Quick disambiguation before anything else: this is about OpenAI’s GPT-6 Astra, not Google DeepMind’s Project Astra (the real-time multimodal assistant built on Gemini). Two unrelated products currently sharing a name.
With that out of the way: GPT-6 Astra (OpenAI, released September 3, 2026) and Claude Fable 5.1 (Anthropic) are the two most direct flagship-vs-flagship releases on the table right now. Skipping the benchmark leaderboard entirely — that’s a race with a new “winner” every few weeks and it tells you almost nothing about how to actually use either model. What’s more interesting is how differently each one is built to work, and what that costs per token.
Two different bets on what “capability” means
GPT-6 Astra is positioned around raw, general-purpose intelligence pushed as far as OpenAI is willing to ship it. Per OpenAI’s own published system card, it’s the first OpenAI model classified as hitting the “critical” cybersecurity threshold under the company’s preparedness framework — meaning it can find and exploit unknown vulnerabilities in well-protected systems without step-by-step human guidance. That’s not a benchmark score, it’s a capability classification serious enough that OpenAI shipped the public release with prompt restrictions in cybersecurity-adjacent areas.
Claude Fable 5.1 is built around a narrower, more explicit bet: sustained autonomous work over long horizons. Anthropic’s own framing is “ambitious, long-running, asynchronous work” — a model meant to be set loose on a task for hours or days, plan its own approach, use tools, recover from its own failures, and check in with updates rather than complete a single turn well. The published examples lean into this directly: Stripe reportedly used it to carry out a 50-million-line codebase migration essentially unsupervised, and Anthropic describes Mythos-tier variants running over a week of largely autonomous scientific research.
Same underlying shift — both companies are optimizing for agentic, tool-using, long-horizon behavior rather than single-shot Q&A — but Astra is framed as a raw capability jump and Fable 5.1 is framed as a sustained-autonomy jump. Different products, even when the headline numbers look similar.
Token economics: nearly identical sticker price, different transparency
This is the part worth actually sitting with:
50 — GPT-6 Astra’s price per million input/output tokens.
50 — Claude Fable 5.1’s price per million input/output tokens.
Identical, standard-tier, at least at list price. Where they diverge is in the discount structure and in what’s actually published:
- GPT-6 Astra: cached input drops to **12.50/M, batch/flex mode halves the price to 25, and a documented 1,050,000-token context window with up to 128,000 completion tokens. Prompts over 272K tokens step up to a higher 75 tier for the whole request.
- Claude Fable 5.1: cache reads are cheaper still at $0.25/M (a 75% cut from Fable 5’s cache pricing), but Anthropic has not published a hard context-window number — the public material says it “stays focused across millions of tokens in long-running tasks,” which is a real, load-bearing capability claim without a citable ceiling attached to it.
That gap matters if you’re actually budgeting a workload: Astra gives you a number to plan context-window usage against; Fable 5.1 gives you a behavior claim you have to validate empirically against your own task length.
Identical sticker price, very different bill
Same per-token price doesn’t mean same cost per task, and this is where the two models actually separate. Independent task-level measurements from Artificial Analysis (matched to the same capability score, not a raw leaderboard) show Astra at max reasoning effort using about 27k output tokens per task, against 78k for Claude Fable 5.1 at max effort with fallback — roughly a third as many tokens for the same result.
That efficiency gap shows up directly in the bill: on Artificial Analysis’s Intelligence Index, Astra comes to about 3.76 for Fable 5.1 — Astra matching Fable 5.1’s score at roughly 40% of the cost. The gap narrows but doesn’t close on coding-specific work: on the Coding Agent Index, Astra runs about 60% of Fable 5.1’s per-task cost at max effort.
Worth being precise about why: Astra’s reasoning effort is a dial (light, medium, high, extra high, ultra), and OpenAI has explicitly tuned it to think in fewer tokens per point of capability than its predecessors — Artificial Analysis found it uses roughly a third of the tokens GPT-5.6 Sol needed for comparable scores. Fable 5.1’s higher token spend tracks with its own design goal from earlier in this piece: it’s built to stay engaged across long, multi-step, tool-using sessions, and that sustained engagement costs tokens that a single-shot task doesn’t.
So: identical list price, but if you’re paying per completed task rather than per token in isolation, Astra is currently the cheaper model to run — the opposite of what the matching sticker price would suggest.
What this means if you’re choosing between them
If your workload is bounded and context-window-sensitive — long documents, large codebases you need fully in-context in one shot — Astra’s published 1.05M-token ceiling is the easier number to design around. If your workload is open-ended and agentic — a job you want to hand off and check back on hours later — Fable 5.1’s design center is closer to what you actually need, even without a headline context number to point to.
Neither is objectively “more advanced.” They’re optimized for different shapes of work, priced almost identically, disclosed unevenly, and — despite matching sticker prices — meaningfully different in what they actually cost to run. Anthropic hasn’t published a hard context-window ceiling for Fable 5.1 anywhere in its public materials; if that changes, I’ll update this post with the number.
Sources: OpenAI — GPT-6 Astra, OpenAI — GPT-6 Astra System Card, Deployment Safety Hub, OpenAI — API Pricing, Google DeepMind — Project Astra, Anthropic — Claude Fable 5 and Mythos 5, Anthropic — Claude Fable, Artificial Analysis — Benchmarking GPT-6 Astra, Artificial Analysis — GPT-6 Astra vs Claude Fable 5.1
Opinions are my own.