Between September and October, Anthropic and OpenAI renewed their AI model lineups. Anthropic, which makes Claude, launched Fable 5.1 and the 5.5 family (Opus, Sonnet and Haiku). OpenAI, which makes ChatGPT, launched the GPT-6 generation, with the Astra, Sol and Luna models, and has already updated Sol to version 6.1. These models run behind the apps and also reach companies through the API, for use inside their own systems.
What each company launched
Anthropic (Claude)
- Claude Fable 5.1 (September 15): from Anthropic’s highest line, presented by the company as the world’s most advanced model for coding and knowledge work. The same model, with different safeguards, exists as Claude Mythos 5.1, available only to approved organizations working in cybersecurity and the life sciences.
- Claude Opus 5.5 (September 22): opens the 5.5 family, costs 20% less per token than Opus 5 and, in the tests Anthropic published, comes out ahead of Fable 5.1 in agentic coding, computer use and knowledge work.
- Claude Sonnet 5.5 (September 28): faster and cheaper than Opus, meant for well-defined everyday tasks, bug fixing and creating documents, presentations and spreadsheets.
- Claude Haiku 5.5 (October 7): the small model, Anthropic’s fastest and cheapest, built for high volumes such as summaries, classification and real-time customer support.
OpenAI (ChatGPT)
- GPT-6 Astra (September 3): OpenAI’s most capable model, which the company presents as the most intelligent in the world, with a focus on computer use, coding and science.
- GPT-6 Sol and GPT-6 Luna (September 22): faster and cheaper versions, trained with methods similar to Astra’s. Luna is the cheapest, built for everyday work at scale.
- GPT-6.1 Sol (September 29): an update to Sol that, according to OpenAI, comes close to Astra in agentic coding, computer use and professional work at one fifth of the price.
- GPT-6 Cyber (November 5): a model for security research, available for now to a group of approved customers of the Daybreak Red program.
Side by side
By API price, the two lineups fall into tiers, and in three of them the price per token is identical:
- Most expensive tier: Claude Fable 5.1 and GPT-6 Astra, at $10 per million input tokens and $50 per million output tokens.
- Middle tier: Claude Sonnet 5.5 and GPT-6.1 Sol, at $2 and $10.
- Cheapest tier: Claude Haiku 5.5 and GPT-6 Luna, at $0.10 and $0.50 (for Haiku, on requests of up to 100,000 tokens).
Claude Opus 5.5 sits in between, at $4 and $20, with no direct equivalent at OpenAI. One difference that matters for agents is reused context (cache): in the most expensive tier, Fable 5.1 charges $0.25 per million tokens read from the cache, and Astra charges $1.
In the apps
The rollout in the apps was different. In ChatGPT, Astra was released to the Plus, Pro, Business and Enterprise plans; Sol and Luna debuted in ChatGPT Work and Codex on the paid plans, and Free and Go users got Luna in the desktop app. In Claude, Fable 5.1 is in the app, in Cowork and in Claude Code, and Opus 5.5 arrived with higher usage limits on the Pro, Max, Team and Enterprise plans.
Who comes out ahead
Each company publishes tests in which it leads. Anthropic says Opus 5.5, at its default effort level, beats GPT-6 Astra on FrontierCode, a coding test, at about one fifth of the cost per task. OpenAI says GPT-6.1 Sol beats Opus 5.5 by 2.2 points on AutomationBench, which measures business workflows, at about a third of the cost. The numbers come from whoever sells the model, in settings each one chose: they are a clue, not a verdict.
What this changes for companies
- Choose by task, not by brand: the cheapest models were built for volume work such as summaries, classification and customer support, and cost one hundredth of the most expensive per input token. The most expensive are for the hardest work.
- Compare the cost per task: each model spends a different number of tokens, and both let you adjust the effort level, which changes the result and the bill. A sample of your real tasks, judged by the same criteria, says more than any ranking.
- Make switching models easy: in little more than a month, both companies renewed their lineups. A system that switches models through configuration, with tests that measure quality at each switch, makes the most of new releases without redoing the project.
- Keep the usual care: with customer data, check in the contract how the provider handles and stores information, and keep human review, logging and stop criteria, the safeguards we apply in AI and automation projects.
GVolo teamPublished November 6, 2026