Analyst: Anthropic's Opus 5.5 cuts AI coding costs by 40%, beats OpenAI GPT-6 Astra at one-fifth the price
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
Aakash Gupta reports that Anthropic's new Opus 5.5 model will significantly reduce AI coding costs, far exceeding the commonly cited 20% price drop on regular tokens. The key savings come from a 60% reduction in the cost of re-reading files, which constitutes most of an AI agent's expenses. Additionally, the model requires fewer steps to complete tasks, with one early tester fixing a 200,000-line codebase in under 3 hours compared to over 20 hours for the old Opus. Anthropic states the typical bill drops about 40% when combining these factors. In a test measuring whether AI code would be approved by a real team, Opus 5.5 on its normal setting beat the best score ever posted by OpenAI's GPT-6 Astra, at roughly a fifth of the cost per task. Gupta notes this is the first time in the AI race that the top-performing coding model is also the cheaper one to run, posing a challenge for OpenAI to respond with a price cut.
Source report
The widely cited 20% price reduction on Anthropic's new Opus 5.5 only tells part of the story — the real savings for AI-assisted coding are significantly larger.
Where the Savings Come From
The 20% figure applies to regular tokens. However, when an AI agent works on code, the bulk of the cost comes from repeatedly re-reading the same files before every single step. That re-reading just became 60% cheaper.
The model also requires fewer steps to complete tasks. One early tester reported that Opus 5.5 fixed a 200,000-line codebase in under 3 hours — a job that took the previous Opus model over 20 hours.
Stack these two improvements together, and Anthropic estimates the typical bill drops by roughly 40%.
The OpenAI Comparison
The most striking benchmark involves a test that measures whether an AI's code would actually be approved and merged by a real development team. Opus 5.5, on its normal setting, beat the best score ever posted by GPT-6 Astra — and did so for about one-fifth the cost per task.
That's a better result for 20 cents on the dollar.
A Shift in the AI Race
For most of this AI race, the model that topped the leaderboard was also the one you paid a premium for. Opus 5.5 marks the first time in a while that the model winning coding tests is also the cheap one to run — a challenge that OpenAI cannot easily answer with a price cut of its own.
Source
aakashguptaNeutral / independent
Part of this Story
Anthropic launches Claude Opus 5.5 with Fable 5.1-level performance at 40% lower cost