Astra and Fable 5.1 Share a Rate Card, Not a Bill

Same 10 in and 50 out per million tokens. One burns 3.3x the tokens per index task and bills 2.4x. One cache line sits 4x apart.

OpenAI raised GPT-6 Astra's price 2.5x over GPT-5.6 Sol and, in the same launch post, told customers it costs less. Anthropic held Claude Fable 5.1 at Fable 5's price and, by Artificial Analysis's measurement, it costs 20 percent more per task than Fable 5 and 2.4x more than Astra. Both statements are true, and they are true for the same reason: at the frontier, the price per token is a tie, and the bill is decided by how many tokens each model spends to finish a job.

That is the comparison this issue runs. Not the system cards, not the cyber gating (we covered OpenAI's pacing framework on Aug 20). This one is about the two rate cards, the two token budgets behind them, and the one mechanism inside each model that explains its budget.

Line for line, one line differs

Fable 5.1 shipped on Sept 1, Astra on Sept 3. Put the two published rate cards side by side and the input, output, and cache write prices match to the cent. The cache read does not: Anthropic priced a Fable 5.1 cache hit at 0.025x the input rate, 0.25 per million. OpenAI kept its standard 0.1x, 1.00 per million. Then there is a rule Anthropic does not have: any Astra request above 272K input tokens is billed at 2x input and cache rates and 1.5x output, for the whole request.

Two footnotes on the sticker. OpenAI's 2.5x is against Sol's 4 and 20 rate, which OpenAI itself called promotional. And Anthropic's cache cut is the whole of its price change: fresh input, output, and cache writes did not move from Fable 5.

Same price per gallon, different miles per gallon

Artificial Analysis runs every effort setting of a model through its Intelligence Index and records the output tokens the run consumed. Read those totals as fuel burn. Astra at max effort burned 49M tokens for a score of 55 on the v4.2 index. At low effort it burned 5.4M for 49. Fable 5.1 at max effort burned 160M for 57 on the same index. On the earlier v4.1 run, its five effort levels spanned 13.1M to 143.7M output tokens, an 11x range, for scores of 58 to 66.

Two ladders, two philosophies. Astra's is flat and thin: 4.1x in cost from low to max for six points of score, and 49M tokens at the top. Fable 5.1's is steep: 160M tokens at the top, and on the v4.1 run the last point, from xhigh to max, cost 1.04 per task on its own. Anthropic's own comparison makes the direction explicit: Fable 5.1 spends roughly 1.7x the output tokens of Fable 5 to gain four index points, which is why its per-task cost rose 20 percent while its rate card stayed flat.

Where Artificial Analysis put them on the same road

On the Sept 3 v4.1.1 run, before the index was revised, Artificial Analysis scored Astra at 61, level with GPT-5.6 Sol and five points behind Fable 5.1 at 66. Astra used about 10 percent fewer output tokens than Sol at max effort, yet cost 75 percent more per task, because the price rose 2.5x. On intelligence per token, Astra defines a new frontier. On intelligence per unit of cost, it sits behind its own predecessor. The revised v4.2 index keeps the ranking and adds the price: Fable 5.1 first at 57 and 6.12 per task, Astra second at 55 and 2.57, a 2.4x gap on identical rate cards.

The coding index is a different story. In the Codex harness, Astra scores 67 on the Coding Agent Index, level with Fable 5 and Opus 5 in Claude Code, at less than half the per-task cost of Fable 5. It got there by using one third of Sol's tokens and one fifth of Opus 5's at xhigh. Fable 5.1 in Claude Code leads that index at 70. So the picture is: Fable 5.1 is ahead on both indices, and Astra reaches the previous frontier at a fraction of the token bill.

Notes instead of summaries, and shorter thinking

OpenAI's launch post names the mechanism. Long agent sessions used to survive a full context window by compaction: summarize the transcript, throw the original away, continue from the summary. Each compaction can drop why a fix failed or how a component behaves. With Astra in Codex, the model keeps notes across context windows and earlier windows stay searchable, so requirements and test results from previous windows can be retrieved without being compressed into one summary again. It is experimental, enabled in the Codex config, and OpenAI says it becomes the default for Astra in the coming weeks.

