OpenAI released GPT-6 Astra on September 3, 2026. Anthropic released Claude Fable 5.1 two days earlier, on September 1. Both companies are now pitching the same idea from opposite directions: an AI that can carry a task from start to finish without much hand-holding.
The specs alone don’t settle which one to use. The workflow does.
What GPT-6 Astra Actually Changed
OpenAI built Astra around computer use. Greg Brockman, the company’s president, called it a “generational leap” at the launch briefing and said it could eventually mark the arrival of general-purpose AI capability. That’s a big claim, and it’s worth reading skeptically — but the underlying feature set backs up the framing to some degree. OpenAI reports the model scored 72.6% on OSWorld, a benchmark for navigating real desktop and browser environments, and describes Astra as its strongest computer-use model to date.
In practical terms, that means Astra can click through interfaces, fill out forms, and move between applications with less step-by-step instruction than earlier models needed. OpenAI has also flagged that the jump in capability comes with a “Critical” cyber-risk rating internally, which is why the rollout started with an application-only program for cyber defenders before reaching ChatGPT Plus, Pro, Business, and Enterprise tiers.
Pricing sits at $10 per million input tokens and $50 per million output tokens.
What Claude Fable 5.1 Actually Changed
Anthropic’s update runs in a different direction. Fable 5.1 ships with a 1-million-token context window and up to 128,000 tokens of output, and its “adaptive thinking” stays on by default — the model adjusts how much internal reasoning it applies based on how hard a task looks, without a manual toggle. You can steer the intensity through a reasoning-effort setting, but you can’t turn reasoning off entirely.
The more consequential change is pricing on cached input. Cache reads dropped to $0.25 per million tokens, a quarter of what Fable 5 charged. For a long-running agent that keeps re-reading the same repository or document context hundreds of times in a session, that’s the cost that actually adds up — more than the output pricing does. Anthropic states this can cut typical workload costs by roughly 25% and agentic workflow costs by closer to 45%, though those figures depend heavily on how much context a given workload reuses, and are worth testing against your own usage rather than taking at face value.
Base input and output pricing stayed the same as Fable 5: $10 and $50 per million tokens.
Where Each One Actually Wins
Astra’s edge shows up in live environments. Debugging inside an active codebase, pulling data through a browser, assembling a rough cut in a desktop video editor — tasks that require acting on something outside a chat window play to its computer-use design.
Fable 5.1’s edge shows up in scale and static review. Reviewing an entire codebase for a subtle bug, working through a document that runs into hundreds of pages, holding a multi-file research project in view at once — that’s what a 1-million-token window is actually for. Teams that need to compare Fable 5.1 against other models on a task they already know well can try it through Claude Fable 5.1 directly, without setting up a separate account first.
Neither model generates video natively. Both hand that off to a separate system — Astra by controlling a desktop editor directly, Fable 5.1 by writing the script or shot list upstream of a dedicated video model.
Social listening works similarly for both. Each supports web search that can retrieve publicly indexed posts from platforms like X or Reddit, but neither is a live feed of platform activity — results depend on what’s indexed at request time, so anything load-bearing still needs a check against the platform itself.
Picking Between Them
If the work involves acting inside a browser or a live system with minimal supervision, Astra’s design fits. If the work involves sitting with a large amount of static information — a contract, a repository, a research corpus — and reasoning through it carefully, Fable 5.1’s context window and always-on reasoning fit better.
Most teams doing serious work with both aren’t picking one and discarding the other. A growing pattern is using one model for research and information-gathering, a second for drafting or coding, and comparing outputs on the same brief before committing either to production. Platforms like ChatGoat AI let users run that comparison in one place instead of maintaining separate accounts for each provider.
The newest model isn’t automatically the right model. The task in front of you still decides that.
Related: Why Did Chat GPT, Claude & Grok Fail at the Same Time?
