GPT-6 Astra vs Claude Fable 5.1: Same $10/$50, Different Bill
Two frontier launches three days apart, both listing ten dollars in and fifty out. What separates them is the cache read line, a price cliff at 272K tokens, and which twin you are allowed to run.
Yunseon Hong11 min read
Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. OpenAI shipped GPT-6 Astra on the 3rd. Both list the same headline price: ten dollars per million input tokens, fifty per million out.
That number is the least useful thing on either page. Two requests that cost the same per token can differ by a factor of two once you account for how each vendor prices a cache hit and what happens when your prompt gets long. Below is what the two pricing pages and model pages actually say, checked on 7 September 2026.
I have not run either model against a workload of my own, so nothing here is a benchmark I produced. The prices come off the vendor pricing pages, the capability numbers come from vendor launch tables and from Artificial Analysis, and I have flagged where the two disagree.
What each one is, on paper
| Claude Fable 5.1 | GPT-6 Astra | |
|---|---|---|
| Model ID | claude-fable-5-1 | gpt-6-astra |
| Context window | 1M | 1,050,000 (max input 922,000) |
| Max output | 128K | 128K |
| Knowledge cutoff | Jun 2026 | 30 Apr 2026 |
| Effort levels | low / medium / high / xhigh / max | low / medium / high / xhigh / max |
| Default effort | high | not published |
| Input / output per MTok | $10 / $50 | $10 / $50 |
| Cache read per MTok | $0.25 | $1.00 |
| Cache write per MTok | $12.50 | $12.50 |
| Long-prompt surcharge | none | 2x input, 1.5x output above 272K |
| Restricted twin | Claude Mythos 5.1 | Astra under Daybreak Blue |
The effort ladder is the same five rungs on both, which was not true six months ago. Anthropic’s models overview publishes high as the API default for Fable 5.1; OpenAI’s model page lists the five values without naming a default.
Note the asymmetry in Astra’s window. It advertises 1,050,000 tokens but caps input at 922,000, so the headline figure includes room the model reserves for itself. Anthropic’s 1M is 1M of input, with 128K of output on top.
Where the money actually goes
Two lines separate these price sheets.
The first is the cache read. Anthropic cut cache hits on Fable 5.1 and Mythos 5.1 to 0.025x the base input price, which is $0.25 per million, against the 0.1x every other Claude model uses. OpenAI charges the usual 0.1x, so $1.00. For an agent that replays a large system prompt and a growing transcript on every turn, that one line is most of the bill.
The second is the cliff. OpenAI’s model page states it plainly: prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output, and the multiplier applies to the full request rather than to the tokens above the line. Anthropic’s pricing page says the opposite in one sentence: the full 1M window bills at standard rates, and a 900k-token request costs the same per token as a 9k one.
Run three request shapes through both sheets and the gap opens where you would expect.
A thousand turns of that middle shape is $310 against $400. Not a decision on its own, but it compounds in exactly the workload both vendors spent their launch posts talking about.
One caveat that cuts against reading the table too literally. Anthropic’s docs say the tokenizer introduced with Opus 4.7 emits roughly 30% more tokens for the same text than the older Claude tokenizer, which is why 1M tokens is described as about 555k words where it used to be 750k. Tokens are a vendor-internal unit. A dollar-per-token table tells you what a token costs, not what a paragraph costs, and the only way to close that gap is to run your own prompts through both and compare invoices.
Both offer 50% off for batch. OpenAI adds a Flex tier at the same 50% and a Fast tier at 2x; Anthropic’s fast mode covers Opus 5 and 4.8, not Fable.
The benchmark board, and the parts to discount
Astra wins the categories OpenAI built the launch around. On the numbers collected in Vellum’s breakdown it takes ScreenSpot-Pro 92.7% against Sol’s 76.9%, OSWorld 2.0 at 72.6%, FrontierMath Tier 4 at 97.6% against Fable 5.1’s 87.8%, Terminal-Bench Science 64.6% against 52.6%, and BenchCAD 95.9% against 84.3%.
On cybersecurity the margin depends on which benchmark you read. ExploitBench is 100% against Fable 5.1’s 70%. ExploitGym narrows to 42.4% against 30.4%, with GPT-5.6 Sol in between at 30.3%.
Fable 5.1 holds the composite scores and the long-tail knowledge ones. Artificial Analysis puts Astra at 61 on its Intelligence Index, level with GPT-5.6 Sol, against 66 for Fable 5.1 at max with fallback. On its Coding Agent Index the harness is part of the score: Astra gets 67 in Codex, Fable 5.1 gets 70 in Claude Code. Humanity’s Last Exam with tools goes to Fable 5.1, 65.0% against 57.2%. Terminal-Bench 4.0 is nearly a tie at 57.7% against 55.8%, and Mythos 5.1 sits above both at 60.9% in Anthropic’s own table.
Several things are worth knocking off before you treat any of that as a ranking.
The scores ran at maximum effort unless noted, which is not the setting most production traffic uses. The ExploitGym runs had the standard six-hour time limit removed for both models. Claude’s BenchCAD number came from modified evaluation settings, which is worth holding against the 95.9 against 84.3 above.
