Opus 5.5 and GPT-6 Sol Land on the Same Day, and Both Bet on Price
Anthropic and OpenAI released new flagship-class models within hours of each other. The benchmarks are close; the price cuts are the point.
On 22 September, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna. Two of the biggest labs shipping on the same day is not new, but this time both launches led with the same message: you will pay less for more.
Opus 5.5 is the first model in Anthropic's Claude 5.5 family. It costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, with cache reads cut from $0.50 to $0.20. Anthropic says that works out to about 40 per cent less on typical workloads, and that output is more than 30 per cent faster. It is available on Anthropic's platform as well as AWS, Google Cloud and Microsoft Azure, with a one-million-token context window listed in Anthropic's model documentation.
OpenAI's GPT-6 Sol costs $2 and $10 per million tokens, and the small GPT-6 Luna just $0.10 and $0.50, roughly half the price of the GPT-5.6 models they replace. Both have a 1,050,000-token context window and can write up to 128,000 tokens in one response, and OpenAI told VentureBeat the prices are permanent rather than promotional, DataNorth reported.
Reading the benchmark tables
Anthropic's claim is that Opus 5.5 performs at the level of Claude Fable 5.1, its larger and pricier top model, on most work. Its launch post backs that up with its own numbers: 66.4 per cent on Terminal-Bench 4.0, against 55.8 per cent for Fable 5.1 and 52.3 per cent for Opus 5; 40.0 per cent on AutomationBench against 31.4 per cent for Fable 5.1; and 81.8 per cent on OSWorld 2.1 for computer use. Fable 5.1 costs $10 and $50 per million tokens, two and a half times the price.
OpenAI says Sol beats both GPT-6 Astra and Claude Opus 5 on business task automation. Note the comparison: Opus 5, not Opus 5.5, which did not exist when OpenAI's charts were drawn. On the same benchmark Anthropic reports a large jump for Opus 5.5, so OpenAI's chart was out of date within hours.
The usual caveats apply. Every number here is self-reported, measured with each lab's own harness and settings. Anthropic's results depend on its effort parameter, which now runs from low to max with medium as the default. A model run at max effort can be both better and more expensive than the headline price suggests.
The real competition is cost per finished task
Strip away the charts and both companies are making the same argument: frontier-level capability is getting cheaper, and quickly. Anthropic is effectively offering Fable-class results at Opus prices; OpenAI is offering a Sol that it says beats its own flagship on business automation at a fifth of Astra's price.
For anyone building agents, list prices per token are now the least interesting number. What matters is how many tokens a model burns to finish a task, how often it finishes it, and how much caching can absorb. Anthropic's steep cut to cache reads is aimed squarely at agent workloads that resend long contexts many times; OpenAI's Luna is aimed at the extraction and summarisation jobs that run millions of times a day.
Anthropic also put weight on safety: it reports its best scores yet on an automated behavioural audit of nearly 2,000 scenarios, and an 85 per cent drop in attempts to get around boundaries compared with Opus 5. Most cybersecurity requests are routed to an older model unless the customer is in its Cyber Verification Program. That is a material limitation if you build security tools, and worth testing before you migrate.
What to watch
- Sonnet and Haiku 5.5. Anthropic says both arrive in the coming weeks. If Sonnet 5.5 lands near Sol's price, the mid-tier will become the most contested part of the market.
- Undocumented variants. DataNorth notes that Sol Pro and Luna Pro appeared on OpenRouter at launch without documentation. Expect a tiered lineup to follow.
- Independent evals at matched effort. Until someone runs both models through the same harness at comparable cost, the ranking is still marketing.
Sources