Anthropic launched Claude Opus 4.7 on April 16 across API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, with GitHub Copilot Pro+, Business, and Enterprise users receiving the update simultaneously. This is the strongest publicly available model, positioned below the internal Mythos model, which is restricted to federal agencies due to zero-day vulnerability risks.
Three Core Upgrades: Coding, Vision, Context
Coding leads the improvements. Opus 4.7 scores 64.3% on SWE-bench Pro, up 10.9 points from 4.6's 53.4%, surpassing GPT-5.4 at 57.7% and Gemini 3.1 Pro at 54.2%. SWE-bench Verified jumps from 80.8% to 87.6%, and CursorBench rises from 58% to 70%. Multi-step agentic workflows improve by 14%, with tool usage errors cut to roughly one-third. For vision, it supports up to 2576px/3.75MP—three times higher than 4.6—with a 13% gain on vision benchmarks. The context window expands to 1M tokens at no extra cost, keeping the $5/$25 per million tokens pricing and 128K max output, enabling full codebase analysis without chunking.
API Changes: xhigh Level, Adaptive Thinking, Removed Sampling
API changes stir the most controversy. A new xhigh (extra high) effort level targets high-intensity tasks below max reasoning. The /ultrareview command is added for deep code review. Extended Thinking is replaced by adaptive thinking, where the model decides whether and how long to think; the old budget_tokens parameter now returns a 400 error. More critically, temperature, top_p, and top_k are removed, breaking deterministic outputs and structured data pipelines. Built-in cyber safeguards automatically block prohibited high-risk requests.
User Impressions and Community Backlash
In practice, a cross-file refactor spanning 8 files completed successfully with test updates, boosting success rates from 60% to 75-85%. Tool call errors dropped notably, and vision accuracy improved from 75% to ~90% for PDF tables. But creative writing and advanced math show no clear gains; the model sticks more strictly to explicit instructions, reducing “filling in” behavior. Community backlash focuses on three points: the new tokenizer inflates bills by 30-35% despite unchanged per-token pricing; removed sampling parameters force test suite rewrites; adaptive thinking keeps reasoning hidden by default (opt-in required). Anthropic claims adaptive mode wins on average, but power users demand predictability, not average wins.