Claude Opus 4.7 - A New Frontier, in Performance … and Drama

TL;DR

Claude Opus 4.7 leads on real-world professional tasks but sparks controversy over compute throttling, benchmark regressions, and the Anthropic-OpenAI rivalry.

Key Points

• 1.Opus 4.7 leads on professional tasks but has notable benchmark regressions. It outperforms Opus 4.6 on most coding and knowledge benchmarks and tops vibe-coding tests, yet underperforms the 10x cheaper Gemini 3 Flash on OCR tasks and scores worse than Opus 4.6 on agentic web search and Simple Bench due to adaptive thinking misjudging task difficulty.
• 2.Anthropic intentionally capped Opus 4.7's cybersecurity and thinking capabilities. Per page 48 of the system card, they deliberately reduced vulnerability-reproduction ability during training; separately, 'medium effort' thinking is now the default, meaning users must manually set high or max effort — a change confirmed after AMD's senior AI director showed thinking token usage dropped by 75%.
• 3.Anthropic hit an AI industry first with a $1 trillion valuation, but financials are disputed. A leaked OpenAI internal memo claims Anthropic's revenue run rate is overstated by ~$8B, putting it closer to $22B — still behind OpenAI — while also alleging Anthropic lacks sufficient compute, causing throttling and reliability issues for users.
• 4.The Claude Mythos 4x productivity claim was based on a deeply flawed opt-in survey. The survey asked engineers how much more output they produced (not time saved or quality improved) and was opt-in rather than random-sampled, making it unscientific — a significant credibility problem given CEO Dario Amodei's public claims about imminent 50% white-collar unemployment and recursive self-improvement.
• 5.External security labs found Mythos's headline vulnerabilities are replicable by other models with the right scaffold. Vidok security lab reproduced nearly all of Anthropic's showcased Mythos bug-finds using Opus 4.6 or GPT-5.4, framing the real shift as economics — finding vulnerability signals is getting cheaper — not a single lab holding a uniquely magical model.
• 6.A 9-year personal rivalry between Dario Amodei and Greg Brockman underpins the Anthropic-OpenAI coding war. Amodei clashed with Brockman at OpenAI as early as 2016 over mass layoffs and the idea of selling AGI to nuclear powers including Russia and China; Brockman now leads Codex, and recently admitted OpenAI lost the coding edge by training on abstract competitions rather than messy real-world codebases — the exact area Anthropic prioritized.