W.165OpenAI priced GPT-6.2 Astra at $8 and $40, the same as Anthropic’s Fable 5.2
OpenAI released GPT-6.2 Astra on November 17 at $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra and identical to the list price of Anthropic’s Claude Fable 5.2. GPT-6.2 Sol stays at $2 and $10. OpenAI’s claim that Astra does a benchmark at about 5% of Fable’s cost therefore rests on tokens used per task, not on price. On eight benchmarks from OpenAI’s charts, as relayed by outlets, Fable cost 2.1 to 20.8 times as much per task; it scored higher on three of them.
What OpenAI released
Two models, one of them cheaper than its predecessor. OpenAI’s pricing page lists GPT-6.2 Astra at $8 per million input tokens, $0.40 for cached input and $40 for output. GPT-6 Astra is listed at $10, $1.00 and $50. Input and output fell 20%; cached input fell 60%. GPT-6.2 Sol is $2, $0.10 and $10, the same as GPT-6.1 Sol.
Astra’s list price now equals Fable 5.2’s. Anthropic’s pricing documentation lists Claude Fable 5.2 at $8 and $40. Fable 5.1 is listed at $10 and $50, so Anthropic’s step down was also 20%. Fable 5.2 was released on November 9, according to AlphaCorp, eight days before OpenAI’s launch. Sol’s $2 and $10 equal Claude Sonnet 5.5’s.
The two differ below the headline price. Astra’s cached input, at $0.40, is 2 times Fable’s cache read of $0.20. Above 272,000 input tokens OpenAI doubles Astra’s input rate and charges 1.5 times for output on the whole request; Anthropic’s page lists no long-context surcharge for Fable 5.2.
We did not read the announcement itself. OpenAI’s launch page did not open for us. Prices are from its pricing and model pages. Benchmark figures below are from OpenAI’s charts as two outlets relayed them, and from OpenAI’s forum post.
GPT-6.2 Astra and Fable 5.2 list at the same $8 and $40
Dollars per million output tokens at each company’s standard list price, with the input and cached-input prices under each label. Pink: GPT-6.2 Astra. Strong grey: OpenAI’s GPT-6.2 Sol. Grey: Anthropic’s current models. Faint grey: the two predecessors, GPT-6 Astra and Claude Fable 5.1. From OpenAI’s and Anthropic’s pricing pages as read on November 18, 2026.
List prices, $ per million tokens
| Model | Company | Input | Cached input | Output |
|---|---|---|---|---|
| GPT-6.2 Astra | OpenAI | 8.00 | 0.40 | 40.00 |
| GPT-6 Astra | OpenAI | 10.00 | 1.00 | 50.00 |
| GPT-6.2 Sol | OpenAI | 2.00 | 0.10 | 10.00 |
| GPT-6.1 Sol | OpenAI | 2.00 | 0.10 | 10.00 |
| Claude Fable 5.2 | Anthropic | 8.00 | 0.20 | 40.00 |
| Claude Fable 5.1 | Anthropic | 10.00 | 0.25 | 50.00 |
| Claude Opus 5.5 | Anthropic | 4.00 | 0.20 | 20.00 |
| Claude Sonnet 5.5 | Anthropic | 2.00 | 0.10 | 10.00 |
The cost claim
OpenAI says Astra beats Fable 5.2 at “just 5% of the cost”. That is its wording as two outlets quote it. On Agents’ Last Exam, the accounts give Astra 78.3% at $2.89 a task and Fable 5.2 72.8% at $60.05. That is 5.5 points more at 4.8% of the cost. Astra is 5.2 points above GPT-6 Astra, per OpenAI, which implies about 73.1% for the older model.
With equal list prices, the gap is tokens. A cost per task 20.8 times higher at the same price per token means roughly that many more billed tokens, or a different mix of input, cached and output tokens. The one published token count we found is from GDPval: 222,344 output tokens for Fable against 35,812 for Astra, 6.2 times. At $40 per million that is $8.89 against $1.43 of output.