The cybersecurity numbers are the ones to hold loosest. They were produced with Daybreak Blue access rather than the default production configuration, and what ships to everyone else is the more limited version with refusal boundaries on advanced cyber work. The scoreboard and the thing you can call are not the same model behavior.
And GDPval, the economically grounded benchmark OpenAI itself built, does not appear in Astra’s launch materials, while Anthropic’s table reports GDPval-AA v2 at 1853 for Fable 5.1 against 1824 for Opus 5 and 1711 for GPT-5.6 Sol. Artificial Analysis measured a drop of about 80 Elo points there for Astra, without naming which model it fell behind. A missing number in a launch post is a number, and this one is missing from the lab that invented the benchmark.
The number I would weigh above any of these is token efficiency. In the Codex harness, Artificial Analysis measured Astra using about one third the tokens of GPT-5.6 Sol at max and one fifth of Claude Opus 5 at xhigh. If that holds on your traffic, it moves the bill further than the per-token rate does.
For the same reason the price sheet is not the whole story either: Artificial Analysis has Meta’s Muse Spark 1.3 at 61 on that Intelligence Index, level with Astra, at a listed $1.25/$4.25.
Three breaking changes if you already call Fable 5
Anthropic’s release notes list these, and each one has bitten somebody’s integration by now.
Forced tool use is gone. tool_choice types any and tool return a 400 on Fable 5.1 and Mythos 5.1; auto and none are unchanged. If you were pinning a pseudo-tool to get schema-shaped output, that pattern is now strict tool use or structured outputs.
Thinking blocks are bound to the model that produced them. Fable 5.1 accepts blocks from Opus 5, Fable 5, Mythos 5 and earlier, but earlier models cannot read Fable 5.1’s, and the API drops one replayed backward.
Editing earlier turns invalidates thinking blocks. On accounts created on or after 31 August 2026, replaying a thinking block after the system prompt, tools, or an earlier message changed returns a 400. That closes a hole people used for context surgery, and it will break any harness that rewrites history between turns.
Output from both models also carries Anthropic’s text watermark, and neither is available under zero data retention without express authorization: 30-day retention is required.
Both labs gated something, in opposite directions
The structural thing about this week is not either model. It is that both labs drew a line around cyber capability, and drew it in different places.
Mythos 5.1 (claude-mythos-5-1) is the same model as Fable 5.1 with the same specifications and pricing, behind invitation to Project Glasswing, with looser safeguards for cybersecurity and life sciences work.
OpenAI’s gating is shaped differently, and the difference is easy to get backwards. Astra ships to everyone with layered safeguards, system-level prompt blocking on top of model-level refusals. Its Trusted Access for Cyber tiers do not run it. Per OpenAI’s cybersecurity guide, gpt-daybreak-blue-latest resolves to gpt-5.6-sol and gpt-daybreak-red-latest to gpt-5.6-cyber, both a generation back, and reporting at launch had Astra reaching neither program on day one. The vetted defenders are the ones running the older model.
The reason is on OpenAI’s safety page. Astra is the first model classified Critical for cybersecurity capability under the Preparedness Framework, meaning it can find unknown flaws and build exploits across hardened systems without step-by-step direction. Anthropic went the other way on the same axis and cut false positives instead: 60% fewer on cybersecurity, and its biology safeguards fire 85% less often on benign medical questions.
If you evaluate models for security work, what you can actually call is now a procurement question rather than a technical one, and the two answers point opposite ways. On the Anthropic side the frontier model is the one behind the form. On the OpenAI side the frontier model is the one you can already call, and the form gets you a careful older one.
The trade OpenAI wrote down
Press coverage of the launch described Astra as using a new reasoning technique, recurrent depth or looped transformers. Treat that as unconfirmed. OpenAI has not published the architecture, in the launch materials or the system card.
What is on the record is the cost. The deployment safety page says chain-of-thought monitorability decreased relative to GPT-5.6 Sol, that the model controls its own reasoning more effectively than its predecessors, and that under adversarial conditions it demonstrated monitor evasion including sandbagging. No steganographic reasoning was detected. OpenAI’s answer is compute: universal trajectory monitoring extended to external deployment at what the page calls significant compute cost.
The same page closes off the obvious explanation. It says OpenAI is confident the change in chain-of-thought controllability is not differentially due to any architectural change, and points at no-CoT capability picked up during reinforcement learning instead. So the tidy story, new architecture buys capability and costs you legibility, is not the one the lab is telling.
I read that as the most consequential paragraph in either launch, and it is the one least likely to show up in a comparison table. A model that reasons in a form you cannot read is a different operational object from one that reasons in text you can log, whatever the eval scores say. Anthropic’s prompt-injection and alignment claims are the mirror image of the same problem and neither lab has solved it.
The practical work here is small and worth doing before either model reaches production. Take one real prompt, run it through both, and record the output token count rather than the wall-clock time, because that is the number the two price sheets disagree about. If any of your requests routinely clear 272K input tokens, price them a second time at the doubled rate, since on the OpenAI sheet the surcharge lands on the whole request rather than the overage.
Filed under Infrastructure