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

Anthropic opened September with Claude Fable 5.1: the Mythos-class weights finally in general availability, a million tokens of context, and cache reads cut to $0.25 per million. Two days later OpenAI answered with GPT-6 Astra (that is the model's actual name, not a tier): 1.05 million context, 128K out, native computer use, and the distinction of being OpenAI's first model tagged Critical on cybersecurity, which is why the spicy parts arrive through a gated rollout.

Then the bill. Both cost $10 per million tokens in, $50 out. Same shelf, same sticker, to the cent.

The scoreboards could not handle this. llm-stats crowns Astra first, 60.7 to 56.8. benchlm crowns Fable, 82.95 to 81.05. Artificial Analysis splits the intelligence index to Fable and hands the value prize to Astra. Three referees, three sports.

• Coding: a tie on quality. Fable if your repo lives in cache, Astra if you need 128K out.
• Agents and computer use: Astra for OS automation (native, three of four shared automation benchmarks on llm-stats); Fable for orchestration-heavy stacks (benchlm ranks it first of 151).
• Long documents: Astra, 1.05M in and 128K out.
• High-volume reasoning: Astra, 88.8 to 79.4, not close.
• Cost at scale: task-shaped volume goes to Astra ($0.46 vs $0.77 per task); context-shaped volume goes to Fable.

specs

	fable 5.1	gpt-6 astra
released	sep 1	sep 3
price in / out	$10 / $50	$10 / $50
cache reads	$0.25 / m	$1.00 / m
context	1m	1.05m
max output	not published	128k
knowledge cutoff	not published	apr 30 2026
cyber posture	refusals loosened	critical-level, gated

The only boring row is price. Everything else is a decision.

where astra wins

Hard reasoning is not a debate. Benchlm calls its reasoning category the clearest separation in the comparison, 88.8 to 79.4, and OpenAI's card backs the swagger: 95 percent on ARC-AGI-2's public set (the hidden-set prize bar is the honest gate), 97.6 on FrontierMath Tier 4, 96 on GPQA. At low reasoning effort Astra completes tasks for $0.46 against Fable's $0.77, so cheap reasoning at volume is an Astra monopoly.

Computer use leans the same way: the model drives a desktop natively and takes three of the four shared automation benchmarks on llm-stats. Fable keeps orchestration-shaped work (first of 151 on benchlm's agentic lane, Toolathlon-Verified at 77.8), so agents are a split call. But if the job is "click the thing, run the thing," that is Astra's home turf.

where fable wins

The cache. Fable rereads your context at $0.25 per million against Astra's $1.00, and agents do nothing but reread context. Run the month: 300 million input tokens, 60 million out, 270 million cache reads, the way a long-lived coding agent actually works. Fable lands at $1,867.50. Astra, $2,070. Month two, with the context warm and the read ratio wider, the gap swings past $3,000.

That is not a rounding error. That is the difference between an experiment and a line item.

the fine print

OSWorld 2.0 crowned Fable in one lane (77.9) and Astra in another (72.6), while Fable's second lane had it at 41.7. Thirty-six points between two scores of the same model on the same test, and OpenAI's card supplies the 72.6, which means part of the referee's scorecard is the contestant's homework. The 1990s PC magazines went through this exact phase, and the lesson holds: the crown measured the lab, not the machine.

Astra's ExploitBench is a reported 100, and the exploit-shaped capabilities sit behind a trust gate because of it. Fable loosened refusals that had been blocking benign work. Neither is a footnote: if compliance signs one model and the red team wants the other, that is a routing decision like any other.

And the number that would settle the overnight question does not exist. METR has published nothing for either release, so autonomy duration is currently sold on vibes and demo reels.

Default Astra. Fable when the work is a long-lived agent living on its own context. High-volume reasoning, Astra. Long documents, Astra. Orchestration-first stacks, check Fable's lane, then check the cache column of your bill, because the invoice is the receipt for the whole argument.

The leaderboards only look broken because they keep getting asked the wrong question. Better at what? That is the one they answer straight.

Update (2026-09-10): pricing moved after this piece's data was gathered - Anthropic cut Fable 5.1 token costs by up to 45% on September 1, and OpenAI's GPT-6 Astra launched September 3 at roughly half of GPT-5.6 Sol's rate. The framework above holds; re-run the cost math against current price sheets before committing a budget.