OpenAI ran the comparison. AlphaCorp says the cost-per-task figures come from OpenAI’s own runs, including its runs of Fable 5.2, though it attributes Fable’s Terminal-Bench-Science cost to a run by Artificial Analysis. On Agents’ Last Exam, it says, Fable was plotted at its maximum effort setting and Astra at its “xhigh” setting, and Fable’s result excludes nine tasks that used a fallback model. Data Studios says OpenAI had published no methodology for the cost figures when it wrote. Anthropic has not, in what we read, responded.
The ratio varies widely across benchmarks. On the eight charts for which AlphaCorp gives both models’ scores and costs, Fable’s cost per task is 20.8 times Astra’s on Agents’ Last Exam, 2.4 times on FrontierCode 2.0 and 2.1 times on the Vals Science Index, where Fable was plotted at a lower effort setting. On the Artificial Analysis index, which that firm publishes, it is 6.6 times.
On OpenAI’s charts, Fable 5.2 cost 2.1 to 20.8 times as much per task
Claude Fable 5.2’s cost per task divided by GPT-6.2 Astra’s, our division, on eight benchmarks from OpenAI’s launch charts as relayed by AI Catchup and AlphaCorp. Pink: Agents’ Last Exam, the benchmark behind OpenAI’s “just 5% of the cost” claim. Grey: the others. Dashed line: 1, where the two would sit if they used the same tokens, since their list prices are equal. Costs are from OpenAI’s runs except where a third party ran the benchmark; effort settings differ by chart.
Scores and cost per task
| Benchmark | Astra | Fable 5.2 | Difference | Astra $ a task | Fable $ a task | Ratio |
|---|---|---|---|---|---|---|
| Agents’ Last Exam v1.1 | 78.3 | 72.8 | +5.5 | 2.89 | 60.05 | 20.8× |
| BenchCAD 2.1 Core | 82.5 | 83.5 | −1.0 | 5.43 | 50.77 | 9.3× |
| Artificial Analysis Intelligence Index | 62.1 | 64.6 | −2.5 | 4.00 | 26.54 | 6.6× |
| Terminal-Bench-Science 0.1 | 77.1 | 73.3 | +3.8 | 7.15 | 45.21 | 6.3× |
| Harvey LAB-AA v1.1 | 15.0 | 7.2 | +7.8 | 6.35 | 39.84 | 6.3× |
| FrontierCode 2.0 | 44.2 | 49.1 | −4.9 | 5.97 | 14.19 | 2.4× |
| AutomationBench 1.0.6 | 52.4 | 51.4 | +1.0 | 0.83 | 4.64 | 5.6× |
| Vals Science Index v1 | 68.4 | 63.7 | +4.7 | 4.17 | 8.79 | 2.1× |
The scores
Astra scores higher on five of the eight, Fable 5.2 on three. On FrontierCode 2.0 Fable leads Astra by 4.9 points, 49.1% to 44.2%. On the Artificial Analysis index it leads by 2.5, and on BenchCAD by 1.0. Astra leads on Agents’ Last Exam, Terminal-Bench-Science, the Vals Science Index, AutomationBench and Harvey’s legal benchmark, where its 15.0% is about twice Fable’s 7.2%. AlphaCorp counts six to three in Astra’s favour across OpenAI’s nine head-to-head charts; the ninth, GDPval, has no cost figure.
The larger and more consistent difference is cost. Astra is cheaper per task on all eight. It is ahead on score on five and behind on three. These are charts OpenAI chose; a different selection could read differently.
Sol is the cheaper route to nearly the same coding score. On FrontierCode 2.0, the accounts give Sol 43.8% at $1.29 a task, 0.4 points below Astra at 22% of its cost. Sol’s list price is 25% of Astra’s. On Agents’ Last Exam Sol scored 73.0% at $1.32.
Early reception from heavy users was largely negative, VentureBeat reported. Its November 17 account says several users who tested the models before launch criticised their behaviour on delegated and long-running work, one writing that they “were essentially unusable”, while others praised Astra for animation and video work. The complaints it describes are about quality and steering, not about price.
