Claude Fable 5.1 vs GPT-6 Astra: Same Price, Different Bill

Anthropic shipped Fable 5.1 on September 1. OpenAI shipped GPT-6 Astra on September 3. Both list at $10 per million input and $50 per million output.

That is not two labs arriving at the same number by accident. That is a price point, and once the sticker matches, the decision moves somewhere else.

The cache line is the whole story

Anthropic cut cache reads by 75% — $1 per million down to $0.25. Standard tokens did not move.

If you have never costed an agent workload, that sounds like a footnote. It is not. An agent re-reads the same context on every turn: your codebase, your schema, your instructions. Those are cached reads. Anthropic puts the effective saving near 25% for typical use and as high as 45% for agent-heavy applications, roughly $1.40 per task.

Astra’s cached input, per OpenAI’s own model page, is $1 per million. That is the exact number Fable 5 was charging last week. So at the one line that actually moves an agent bill, Fable 5.1 is 4x cheaper, and everything else on the sticker is a tie.

The arithmetic, with my numbers

Let me cook with a workload shape I actually run. A coding agent holds a 50,000-token context (repo map, schema, instructions) and makes 40 tool-call turns before it finishes. That is 2,000,000 cached-read tokens per task, and I am ignoring cache writes and uncached input so the shape stays visible.

Fable 5 cached reads: 2,000,000 at $1 per million is $2.00. Fable 5.1: 2,000,000 at $0.25 per million is $0.50. Astra at $1 per million: $2.00.

Now the part the launch posts skip. At max effort, Fable 5.1 costs about 20% more per task than Fable 5, because it burns roughly 1.7 times the output tokens (Artificial Analysis has the per-effort receipts). Say Fable 5 wrote 20,000 output tokens for $1.00. Fable 5.1 writes 34,000 for $1.70.

Total: Fable 5 at $3.00, Fable 5.1 at $2.20. The cache cut saved $1.50, the extra tokens cost $0.70, you net out 27% cheaper. That is the agent shape, and it is why Anthropic’s 25% to 45% claim holds up on paper.

Flip the shape. Short prompt, 5,000-token context, same 40 turns: 200,000 cached reads. Fable 5 pays $0.20 plus $1.00 of output, $1.20 total. Fable 5.1 pays $0.05 plus $1.70, $1.75 total. That is 46% MORE. Short prompts and long answers get the token increase without the cache relief, and on that shape 5.1 is a downgrade on cost.

That is the thing to check against your own traces before you switch. I laid out how to read these bills in what AI coding agents actually cost.

Effort is a dial, and the dial is the bill

Both models expose five effort levels: low, medium, high, xhigh, max. On Fable 5.1, Artificial Analysis measured cost per Intelligence Index task at $1.70 on low and $6.12 on max, a 3.6x spread, for an index score of 51 versus 57. Six points of index for 3.6 times the money.

Astra at max effort came in at $2.57 per task with an index score of 55, third of 202 models. So on the independent index, Astra is cheaper per task than Fable 5.1 at max and pricier than Fable 5.1 at medium (53, $2.14). Same neighborhood. Not different leagues.

Speed is where Astra gets weird. The same page clocked a 322.48 second time to first token at max effort, against a 3.36 second median for its price tier. That is a five-minute wait before the first token lands… fine for a batch job, not fine for anything a human is watching.

Astra’s numbers are the best I have seen published

98.6% on ARC-AGI-3. 97.6% on FrontierMath Tier 4 v2. 96% GPQA Diamond. 95.9% BenchCAD. 74.1% DeepSWE v1.1. 72.6% OSWorld. The full launch table also carries a Critical cyber rating and 100% on ExploitBench.

Those are OpenAI’s numbers, on launch day. I am reporting them, not confirming them — the independent index is where they get graded by someone who did not build the model. That is not scepticism about OpenAI specifically. It is the same rule that applies to every launch table, including Anthropic’s.

What the three that matter actually test, one line each. ARC-AGI-3 drops an agent into turn-based puzzle environments with no instructions and scores what it figures out. OSWorld V2-Offline scores work across real desktop applications. DeepSWE v1.1 is a 113-task agentic coding test.

Read the fine print on all three. OpenAI ran ARC-AGI-3 through a Responses API setup that keeps reasoning between turns, and The New Stack points out OpenAI had already shown that setup alone can raise the score without touching the model. On DeepSWE, OpenAI’s chart used a 67.4% Fable 5.1 result while the public leaderboard has Gemini 3.8 Flash and Claude Opus 5 at 74%, and Meta reported 75.4% for Muse Spark 1.3 at max. The uncertainty ranges overlap. Nobody leads that board.

And the cyber results reflect Daybreak Blue access, not the production configuration most of us will get. The model you can call refuses exploit discovery. The model that scored 100% on ExploitBench does not.

Access is a spec

Fable 5.1 hit the standard API on day one. Astra went to Daybreak organizations, with everyone else — Plus, Pro, Business, Enterprise, the API, AWS — pointed at “the coming days.”

A model you cannot call is not in your stack. If you are choosing this week, that is not a tiebreaker, it is the decision.

Daybreak, in practice, is OpenAI’s program for organizations already approved for its most capable models. If your company is not on that list, the number you can compute for Astra today is zero, and “coming days” is not a date you can put in a sprint plan. Fable 5.1 has a model string and a bill you can run tonight. I said the same thing about GPT-5.6 Sol in July: a launch chart is not availability.

What 1M context and 128K output actually buy you

Both models take a million tokens in and 128K out. In practice that is a mid-sized repo plus its docs in one prompt, and an output long enough to write the migration, the tests and the rollback in one pass instead of three.

The catch is on the bill. OpenAI prices prompts over 272K input tokens at 2x input and cache rates and 1.5x output for the whole request. Stuff 400K tokens into Astra and the $10 / $1 / $50 line becomes $20 / $2 / $75. The 1M window is real. Using most of it is a different price list.

Who should pick which

Agents over a fixed context, shipping now: Fable 5.1, on the cache math, which you can compute against your own usage today. That is continuous with what I said when Fable 5 came back.

Hard reasoning, and you have Daybreak access: Astra, with the launch table treated as a claim until the independent numbers land.

Short prompts, long answers, max effort: stay on Fable 5 or drop 5.1 down to medium. Interactive tools with a human watching the cursor: not Astra at max effort, not with a five-minute first token.

Coding agents where the leaderboard is the argument: it is a coin flip at 74%, so pick on price and access, which both point the same way this week.

Everyone else: nothing broke on September 1. The frontier moved, the price didn’t, and the thing that changed your bill was a discount on cached reads — not a benchmark.

#TheAIMogul