Claude Opus 4.7 Is Out — 87.6% on SWE-bench, New Agentic Controls, Same Price
Anthropic's latest flagship model scores 87.6% on SWE-bench Verified and introduces a new xhigh effort tier and token budget system for agentic loops — all at the same $5/$25 per million tokens as Opus 4.6.
Anthropic released Claude Opus 4.7 on April 16. SWE-bench Verified sits at 87.6% — up 6.8 points from Opus 4.6’s 80.8% and 7 points ahead of Gemini 3.1 Pro’s 80.6%. On Anthropic’s internal 93-task coding benchmark, Opus 4.7 resolved 13% more tasks than Opus 4.6, including four problems that neither Opus 4.6 nor Sonnet 4.6 could crack. Pricing didn’t move: $5 per million input tokens, $25 per million output.
New agentic controls
Two developer-facing additions stand out. The first is an xhigh effort level — a new tier sitting between high and max — giving finer control over the reasoning-vs-latency tradeoff on hard problems without burning max compute on every call.
The second is a task budget system. Pass Claude a rough token target for a complete agentic loop and it receives a running countdown. The model reads that counter, prioritizes work accordingly, and aims to finish the task gracefully before the budget runs out. In practice this gives developers a way to cap runaway agentic tasks without relying on hard external timeouts.
Vision improvements
Maximum image resolution increased to 2,576 pixels on the long edge — roughly 3.75 megapixels, more than 3x the previous ceiling. The improvement is most visible on handwritten technical diagrams, whiteboard photos, and dense infographics where earlier models blurred detail.
Benchmark snapshot
| Benchmark | Opus 4.7 | Opus 4.6 |
|---|---|---|
| SWE-bench Verified | 87.6% | 80.8% |
| MCP-Atlas (agentic) | 77.3% | — |
| CyberGym | 73.1% | — |
For comparison, GPT-5.4 scores 66.3% on CyberGym. Claude Mythos Preview, Anthropic’s unreleased frontier model, sits at 83.1% on the same benchmark.
Behavioral change worth noting
Opus 4.7 executes the exact text provided. Opus 4.6 had a tendency to fill in gaps when instructions were ambiguous. Prompts that relied on that implicit inference behavior will produce literal but potentially wrong results in 4.7. This is strictly better for production systems that need predictable outputs — but test your existing prompts before upgrading in production.
Tokenizer update
Anthropic shipped a revised tokenizer alongside Opus 4.7. Depending on content type, token counts may run 1.0–1.35x higher than on the old tokenizer. The $5/$25 pricing is unchanged, but budgeted workflows should account for the potential count increase.
Availability
Opus 4.7 is live across Claude.ai, the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The context window stays at 1 million tokens.
Anthropic simultaneously acknowledged that Opus 4.7 trails its unreleased Claude Mythos model — the one that leaked in March — on several advanced reasoning tasks. That’s now the stated ceiling for what’s publicly available.