Astra leads Fable 5.2 on five of eight benchmarks and trails on three
GPT-6.2 Astra’s score minus Claude Fable 5.2’s, in points, on the same eight benchmarks, with both scores under each label. Pink: Astra ahead. Black: Fable ahead. Seven are percentages; the Artificial Analysis index is an index score. From OpenAI’s launch charts as relayed by AI Catchup and AlphaCorp.
What the list price leaves out
The same model lists from $20 to $240 per million output tokens. OpenAI’s pricing and model pages set batch at half of standard and Fast mode at twice. Ultrafast is three times standard through February 28, 2027 and six times from March 1: $120 per million output tokens now and $240 then. Requests over 272,000 input tokens pay $16 and $60 at the standard tier; the model page says the fast tiers’ multiples apply at the same context length, so long requests in those tiers would cost more than the figures here.
Distribution is wider than the API. The forum post says both models are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu accounts. Microsoft’s chief executive said at Ignite on November 17 that it would bring GPT 6.2 to Foundry; the published remarks give no date. The model page lists a context window of 1,050,000 tokens.
OpenAI rated Astra at its High threshold for AI self-improvement. The system card says both models meet its Critical threshold for cybersecurity capabilities, and that Astra meets the High threshold for self-improvement while Sol falls below it.
Astra’s listed output price runs from $20 to $240 per million tokens by service tier
GPT-6.2 Astra, dollars per million output tokens by service tier, with the input price under each label; our multiplication of the standard price by the multiples on OpenAI’s pricing and model pages. Pink: standard. Faint grey: batch. Grey: long context and Fast mode. Strong grey: Ultrafast now. Black: Ultrafast from March 1, 2027. Batch, Fast mode and Ultrafast are shown for requests up to 272,000 input tokens.
What this changes on inthesingularity.ai
- What does Fable 5.2 do to the price of the frontier? Lead: OpenAI matched Fable 5.2’s $8 and $40 eight days after its release, with the same 20% step down from $10 and $50.
- Who sets the price of a token? Lead: the two labs’ top models and their mid-tier models (Sol and Sonnet 5.5, at $2 and $10) now carry identical list prices; the competition OpenAI is advertising is cost per task.
- Lead: a series of cost per task on named benchmarks by model, with who ran each, alongside list prices. Tokens per task is the variable the list price hides.
- Who makes money serving open models? The Ultrafast tier’s move from three to six times standard on March 1, 2027 is a dated price rise worth tracking.
Sources and method
| Item | Figure | Source |
|---|---|---|
| Prices | GPT-6.2 Astra $8 input, $0.40 cached input, $40 output per million tokens; GPT-6 Astra $10, $1.00, $50; GPT-6.2 Sol and GPT-6.1 Sol $2, $0.10, $10; Ultrafast at three times standard through February 28, 2027 and six times from March 1; batch and Fast mode tiers | OpenAI API pricing page, read Nov 18 |
| Prices | Above 272,000 input tokens, input and cache rates double and output is 1.5 times, on the whole request; batch at 50%; context window of 1,050,000 tokens and maximum output of 128,000 | OpenAI model page, GPT-6.2 Astra |
| Prices | Claude Fable 5.2 $8 input, $0.20 cache read, $40 output; Fable 5.1 $10, $0.25, $50; Opus 5.5 $4, $0.20, $20; Sonnet 5.5 $2, $0.10, $10; no long-context surcharge on Fable 5.2 | Anthropic pricing documentation, read Nov 18 |
| The launch | Announcement (not read; the page did not open) | OpenAI, Nov 17 |
| The launch | Both models available in the API, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu; Astra at less than a third of GPT-6 Astra’s cost per task in general and less than half on Agents’ Last Exam; Ultrafast availability | OpenAI developer forum, Nov 17 |
