OpenAI shipped GPT-6 Astra on September 3rd — their biggest training run ever, 100,000+ GPUs at the Stargate site, and enough confidence that Greg Brockman used the phrase "generational leap". It lands directly opposite Anthropic's Claude Fable 5.1. Both flagships now cost exactly $10 per million input tokens and $50 per million output — which makes this the first frontier-model matchup where the sticker price tells you nothing.
I build and run LLM systems for a living — agents, voice AI, RAG — so here's the comparison I actually wanted to read: not "which one is smarter", but which one to reach for, for which job, and what each will really cost you.
The spec sheet, side by side
| GPT-6 Astra | Claude Fable 5.1 | |
|---|---|---|
| Context window | 1.05M tokens | 1M tokens |
| Max output | 128K | 128K |
| Input / output per 1M | $10 / $50 | $10 / $50 |
| Cache reads per 1M | $1.00 ($2.00 past 272K) | $0.25 |
| Long-context surcharge | 2× input, 1.5× output past 272K | none |
| Knowledge cutoff | April 30, 2026 | June 2026 |
Two rows in that table matter far more than the headline price, and they're the two nobody puts in the launch tweet.
Cache reads are the real price of an agent
An agent loop resends its whole history every turn. After the first turn, most of what you're billed for is cache reads — the system prompt, the tool definitions, everything the model already saw. On my own agent workloads, cached input routinely outweighs fresh output several times over.
At $0.25 vs $1.00 per million cached tokens, Fable 5.1's cache reads are 4× cheaper — and past 272K of context, Astra's cache reads double to $2.00 while a surcharge kicks in on everything else. Fable has no long-context penalty at all. If your workload is a long-running agent with a fat context, the identical sticker prices diverge fast, in Anthropic's favour.
But there's a twist in the other direction: Astra is dramatically more token-efficient at reasoning. Artificial Analysis measured $1.67 per completed task for Astra against $3.76 for Fable 5.1 at max effort — Astra simply spends fewer tokens getting to its answer. So the honest cost picture is workload-shaped: short, hard, self-contained tasks favour Astra; long-context agent loops favour Fable. Run your own numbers — my LLM cost estimator exists for exactly this kind of arithmetic.
The benchmarks split cleanly — and that's the useful part
Neither model wins. They win different things, consistently enough that the split looks real rather than noisy:
Astra's column: math and hard science (97.6% on FrontierMath Tier 4 vs 87.8; 96.0% GPQA Diamond vs 93.7), computer use (72.6% OSWorld 2.0, 92.7% ScreenSpot-Pro), professional artifacts — slides, spreadsheets, CAD — and cybersecurity, where it's the first model OpenAI has rated at its own "critical" threshold. It also edges the disclosed software-engineering runs (74.1 vs 67.4 on DeepSWE v1.1, 57.7 vs 55.8 on Terminal-Bench 4.0).
Fable's column: overall reasoning depth — it ranks first of 202 entries on the Artificial Analysis Intelligence Index at 66, with Astra eighth at 61 — plus tool-augmented research (65.0 vs 57.2 on Humanity's Last Exam with tools), and the fresher knowledge cutoff by two months.
Worth saying plainly: the coding picture is genuinely contested — different benchmark suites crown different winners, which usually means the models are close enough that your codebase is the only benchmark that matters.
The differences that bite in production
Benchmarks aside, these are the things that changed how I'd integrate each:
- Astra ships gated. At launch it's off by default for API organizations — an enterprise admin has to enable it, and the cybersecurity-heavy capabilities rolled out first through OpenAI's Daybreak program. Plan for procurement friction, not just an API key.
- Fable's API has opinions. Thinking is always on (you control depth with an effort setting, not a toggle), forced tool choice returns a 400 — you prompt for the tool instead — and conversation histories are effectively append-only. If your harness edits old turns, it needs rework before Fable 5.1.
- Refusals are a real code path on both. Frontier models now decline requests their safety classifiers flag — with structured categories, not just errors. Anthropic ships server-side fallbacks that re-route a refusal to a smaller model automatically; if you build on either flagship, handle the refusal path on day one, not after the first production incident.
What I'd actually pick
- Long-running agents with big contexts (the kind that dominate my own work): Fable 5.1 — the cache economics and the absence of a long-context surcharge compound on every turn.
- Math-heavy, science-heavy, or short one-shot hard problems: Astra — better scores and lower measured cost per task.
- Computer-use automation and document/CAD generation: Astra, clearly.
- Anything security-sensitive: test both, but budget review time — Astra's cyber capability is precisely why it ships gated.
- Most products: neither, most of the time. The flagship tier is for the hardest 5% of your requests; the mid-tier models from both labs handle the rest at a fraction of the cost. Route, don't default.
The genuinely new thing about this matchup isn't either model — it's that pricing has converged to the point where architecture, not list price, decides your bill. That's a better world for people who read the pricing page past the first row.
Sources: OpenAI's GPT-6 Astra announcement · Axios · CNBC on the cyber threshold · Artificial Analysis head-to-head · DataCamp's benchmark breakdown. Numbers are as published in the week of launch — frontier benchmarks move; check the live sources before quoting.