The adaptive thinking mechanism is where I'd urge caution. The model deciding how much effort to apply to your task sounds fine until you realize it misreads the difficulty of exactly the tasks where you need it most — questions that look simple but aren't. That's not a small edge case. That's the core of most real-world engineering decisions. Prompts that are genuinely ambiguous don't announce themselves as ambiguous. Neither do the bugs that are buried in exactly the kind of messy real-world codebase that Claude was trained on.
The compute situation is the variable that almost no one's pricing in. Anthropic is growing fast enough that their infrastructure is visibly straining — throttling, reduced thinking depth as a default, rate limit complaints from high-volume users. The models are getting better. The delivery infrastructure has to keep pace with the model improvements. Right now there's a gap. How Anthropic closes it — whether through the Amazon partnership, additional fundraising, or architectural efficiencies — will matter as much as the next model release.
On Mythos: exceptional, probably. But "it changes the economics" is a more defensible claim than "it does things no other model can." The security researchers who tried to replicate those results with scaffolded Opus 4.6 are more informative than the internal survey about 4x productivity. Strong model. Overstated case.
Key Takeaways
- Coding benchmarks are the strongest part of the release — SWE-bench Pro up 10.9 points, tool errors down by two-thirds.
- Adaptive thinking is mandatory for extended thinking. You cannot override it. Verify outputs on tasks that look simple but aren't.
- Two regressions in the system card: cybersecurity vulnerability reproduction (intentional) and a long-context MRCR configuration (documented, benchmark being retired).
- Vision resolution tripled — major for computer use and dense diagram parsing. Not best-in-class for document OCR.
- Tokenizer refactor may increase token counts 1.0–1.35x. Same price per token, potentially higher total cost.
- Claude Code gains: Routines, /ultrareview, Dispatch, Task Budgets, xhigh effort tier — all practical additions.
- Compute constraints are shaping product defaults. Medium effort is now the default. Track rate limits before committing to high-volume builds.
- The Dario-Brockman rivalry explains meaningfully why Claude Code beat Codex on real-world codebase tasks — data strategy, not just raw model capability.
FAQ
Should I upgrade from Opus 4.6 to 4.7 immediately?
For coding and agentic tool use — yes, the improvement is real. For web retrieval, document OCR, or long-context retrieval-heavy workflows, run parallel tests first. The BrowseComp regression and long-context MRCR issue are specific enough that they'll affect some workflows and not others. Don't assume a clean upgrade.
What is adaptive thinking and how does it affect my prompts?
The model assesses how complex it believes your task is and allocates reasoning time accordingly. If it misjudges a task as simple, it will apply less thinking depth than you'd expect. For critical tasks, be explicit about complexity in your prompt. The xhigh effort setting in Claude Code pushes it toward deeper reasoning by default.
What is the new xhigh effort tier?
A new inference setting between high and max, giving finer control over reasoning depth versus latency. Claude Code defaults to xhigh for all plans. :antCitation[]{citations="8d4d8afa-2484-418d-a7df-33771d7db33e"} For API users, you'll want to set effort to high or xhigh explicitly for complex coding and agentic tasks rather than relying on the default.
Has the price changed?
Pricing stays the same as Opus 4.6 — $5 per million input tokens and $25 per million output tokens. :antCitation[]{citations="9d8d6b43-ba8a-4918-82f1-92579c708313"} However, the tokenizer refactor means some content types may use more tokens than before. Real-world costs could be higher depending on your content mix.
Is Claude Mythos Preview available to the public?
No. Mythos Preview remains limited to select enterprise and government partners. Anthropic has stated this is intentional — they're testing new cybersecurity safeguards on less capable models before expanding access. Opus 4.7 is the most capable generally available Claude model as of April 2026.
Where is Opus 4.7 available?
Available across Claude products, the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. :antCitation[]{citations="ada1567b-3f8b-4e50-821e-cf876c34bc1c"} The model ID for API access is claude-opus-4-7.
What happened to Opus 4.5 and 4.6?
Opus 4.5 was quietly removed, and Opus 4.6 is being deprecated — a pattern that drew significant user backlash, similar to the criticism OpenAI received when it deprecated earlier GPT models. Anthropic has acknowledged the issue. It reflects the pace of the release cycle more than any deliberate disregard for users on older versions.
The benchmarks tell you what Opus 4.7 can do under ideal conditions. The system card tells you where Anthropic made deliberate tradeoffs. Both are worth reading before deciding how much of your workflow to route through the new model. The headline numbers favor an upgrade. The specifics — adaptive thinking calibration, tokenizer costs, compute availability — are where the actual decision lives.
For a broader look at how all current Claude models compare on cost and use case, the RevolutionInAI model guide covers Haiku, Sonnet, and Opus tiers in one place.
0 Comments