| The launch | Both models meet OpenAI’s Critical threshold for cybersecurity; Astra meets its High threshold for AI self-improvement and Sol falls below it | OpenAI system card, Nov 17 (read in part) |
| Benchmarks | Agents’ Last Exam v1.1: Astra 78.3% at $2.89 a task, Sol 73% at $1.32, Fable 5.2 72.8% at $60.05; Astra 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Fable 5.2 at “just 5% of the cost”; FrontierCode 2.0, Artificial Analysis index and other charts | AI Catchup, Nov 17 |
| Benchmarks | Cost-per-task figures come from OpenAI’s own runs, including of Fable 5.2; GDPval v2: Astra 50.4% with 35,812 output tokens, Fable 5.2 44.8% with 222,344; Terminal-Bench-Science, Harvey LAB-AA, BenchCAD, AutomationBench and Vals Science Index figures; Fable’s Agents’ Last Exam result excludes nine tasks; nine head-to-head charts, six to three for Astra; Fable 5.2 released November 9 | AlphaCorp, Nov 2026 |
| Benchmarks | No methodology for the cost-per-task claims had been published when it wrote; long-context prices | Data Studios, Nov 17 |
| Reception | Users who tested the models before launch criticised their behaviour; one wrote that they “were essentially unusable”; others praised Astra for animation and video work; the price cut framed as $48 against $60 per million tokens, input plus output | VentureBeat, Nov 17 |
| Distribution | Satya Nadella said Microsoft would bring GPT 6.2 to Foundry | Microsoft, Ignite keynote remarks, Nov 17 |
What we read. OpenAI’s API pricing page and its model page for GPT-6.2 Astra; its developer forum post; the first part of its system card; Anthropic’s pricing documentation; Microsoft’s Ignite post; VentureBeat; two outlets’ accounts of OpenAI’s charts; and a third’s account of pricing and methodology. OpenAI’s announcement page did not open, so every benchmark figure here is secondhand. Arithmetic. Price cuts: 1 − 8 ÷ 10 = 20%; 1 − 40 ÷ 50 = 20%; 1 − 0.40 ÷ 1.00 = 60%; Fable, 1 − 40 ÷ 50 = 20%. Cache: 0.40 ÷ 0.20 = 2.0. Cost ratios: 60.05 ÷ 2.89 = 20.8; 50.77 ÷ 5.43 = 9.3; 26.54 ÷ 4.00 = 6.6; 45.21 ÷ 7.15 = 6.3; 39.84 ÷ 6.35 = 6.3; 14.19 ÷ 5.97 = 2.4; 4.64 ÷ 0.83 = 5.6; 8.79 ÷ 4.17 = 2.1. Agents’ Last Exam: 2.89 ÷ 60.05 = 4.8%; 78.3 − 72.8 = 5.5; 78.3 − 5.2 = 73.1. Score differences: 78.3 − 72.8 = +5.5; 82.5 − 83.5 = −1.0; 62.1 − 64.6 = −2.5; 77.1 − 73.3 = +3.8; 15.0 − 7.2 = +7.8; 44.2 − 49.1 = −4.9; 52.4 − 51.4 = +1.0; 68.4 − 63.7 = +4.7. Tokens: 222,344 ÷ 35,812 = 6.21; 222,344 ÷ 1,000,000 × 40 = 8.89; 35,812 ÷ 1,000,000 × 40 = 1.43. Sol: 1.29 ÷ 5.97 = 22%; 44.2 − 43.8 = 0.4; 10 ÷ 40 = 25%. Tiers: standard $8 and $40 times 0.5, 2, 3 and 6; long context, input times 2 and output times 1.5. Conflicts between sources. The accounts we read quote OpenAI as “just 5% of the cost”; the figures give 4.8%. OpenAI’s forum post says Astra costs less than a third as much per task as GPT-6 Astra across evaluations in general and less than half on Agents’ Last Exam; AI Catchup attaches “less than a third” to Agents’ Last Exam. VentureBeat’s $48 against $60 adds input and output prices and is not a blended price. Limits. Cost per task depends on effort settings, harness and token mix, which differ between charts and are not fully published; OpenAI ran most of the comparisons, including of its competitor’s model. The eight benchmarks are those for which AlphaCorp gives both models’ scores and costs; GDPval, OpenAI’s ninth head-to-head chart, has scores and token counts but no cost, and AlphaCorp describes it as an internal OpenAI benchmark. Tier prices for batch and the fast tiers are for requests up to 272,000 input tokens. List prices were read on November 18 and carry no effective dates on Anthropic’s page. We did not check Google’s prices and leave its models out. The reception paragraph rests on one outlet.
Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.