Claude Opus 4.7 reviewed a month in: 'quiet upgrade that earns its keep at work'

A Medium write-up by independent reviewer Kosuke Okura published May 20 frames Claude Opus 4.7 — shipped on April 16 — as the rare model release whose value emerged through use rather than launch-day benchmarks. The review focuses on three concrete improvements: agentic reliability over multi-step tool-using workflows, long-context reasoning that holds up past 100K tokens of business documents, and tool-use consistency where Opus 4.7 reliably calls the right MCP tool with valid arguments instead of hallucinating tool calls.

The 'quiet upgrade' framing is significant against this week's noise. Google I/O launched Gemini 3.5 Flash and previewed Omni; OpenAI claimed a math conjecture disproof; Anthropic itself hired Karpathy and acquired Stainless for $300M+. Against that, Opus 4.7 isn't generating headlines — but it appears to be generating retention. The review cites real-work scenarios (long legal-doc review, multi-step coding tasks, structured data extraction) where 4.7's outputs are 'boringly correct' in a way that 4.6 wasn't.

The new fact here, against prior briefings that just announced the 4.7 release, is the month-in retrospective: this is the first independent at-scale assessment, not Anthropic's own benchmarks. It validates Anthropic's strategy of shipping incremental model improvements that compound in agentic settings rather than chasing leaderboard wins — exactly the strategy Karpathy joining the pre-training team is meant to extend.

Competitively, this lines up against Ethan Mollick's observation (in today's social signals) that the gap between ChatGPT/Codex and Claude Code/Cowork is closing as OpenAI and Anthropic 'converge on a single experience,' while Google's experiences are diverging. If the AI assistant category is converging on roughly Opus-4.7-shaped capability, the differentiation moves to ecosystem, pricing, and trust — exactly where Anthropic's enterprise positioning is strongest.