GPT-6 Astra vs Claude Fable 5.1: how do they actually compare?
Same $10 and $50 list price, the same score of 53 on Artificial Analysis, and a bill that differs by two and a half times because one model writes three times as much.
GPT-6 Astra and Claude Fable 5.1 are the two most expensive general models on sale, they carry the same list price of $10 per million input tokens and $50 per million output tokens, and as of 11 September 2026 they score exactly the same 53 on the Artificial Analysis Intelligence Index. On the three published benchmarks where both appear, the gap between them is smaller than the error bars, and the real difference is how many tokens each one burns to get there.
That last part is the whole article, so here it is up front. Fable 5.1 cost $13,128.86 to run through Artificial Analysis's evaluation suite. Astra cost $5,324.10 for the same suite, at the same list price, because it emitted 60 million output tokens against Fable's 190 million.
Same sticker. Two and a half times the bill.
So which one should you buy? On the published evidence that is a question about your workload, not about the models.
The price cards are close to identical, and the fine print is not
Both labs publish their rates. Anthropic lists Fable 5.1 at $10 in, $50 out; OpenAI lists gpt-6-astra at $10 in, $50 out. Underneath that, three things diverge.
| GPT-6 Astra | Claude Fable 5.1 | |
|---|---|---|
| Input, per 1M tokens | $10.00 | $10.00 |
| Output, per 1M tokens | $50.00 | $50.00 |
| Cached input read | $1.00 | $0.25 |
| Cache write | $12.50 | $12.50 |
| Long prompts | 2x input and 1.5x output above 272K tokens | full 1M window at standard rates |
| Context window | 1,050,000 | 1,000,000 |
| Max output | 128,000 | 128,000 |
| Knowledge cutoff | 30 April 2026 | June 2026 |
| Input modality | text, image | text, image |
| Output modality | text | text |
Cache reads are where the two diverge most, and it is not a rounding error. Anthropic charges 2.5% of the input rate for a cache hit on Fable 5.1 against the usual 10% multiplier, so a cached token costs a quarter of what the same token costs on Astra (a quarter, not a tenth — the discount is on Anthropic's side here). If you are running an agent that replays a large stable prefix on every turn — which is most agent harnesses — that is the line item you will feel.
OpenAI's long-context rule cuts the other way for anyone actually using the million. Astra's docs are explicit: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request." Anthropic's are equally explicit in the opposite direction, saying a 900k-token request "is billed at the same per-token rate as a 9k-token request".
So the model with the bigger advertised window is the one that charges you for using it.
The benchmark scores are a tie inside the error bars
Three published evaluations have both models in them as of 11 September 2026.
| Benchmark | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Artificial Analysis Intelligence Index v4.3 | 53 (max effort) | 53 (max effort) |
| Terminal-Bench 4.0, resolution rate | 58.2% ± 2.8 (rank 1) | 57.9% ± 3.8 (rank 2) |
| Terminal-Bench 4.0, cost of the run | $3,267.18 | $6,243.50 |
| Cost to run the full Intelligence Index | $5,324.10 | $13,128.86 |
| Output tokens across the Index | 60M | 190M |
| Output speed, tokens per second | 54 | 67 |
| Vending-Bench 2, one simulated year | $15,514.70 ± $1,074 (1st) | not in the published top ten |
Artificial Analysis has them level at 53, out of 146 models evaluated, on an index that is itself a weighted blend of ten evaluations including Terminal-Bench v4.0. Terminal-Bench's own leaderboard (which scores a model-and-agent pair, not a model: Astra ran under Codex, Fable under Claude Code) puts Astra at 58.18% and Fable 5.1 at 57.88% over 330 trials each — a gap of one task out of 330, with confidence intervals of ±2.8 and ±3.8 points sitting on top of it. Calling that a win for anyone is a misreading of an error bar, which is roughly the point we made in Astra and Fable 5.1 tie at 53. The interesting number is what happened to the Flash models when the benchmark changed.
Vending-Bench 2, where models run a simulated vending business for a year, is the one place a real gap shows. Andon Labs has Astra first at $15,514.70 (averaged over five runs) against a human baseline the lab estimates at roughly $63,000; Fable 5.1 does not appear in the ten rows the leaderboard shows by default. We went through that result in The best AI shopkeeper made $15,515, and a competent human would have made $63,000.
Verbosity is the price difference nobody prints
Here is the number we computed, and it is the one we would put on a slide. Fable 5.1 generated 190 million tokens running the Intelligence Index. Astra generated 60 million. Artificial Analysis rates Fable "4 out of 4 units for Verbosity" and Astra 2 out of 4, and the resulting bills — $13,128.86 against $5,324.10 — differ by a factor of 2.5 at an identical per-token rate.
Terminal-Bench shows the same shape on a smaller job. Fable's run cost $6,243.50 and Astra's $3,267.18, roughly 1.9x, for three tenths of a percentage point of accuracy. Fable also took longer per trial, 3,893 seconds against 2,796, despite being the faster model on tokens per second.
Which is the thing to take away if you take away one thing. Per-token price tells you almost nothing about what a model costs to use, and any comparison that stops at the price card is comparing the wrong number.
The case for Fable, which the averages hide
Now the other side, because the averages are unkind to Anthropic here and the underlying numbers are not.
On Terminal-Bench 4.0, Fable 5.1's pass@2 is 0.700 and its pass@5 is 0.788. Astra at max effort scores 0.649 and 0.712 on the same measures. Give each model a second attempt at a failed task and Fable pulls ahead; give it five and the gap is more than seven points. A model that explores widely and expensively is exactly the model you want when you can retry cheaply and keep the run that worked, and a single-shot resolution rate hides that entirely.
Anthropic also aims Fable at a different job. Its own model guide tells you to start with Opus 5 and reach for Fable 5.1 only "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short". That is a lab telling its customers not to buy the expensive one by default (Anthropic's own model guide, not our gloss), which is unusual enough to be worth quoting.
What the numbers do not say
None of these benchmarks measure the thing you are probably buying: whether the model is pleasant to work with in a harness you already have, on your codebase, with your tools. Terminal-Bench is 330 trials of terminal work. Vending-Bench is a simulation with a sales equation that can be gamed. The Intelligence Index is ten evaluations averaged, and averaging is how two very different models arrive at the same 53.
They also say nothing about rate limits, availability, latency under load, or the fact that Astra's knowledge cutoff is the earlier of the two. And they say nothing about what either model refuses.
What would change this answer
A new version of either model resets everything on this page, and one of them probably will. Watch for two things in particular.
First, the Intelligence Index version number. Scores are quoted against v4.3; a v4.4 with a new evaluation in the blend can move a model several points without the model changing at all. Second, verbosity. If Anthropic ships a Fable revision that answers in 60 million tokens instead of 190 million, the cost gap closes overnight and this comparison inverts, which we would rate as likelier than a change in the accuracy gap.
Until one of those happens, the honest answer to "which is better" is that they are the same model at different verbosities, and the one you should buy is the one whose bill you can predict.
