Claude Opus 4.7 Is Here: The Benchmarks, the Pricing, and What Actually Matters

Anthropic shipped Claude Opus 4.7 this week. That is two months after Opus 4.6, and about six weeks after Google's Gemini 3.1 Pro. On paper, it is an incremental release. In practice, it is the kind of point upgrade that does not look exciting in a press release but changes what you can build once you wire it into a real agent loop.

Here is what shipped, what the numbers actually mean, and where I think Opus 4.7 should (and should not) fit into your stack.

---

What Shipped

Opus 4.7 is Anthropic's new flagship hybrid reasoning model. It is aimed at professional software engineering, long-running agentic workflows, and high-stakes enterprise work. The headline details:

• 1M-token context window, generally available, same as Opus 4.6.
• Pricing unchanged at $5 / $25 per million tokens for input and output, with up to 90% cache savings and 50% batch savings.
• Adaptive thinking. The model auto-adjusts reasoning depth based on task complexity.
• A new xhigh effort level that sits between high and max, giving you finer control over the reasoning/cost/latency tradeoff.
• Task budgets (beta). Claude can now prioritize work and cap spend across long agentic runs.
• Vision at 3x the resolution of Opus 4.6. Cleaner slides, cleaner interfaces, cleaner PDF parsing.
• Available on day one on the Claude Platform, AWS Bedrock, Google Vertex AI, and Microsoft Foundry.

Nothing in that list is revolutionary. But task budgets, xhigh, and the vision upgrade are exactly the kind of boring infrastructure features that matter a lot once you are actually running agents in production and paying the bill.

---

The Benchmarks

Benchmarks are imperfect, but they are the only apples-to-apples signal we have. Here is where Opus 4.7 lands against the frontier:

SWE-bench Verified (real GitHub issue resolution):

• Opus 4.7: 87.6%
• Opus 4.6: 80.8%
• Gemini 3.1 Pro: 80.6%

SWE-bench Pro (the harder, agentic variant):

• Opus 4.7: 64.3%
• GPT-5.4: 57.7%
• Gemini 3.1 Pro: 54.2%
• Opus 4.6: 53.4%

On an internal 93-task coding benchmark, Opus 4.7 resolved 13% more tasks than Opus 4.6, including four that neither Opus 4.6 nor Sonnet 4.6 could solve at any effort level.

A 7-point jump on SWE-bench Verified and an 11-point jump on SWE-bench Pro is larger than what you would expect from a point release. For context, the gap between Opus 4.7 and GPT-5.4 on SWE-bench Pro is now roughly the same size as the gap between GPT-4 and GPT-5 at their respective launches. Coding is where Anthropic has chosen to plant its flag, and this release widens the lead.

---

My Honest Take

Unbiased is not the same as neutral. Here is what I actually think after reading the release notes and the leaks.

What is genuinely impressive.

The coding numbers are real, and they will translate to production. A 7-point gain on SWE-bench Verified is not a demo-rigged number. It is the kind of improvement that shows up as "the agent fixed it on the first try" instead of "the agent fixed it after three retries and a human nudge." For forward deployed engineers running agent loops at customer sites, that delta is the difference between a billable hour saved and a billable hour burned.

Task budgets are the feature nobody is talking about, and everybody will end up using. If you have ever watched a long-running agent quietly burn $40 because it got stuck retrying a failing tool call, you know exactly why this matters. Capping agent spend declaratively, instead of through brittle middleware, is the right primitive.

The 3x vision resolution is a bigger deal than the release notes suggest. Most enterprise workloads bottleneck on reading poorly-scanned PDFs and messy invoices, not on generating clever prose. This is where real money gets unlocked.

What is less exciting.

The context window did not grow. 1M tokens is the same as Opus 4.6, and in a world where Gemini has been shipping 2M+ windows, "same as last time" is starting to feel like a ceiling. Most workloads do not need more than 200K, but the headroom matters for codebase-scale reasoning.

"Adaptive thinking" is as much a branding move as a capability move. Every frontier model now auto-adjusts reasoning. Anthropic is just naming it. Fine, but do not mistake the branding for a breakthrough.

Pricing held flat at $5/$25. That is good for existing users, and not so great for anyone hoping Anthropic would close the cost gap with Gemini 3.1 Flash-Lite ($0.25/M input) at the low end. Opus 4.7 is not trying to be cheap. It is trying to be the best. At that tier, it wins. Outside that tier, the economics still favor Gemini or open models for most production volume.

Where I would actually use it.

Agent loops that touch production code. The SWE-bench gains translate. Use Opus 4.7 as the planner or the last-mile fixer, and run cheaper models for the repetitive scaffolding.

Long-horizon agentic runs where the budget is a real concern. Task budgets plus xhigh give you the knobs to actually control spend.

Document-heavy enterprise pipelines. The vision upgrade pays for itself on messy PDFs.

Where I would not.

High-volume, low-complexity chat. Use Haiku 4.5 or Sonnet 4.6. Paying Opus prices for intent classification is malpractice.

Projects already well-served by Opus 4.6. The upgrade is real, but not urgent. Migrate deliberately, not reflexively.

---

The Bigger Picture

Every frontier lab is now shipping point-release upgrades every 6 to 10 weeks. The pace of "wow" releases is slowing. The pace of meaningful, ship-it-today improvements is accelerating. Opus 4.7 is a clean example of what that looks like. Nothing you would put on a billboard, but enough that if you are building agents seriously, you will notice the difference within a week.

The coding lead Anthropic is building is not a moat. But it is a real gap. Until GPT-5.5 or Gemini 3.2 closes it, Opus 4.7 is the model to beat for agentic software engineering, and the default choice when the work actually has to ship.

If you are running a serious agent stack, try it. Compare your own task success rates, not the leaderboards. And budget your tokens.