GPT-6 Astra vs Claude Fable 5.1 vs Gemini 3.8 Flash: Frontier Math, Coding, and Cost

Updated 28 September 2026. This comparison was first published on 5 September and has been rewritten with the current scores. Since then, Epoch AI has published Gemini 3.8 Flash's FrontierMath results, which were missing the first time, and Artificial Analysis has revised its Intelligence Index, moving all three models down. The charts are drawn from the data as it stands today.

Both labs have kept tiering their line-ups since. Our briefing of 22 September reported teams routing routine queries to open-weight models while reserving flagships such as Claude Fable 5.1 and GPT-6 Astra for hard reasoning ¹⁴, and our briefing of 23 September noted that Amazon Web Services added GPT-6 Sol and Claude Opus 5.5 to Amazon Bedrock ¹³.

Commentary has split between capability and cost. Ethan Mollick called GPT-6 Astra ascending in Nethack on its third attempt a startling achievement ⁶. Gemini 3.8 Flash sits apart from both flagships: listed at $1.50 per million tokens blended against $20.00 for the other two, it competes on price and speed rather than on frontier scores ³. Choosing between them means separating the enthusiasm from what the benchmarks measure.

Where the narrative and the numbers disagree

Bindu Reddy argues that Fable 5.1 is far stronger than Claude Opus 5.5 on complex codebases, and under-hyped only because of its price ¹⁰. Against the other two models here, the coding measures bear that out, by margins rather than a gulf. On Artificial Analysis's Coding Index, Claude Fable 5.1 at max scores 81.6, GPT-6 Astra at high 77.1 and Gemini 3.8 Flash at high 76.3 ³. On Terminal-Bench 2.1 the order is the same and the gaps smaller: Fable 5.1 at 91.4%, Astra at 89.9% and Gemini 3.8 Flash at 87.6%.

Mathematics reverses the order, and it is where the new data changes the picture most. On FrontierMath Tier 4, GPT-6 Astra at high leads with 97.6%, Claude Fable 5.1 at max scores 87.8%, and Gemini 3.8 Flash at high - unscored when this article first ran - comes in at 22.0% ². Fable 5.1 also trails its predecessor here: Claude Fable 5 at max scored 90.2%.

Repository repair is closer than any ranking suggests. On DeepSWE, GPT-6 Astra at xhigh scores 74.1%, Gemini 3.8 Flash at high 73.8% and Claude Opus 5 at max 73.6% ¹. Datacurve publishes run-to-run intervals for these scores - ±2.9, ±1.4 and ±3.9 points - and all three overlap, so the order among them is not evidence (what the DeepSWE intervals show). What does separate them is cost: Gemini 3.8 Flash reaches its score at a mean of $2.36 per task, against $4.43 for Astra and $11.84 for Opus 5. On this benchmark, the cheapest of the three repairs code about as well as the dearest.

Formal reasoning and composite intelligence

Artificial Analysis's composite measures put Anthropic and OpenAI close together, with Google behind. Its revised Intelligence Index scores Claude Fable 5.1 at max at 53.4, GPT-6 Astra at max at 52.7 and Gemini 3.8 Flash at high at 40.9 ³; in the first version of this article, the same three read 56.8, 54.7 and 47.1. On Humanity's Last Exam, a broad test spanning many academic subjects, Fable 5.1 at max scores 59.1%, Astra at max 54.7% and Gemini 3.8 Flash at high 47.8%. Among open-weight models, Kimi K3, released under its own licence, comes within a point of Gemini 3.8 Flash.

On FrontierMath Tiers 1–3, the less extreme problems, GPT-6 Astra at max leads with 93.7%, ahead of Claude Fable 5.1 at max at 90.2%, with Gemini 3.8 Flash at high on 68.4% ².

Latency, throughput, and API operating budgets

Price and speed separate them more than any score. Claude Fable 5.1 and GPT-6 Astra are both listed at $20.00 per million tokens blended, and Gemini 3.8 Flash at $1.50 ³. For scale, our analysis of open-weight costs found that $2,111 - a month of one rented H100 - buys about 106 million tokens on either Astra or Fable 5.1, so a self-hosted model would need to sustain 40 tokens a second to be worth the machine.

Gemini 3.8 Flash also generates at 282 tokens a second, against 68 for Fable 5.1 and 63 for Astra ³. Open-weight alternatives are priced near it - GLM-5.2 at $2.15 per million and DeepSeek V4 Flash 0731 at $0.66 - though Artificial Analysis currently publishes no speed for either. For voice and other interactive uses, that speed matters more than a few points of benchmark score.

What to watch

Several critical data points remain unresolved across these models. Specifically, Claude Fable 5.1 has no published DeepSWE score. Datacurve's evaluation suite currently lists Claude Fable 5 at 69.9% at xhigh and Claude Opus 5 at 73.6% at max, but has yet to publish verified runs for Fable 5.1 ¹. If an independent run puts Fable 5.1 ahead of that tied cluster at a reasonable cost per task, it would be the strongest coding case of the three.

The other open question is mathematics. Gemini 3.8 Flash's 22.0% on FrontierMath Tier 4 is the one place in this comparison where it is not competitive. How far a later Flash release closes that gap will decide whether the $20.00 models keep a technical case beyond the hardest reasoning.

Verdict by use

• Choose Claude Fable 5.1 if your work is complex code, terminal-driven agents, or broad knowledge work where it leads Humanity's Last Exam, and your costs can carry $20.00 per million tokens blended.
• Choose GPT-6 Astra if your work is formal mathematics, where it leads both FrontierMath measures, and a measured 63 tokens a second is fast enough.
• Choose Gemini 3.8 Flash for interactive agents, voice tools or cost-constrained code repair: it matches GPT-6 Astra on DeepSWE within the margin of error at a mean $2.36 per task, where Fable 5.1 has no published score yet, and generates at 282 tokens a second. Look elsewhere for frontier mathematics.