Anthropic released Claude Opus 4.7 on 16 April, delivering the most significant upgrade to its flagship model since Opus 4.6 launched in February. The headline numbers are substantial: a 13 per cent coding performance lift on a 93-task benchmark, three times more production tasks resolved on the Rakuten-SWE-Bench suite, and a 21 per cent reduction in errors on document reasoning tasks measured by OfficeQA Pro. Complex workflow tasks see a 10 to 15 per cent improvement, and the model now handles long-running agentic operations with what Anthropic describes as 'more rigour' — following instructions more precisely and verifying its own outputs before reporting back.
The vision upgrade is the most dramatic single improvement. Visual acuity jumped from 54.5 per cent on Opus 4.6 to 98.5 per cent on Opus 4.7, driven by a tripling of maximum image resolution from 1.15 megapixels to 3.75 megapixels (2,576 pixels on the longest edge). Coordinates now map one-to-one with actual pixels, making the model viable for screenshot analysis, document digitisation, and precision UI tasks that previously required dedicated vision tooling. Alongside the vision upgrade, Anthropic introduced a new 'xhigh' reasoning effort level that sits between 'high' and 'max', giving developers finer-grained control over the trade-off between response speed and reasoning depth.
Pricing remains unchanged at $5 input and $25 output per million tokens. The model is available immediately via the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry across all subscription tiers. Anthropic also acknowledged the existence of Claude Mythos — a restricted-access model that outperforms Opus 4.7 on all benchmarks but remains available only to Project Glasswing partners (Apple, Google, and Microsoft) at $25/$125 per million tokens. CNBC and Axios both reported that Anthropic concedes Mythos is more capable, framing Opus 4.7 as the most powerful generally available model rather than the most powerful model full stop.
For context engineers, Opus 4.7 changes the daily calculus in two concrete ways. The vision leap means Claude Code sessions can now reliably parse screenshots, design mockups, and scanned documents without preprocessing — workflows that were fragile on Opus 4.6 should be retested. The self-verification capability matters even more for agentic use: if the model can check its own work on multi-step tasks before returning results, the error rate on autonomous operations drops meaningfully, and the case for trusting Claude with longer-running unattended workflows gets stronger. Combined with Claude Code Routines shipping just two days earlier, the timing positions Anthropic to push Claude into genuinely autonomous production work at a pace the other frontier labs are still catching up to.
