Claude Opus 5.5 vs GPT-6 Sol: Same-Day Launch, Lower Prices

The price of a capable AI model dropped twice on the same Tuesday. On September 22, OpenAI released GPT-6 Sol and GPT-6 Luna at half the price of the models they replace, and Anthropic released Claude Opus 5.5, which it says does the work of its top model, Fable 5.1, for 40% less than Opus 5. Both companies are chasing the same buyer: the business that runs AI at volume and watches the monthly bill.

What OpenAI shipped

GPT-6 Astra stays the flagship. Sol and Luna are the cheaper, faster members of the family, trained with similar methods. API prices per million tokens fell 50% against GPT-5.6:

• GPT-6 Sol: $2 input and $10 output, down from $4 and $20
• GPT-6 Luna: $0.10 input and $0.50 output, down from $0.20 and $1.20
• Cached input reads are discounted 90%

A few hours later OpenAI followed with better prompt caching for GPT-6: higher cache hit rates by default, explicit breakpoints and new diagnostics. Both models are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and Luna reaches Free and Go users in the desktop app.

What Anthropic shipped

Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. Cache reads fell 60%, to $0.20. The headline 40% saving combines that lower price with fewer tokens spent per task, according to Anthropic.

The performance claims are Anthropic's own: Terminal-Bench 4.0 rises from 52.3% to 66.4% over Opus 5, and CursorBench 4.0 from 46.6% to 57.8%. In a code-review test cited by Deloitte, Opus 5.5 caught 72% of known bugs at its lowest effort setting, against 56% for Opus 5 at high effort. Artificial Analysis, an independent benchmarker, gave it a score of 58 on its Intelligence Index, the highest it has recorded. Output is 30% faster than Opus 5, and a fast mode at 2.5x speed costs $8 and $40. Anthropic also says it has fixed the formulaic, jargon-heavy prose critics had dubbed "Claudish".

The benchmarks were a generation old by dinner

OpenAI's launch post positions Sol against Claude Opus 5: level with it on the OSWorld 2.0 computer-use test (60.5% versus 60.3% for Opus 5 at medium effort), ahead of it on AutomationBench at 9% of its max-effort cost. Those comparisons went stale the moment Opus 5.5 appeared. Anthropic's chart, for its part, lines up against GPT-6 Astra and the outgoing GPT-5.6 Sol. Neither side has published a like-for-like result against the other's new release yet.

There is also a catch in Anthropic's efficiency story. The Decoder reports that Opus 5.5 can use roughly 119,000 output tokens per task at maximum effort. Fewer tokens per task depends on the effort setting you choose.

How to read the price gap

On sticker prices alone, a workload of 50 million input and 10 million output tokens a month costs $200 on GPT-6 Sol, $400 on Claude Opus 5.5 and $10 on GPT-6 Luna. That arithmetic ignores the two factors both companies are now competing on: cache hit rate and tokens per task. An app with a long, fixed system prompt can see caching erase much of the headline difference, and a model that finishes in fewer tokens can beat a cheaper one that rambles.

So do not switch on the press release. Replay a few hundred real requests from your logs through both models and record accuracy, tokens used and cache hits. Our guide to LLM cost optimization walks through that setup. Teams on the API should also note that Opus 5.5 ships with "preserved thinking", an anti-distillation safeguard for accounts created on or after August 31, 2026, and thinking can no longer be switched off. Pipelines that read or rewrite reasoning output need a test run before migration.

Sources: Anthropic, Introducing Claude Opus 5.5, OpenAI, Introducing GPT-6 Sol and Luna, OpenAI, Better prompt caching for GPT-6, The Decoder, TechCrunch