Claude Opus 4.7 Is Here: Same Price, Sharper Reasoning, and a Tokenizer That Quietly Raises Your Bill

Anthropic’s new flagship lands with a 64.3% score on SWE-bench Pro, 3x higher image resolution, and a new “xhigh” effort tier. But the real story is what the benchmarks don’t tell you.

Anthropic released Claude Opus 4.7 this morning, and if you’ve been running agentic coding workflows on Opus 4.6, the numbers in the release will read like a quiet victory lap. A jump from 53.4% to 64.3% on SWE-bench Pro is not a rounding error. It is the difference between a model you babysit and a model you hand work to and walk away from.

But the headline benchmarks are only half the story. The other half is buried three paragraphs deep in the technical notes: a new tokenizer, a stricter instruction-following style, a deliberately throttled cybersecurity surface, and a pricing structure that looks unchanged on the invoice but costs more in practice. If you are shipping production software on top of Claude, those four things will decide whether your next bill makes sense.

This is the breakdown you will not find anywhere else today — because we are not going to repeat the press release. We are going to tell you what matters, what to test before you migrate, and where Opus 4.7 actually earns its $25-per-million-output-tokens price tag.

Video: Anthropic — Introducing Claude Opus 4.7 · Open on YouTube

The Benchmark Table Nobody Else Is Showing You in One Place

Most coverage today cherry-picks one or two numbers. Here is the consolidated view across the benchmarks Anthropic published this morning, alongside the competition. Every figure below is from the official release or verified secondary reporting.

1. The Coding Story: From “Supervised Intern” to “Junior Engineer Who Ships”

The headline shift in Opus 4.7 is how far coding performance has moved at the hard end of the distribution. Opus 4.6 was excellent at boilerplate, refactors, and single-file fixes, but it still needed a human in the loop for multi-file PRs, complex refactors across legacy code, and long-running agentic tasks. Opus 4.7 closes most of that gap.

The Rakuten team, which benchmarked against production workloads internally, reports that Opus 4.7 resolves three times the production tasks Opus 4.6 could handle, with double-digit gains in both code quality and test quality. That matches what Anthropic is describing in its own release: the model is designed to handle long-running work with rigor and to verify its own outputs before reporting back.

What this means practically

If you run Claude Code or any agentic coding harness, Opus 4.7 is the first Anthropic model where the output-verification loop actually closes. The model writes code, runs tests it wrote itself, catches its own regressions, and only then reports a task complete. Previous versions could claim completion on a half-finished diff. This one is meaningfully harder to fool — including by itself.

Claude Code also ships with a new /ultrareview command that mimics a senior human reviewer instead of a linter — it flags design flaws and logic gaps, not just syntax. And “auto mode,” the setting that lets Claude make decisions without asking permission at every step, is now extended to Max plan users. For teams already building on top of Anthropic’s stack, that is the quiet productivity unlock.

2. The Vision Upgrade That Actually Matters for Agent Workflows

Every previous Claude capped images at 1,568 pixels on the long edge — roughly 1.15 megapixels. Opus 4.7 raises that ceiling to 2,576 pixels, or about 3.75 megapixels. That is a 3.26x increase in total pixel count.

If you are building anything that reads screenshots, parses diagrams, extracts data from design mockups, or runs a computer-use agent, this changes the game. The old resolution forced you to choose between readable UI text and visible interface chrome. The new resolution lets you send a full-fidelity screenshot and have the model actually read the small print.

Coordinate mapping is now 1:1 with actual pixels, which eliminates the scale-factor math that computer-use workflows previously had to hand-roll. On OfficeQA Pro, the document-reasoning benchmark, Opus 4.7 hits 80.6% accuracy — up from 57.1% on Opus 4.6. That is the kind of jump that reclassifies the task from “experimental” to “shippable.”

3. xhigh, Task Budgets, and the New Control Surface

Opus 4.7 introduces two new knobs that matter if you are running agents at scale.

The first is xhigh, a new reasoning effort level that sits between high and max. Previous Opus models gave you three gears; this one gives you four. The point is that “max” is expensive and slow, and “high” sometimes is not enough for genuinely hard problems. xhigh is the setting for when you want depth without committing to full max latency.

The second is task budgets, currently in public beta. Task budgets let you set a hard token ceiling on an autonomous agent’s execution — meaning a long-running debugging session cannot silently consume your entire monthly spend. This is the feature that makes Opus 4.7 viable for production agentic workloads at teams that care about unit economics. Without a budget ceiling, autonomous loops are a financial risk. With one, they are a line item you can plan around.

4. The Pricing Trap: Same Sticker, Higher Invoice

Here is the part most launch coverage is skipping.

Opus 4.7 is priced at $5 per million input tokens and $25 per million output tokens — identical to Opus 4.6. Prompt caching still offers up to 90% savings, and the Batch API still cuts both sides by 50%. On paper, nothing changed.

In practice, Anthropic shipped a new tokenizer with Opus 4.7 that maps the same text to anywhere from 1.0x to 1.35x more tokens than the old one. Your prompts did not change. Your code did not change. Your bill can still go up by a third if your workload happens to tokenize unfavorably under the new scheme.

This matters most for teams running high-volume, short-prompt workloads — chatbots, classification agents, RAG pipelines — where token overhead dominates actual content. Long-form coding and document analysis will see smaller relative deltas. Either way, the single most important thing you can do before migrating is run your real workload on a sample and measure actual token consumption, not just trust the unchanged sticker price.

The 1M token context window carries over from Opus 4.6 at no long-context premium, and max output per response stays at 128k. Those are both meaningful for agentic workloads where context accumulation is the bottleneck.

5. The Cyber Throttle Is a Feature, Not a Bug

Anthropic has been increasingly public about the tradeoff between capability and cyber safety, and Opus 4.7 is the first generally available model where that tradeoff ships in the weights themselves.

During training, Anthropic deliberately reduced certain cybersecurity capabilities. On top of that, Opus 4.7 includes runtime safeguards that detect and auto-block requests which pattern-match to prohibited or high-risk cyber uses. For most users, this will be invisible. For legitimate security researchers, red teams, and cyber consultancies, it is a workflow concern — and Anthropic has launched a Cyber Verification Program that lets qualified teams apply for exceptions.

The context here is Project Glasswing — Anthropic’s internal initiative around cyber-capable models — and the fact that Anthropic’s more powerful Mythos Preview is being held back from general release specifically because of cyber capability concerns. Opus 4.7, in the company’s framing, is partly a stress-test of the safeguards they want in place before anything Mythos-class ever gets a broader rollout.

Who Should Migrate, and Who Should Wait

Not every workload benefits from jumping to Opus 4.7 on day one. Here is the honest cut.

Migrate immediately if:

You run agentic coding workflows (Claude Code, custom harnesses, CI/CD integration). You rely on computer-use agents that parse screenshots or UI. You run document-intelligence pipelines where OfficeQA-style reasoning matters. You are currently hitting the ceiling of Opus 4.6 and considering a move to a competitor — Opus 4.7 changes that math.

Test before migrating if:

You run high-volume short-prompt workloads where the new tokenizer could inflate your bill meaningfully. You have prompt chains tuned specifically for Opus 4.6’s looser instruction interpretation. You depend on legitimate cybersecurity tooling that might trip the new safeguards.

Stay on Sonnet if:

Cost per call matters more than maximum capability. Sonnet remains the balanced workhorse in the lineup, and most customer-facing chatbot and classification workloads do not need Opus-tier reasoning. Opus is a premium tier for work no other model can handle, and it is priced accordingly.

Model ID on the API is claude-opus-4-7. It is live now on claude.ai, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.