Claude Opus 4.5: Price Cut, Effort Control & Coding Claims

The price moved more than the name suggests

Anthropic announced Claude Opus 4.5 on November 24, 2025, priced at $5 per million input tokens and $25 per million output tokens through the API. The prior Opus model, Opus 4.1, had listed at $15 and $75. The new price is one third of the old one, in both directions at once. [1]

The model shipped across the usual surfaces at once: Anthropic's own apps, the API as claude-opus-4-5-20251101, and Amazon Bedrock, Google Cloud, and Microsoft Foundry. Anthropic also removed the separate usage cap it had applied to Opus inside Claude and Claude Code for eligible users. [1]

Atlas interpretation: A flagship model getting cheaper eight months after the last Opus price cut is not a one-off discount. It reads as a release cadence: each Opus generation, including the one that followed Sonnet 4.5, is priced to replace its predecessor rather than to sit beside it at the old rate. [1][2]

A dial for how much the model spends thinking

Opus 4.5 introduced an effort parameter on the API that lets a caller trade capability against token spend. Anthropic's own comparisons put Opus 4.5 at medium effort matching Sonnet 4.5's performance while using 76% fewer output tokens, and at its highest effort setting exceeding Sonnet 4.5 by 4.3 percentage points on the same suite while still using 48% fewer tokens. [1]

Atlas interpretation: Those figures are Anthropic measuring Anthropic's own models against each other, not an outside benchmark. They are useful mainly as a statement of intent: the company is selling token efficiency as a feature in its own right, alongside raw score, because at these price points the bill for a long agentic session is now a real part of the pitch. [1]

Coding claims that are hard to independently pin down

Anthropic described Opus 4.5 as its best model yet for coding, agentic tasks, and computer use, citing the top score on SWE-bench Verified among frontier models and leading scores on 7 of 8 languages in SWE-bench Multilingual, without publishing a specific SWE-bench Verified percentage in the announcement itself. [1]

Independent hands-on testing complicated the picture. Simon Willison used Opus 4.5 in Claude Code for a real refactoring session, roughly 20 commits and 3,000 changed lines in the sqlite-utils project, then switched back to Sonnet 4.5 partway through and kept working at the same pace. He wrote that he could not say with confidence the task he had posed was able to surface a meaningful difference between the two models. [2]

Atlas interpretation: This is less a knock on Opus 4.5 than a comment on where frontier coding models had gotten to by late 2025. Benchmark gaps of a few percentage points and real-world sessions that feel indistinguishable are not in tension, they are two measurements of the same convergence, and the gap between a flagship model and the mid-tier one released two months earlier had narrowed enough that a working engineer doing ordinary tasks could not reliably tell them apart. [2]