The second lever is inside the reasoning itself, and OpenAI states it as a cost rather than a feature. Its evaluations found Astra's written reasoning harder to monitor than Sol's, and it attributes that to Astra's greater control over what it writes on simpler tasks and its ability to solve problems with fewer written steps. That is the same trait from two angles. The token you are not billed for is also the token the monitor did not get to read. OpenAI says the decline is taken seriously and that monitorability stays a research priority.

Spend more tokens, make re-reading nearly costless

Anthropic made the opposite bet and priced for it. Fable 5.1 writes more: 1.7x Fable 5's output tokens at max effort. Its thinking is always on, with effort as the dial. What Anthropic changed was the cost of the tokens an agent re-reads. Artificial Analysis estimates the cache cut removes about 1.40 per index task, concentrated in the agentic evaluations where most input is cache reads; without the cut, Fable 5.1 max would cost about 5.16 per task instead of 3.76. Anthropic's own estimate is 25 percent lower cost on typical workloads and up to about 45 percent on highly agentic ones.

One more line from Anthropic's pricing page matters for any cross-lab token comparison: Claude 4.7 and later models use a newer tokenizer that produces approximately 30 percent more tokens for the same text. Artificial Analysis counts tokens as each provider bills them, so a Fable token and an Astra token are not the same unit of text. The cost-per-task figures already absorb this; the raw token ladders above do not.

And the fallback. Artificial Analysis evaluated Fable 5.1 with Anthropic's default server-side fallback, which routes safety-flagged requests to Opus 4.8 or Opus 5. Fallback served about 4 percent of output tokens across the index. The 66 is a system score, not a single-checkpoint score. OpenAI's launch table makes the same point from its side: the Fable numbers it reports for ScreenSpot-Pro and ExploitGym come from Mythos, which OpenAI describes as Fable with fewer safeguards.

The launch post sells cost per task, and prints where it loses

Read the comparison table at the bottom of OpenAI's post and count how the wins are phrased. Terminal-Bench 4.0: 57.9 versus Fable 5.1's 55.8, at approximately 63 percent lower estimated API cost per task. Terminal-Bench Science: 64.6 versus 52.6, at approximately 31 percent lower cost. BenchCAD: 95.9 versus 84.3, at approximately 86 percent lower cost. Agents' Last Exam: 59.3 versus Opus 5's 55.5, with approximately 65 percent fewer output tokens. The headline metric OpenAI chose for this launch is not the score. It is the score divided by the bill.

A launch post that prints the competitor ahead on the aggregate index and wins the argument on cost is a post written for a customer who has already stopped counting tokens. That customer exists because both labs made the price per token identical.

Where the two rate cards diverge for an agent

Pure arithmetic from the two rate cards, nothing measured. Picture an agent loop that re-reads one cached prefix every turn and writes nothing new into it. The prefix is a repo plus tool schemas plus transcript.

So the rate cards are not identical after all. They are identical for a chatbot. For an agent that carries a large prefix, Anthropic's card costs 4x less to re-read below 272K and 8x less above it, and Astra then has to win the whole difference back on the output side by writing less. Per the Codex measurements above, it often does. Per the Intelligence Index, it does not yet.

What the mileage label leaves out

Against Astra. Artificial Analysis records regressions next to the gains: a drop of about 80 Elo on GDPval-AA v2, and 2 to 3 point declines on tau-cubed Banking, SciCode, and long-context reasoning. Hallucination on AA-Omniscience fell from 92 percent to 51 percent at max effort, which is a halving and still a coin flip. The public model refuses advanced cyber work such as proof-of-concept exploits; the less restricted version routes through OpenAI Daybreak in the coming weeks.

Against Fable 5.1. Its v4.2 scores below max effort are marked by Artificial Analysis as estimates pending an independent run, so we used the max row and nothing under it. Its per-task cost is the highest the firm tracks. Its launch bench numbers, like every lab's, are its own harness choices. And the fallback means the fastest way to move Fable on a table is to change what the router sends to Opus.

Against this comparison. The two token ladders sit on different index versions because Artificial Analysis revised the index between the two launches. Cost per task on a benchmark is a proxy; your prefix size, cache hit ratio, and effort setting decide your bill, not the index's.

For a decade the frontier race was parameter count. For the last two years it was price per million tokens. Both labs have priced the token identically and moved the fight to how many of them a task takes. The sticker on the window is the same. Read the odometer.