GPT-6.2 vs Claude Fable 5.2: Same Price, Not the Same Cost

GPT-6.2 Astra and Claude Fable 5.2 both cost $8 and $40 per million tokens. At equal scores, a Fable task still costs 3.6 times more. Independent data inside.

AI
9 min read

On November 17, OpenAI released GPT-6.2 as two models: GPT-6.2 Astra, its most capable model, and GPT-6.2 Sol, its everyday model. Eight days earlier, Anthropic had shipped Claude Fable 5.2, its frontier model. The two price sheets show exactly the same numbers: $8 per million input tokens and $40 per million output tokens for Astra, $8 and $40 for Fable.

On paper, it's a draw. In practice, the price per token says nothing about the price of a task, because one model can write five times more than another to reach the same result. That is what Artificial Analysis measures, an independent evaluator that runs every model through the same tests: at equal scores, Fable 5.2's cheapest setting costs $9.08 per task, against $2.52 for GPT-6.2 Astra, 3.6 times as much.

So is GPT-6.2 better than Claude Fable 5.2? No: Fable is still ahead at the top, with 65 points against 62 on the Artificial Analysis Intelligence Index, and also ahead on real research tasks according to Epoch AI. Astra, however, is far more economical. And we think most business use cases do just fine with a cheaper model from the same lab.

What is GPT-6.2, released on November 17?

GPT-6.2 is OpenAI's new generation of models, two months after GPT-6 Astra. OpenAI describes Astra as its model "for your hardest projects", trained to produce slides, spreadsheets and documents that follow the conventions of finance, law or accounting. Sol is pitched as the versatile everyday model. Both accept up to one million tokens of context, several thousand pages.

On price, Astra gets cheaper: its tokens cost 20% less than GPT-6 Astra's, and input tokens already seen (the cache) 60% less. Sol keeps GPT-6.1 Sol's rates: $2 input and $10 output, a quarter of Astra's. In credit-based ChatGPT plans, Astra costs 200 credits per million input tokens and 1,000 per million output tokens, against 250 and 1,250 for GPT-6 Astra; Sol costs 50 and 250.

In ChatGPT, both models are only offered in Work and Codex, not in the regular chat, on Plus, Pro, Business, Enterprise and Edu plans. On Enterprise and Edu, they stay off until an administrator turns them on. The rollout is gradual: on OpenAI's forum, a Business customer in France could not see the models yet the day after the launch. Once enabled, Sol is the default for new Work conversations. How to choose between Astra, Sol and the other ChatGPT models, credit by credit, is covered in our guide to making ChatGPT Enterprise credits last.

GPT-6.2 Astra or Claude Fable 5.2: which one is smarter?

The Artificial Analysis Intelligence Index is a score out of 100 that combines nine evaluations: document-based office work, automation, coding, science, general knowledge. Its value is that it is measured the same way for every model, by a third party that does not sell models. At maximum reasoning effort, Claude Fable 5.2 ranks first out of 92 models with 65 points; GPT-6.2 Astra follows with 62, five points above GPT-6 Astra; GPT-6.2 Sol scores 57.

Other independent evaluators point the same way. According to The Decoder, Epoch AI puts Astra level with Fable 5.2 on its overall index, but behind it on tasks drawn from Epoch's own research, where Astra "played it too safe" and skipped key estimates. On the Vals AI leaderboard, Fable 5.2 is first, Astra seventh and Sol ninth.

Each lab compares its model to an opponent that suits it

The two labs' announcements, for their part, contradict each other. OpenAI claims that on Agents' Last Exam, a test of multistep professional work, Astra beats Fable 5.2 by 5.5 points at 5% of its cost, and that it more than doubles Fable's pass rate on a legal test by Harvey. Anthropic, for its part, publishes a table where Fable 5.2 beats GPT-6 Astra on coding, computer use and reasoning: 74.2% against 57.2% on Humanity's Last Exam, for instance.

Neither table is a fair fight. Anthropic compares Fable to the September GPT-6 Astra, the only one available on November 9. OpenAI picks the tests where Astra shines, and notes in small print that nine Fable tasks were excluded from Agents' Last Exam. On legal work, Vals AI contradicts OpenAI: Astra ranks only 30th on its legal research benchmark. The numbers a lab publishes about its own model remain sales arguments until a third party reproduces them.

Same price per token: why does Fable 5.2 cost more per task?

Cost per task is what a request actually costs once it is completed: the token price multiplied by the number of tokens the model read and wrote, reasoning included. Artificial Analysis computes it over the nine evaluations of its index, for each model and each effort level, that is, the amount of thinking time the model is allowed before it answers.

The results, read on November 18 from the Artificial Analysis comparison:

Model (effort)Intelligence IndexCost per taskTime per task
Claude Fable 5.2 (max)65$26.5441 min
Claude Fable 5.2 (xhigh)62$14.8324 min
Claude Fable 5.2 (high)62$9.0815 min
GPT-6.2 Astra (max)62$4.0013 min
GPT-6.2 Astra (xhigh)62$2.527 min
Claude Fable 5.2 (medium)56$3.806 min
GPT-6.2 Astra (medium)57$1.624 min

With the same token price, the whole gap comes from the number of tokens. To run the full index at "xhigh" effort, Fable 5.2 wrote 144 million tokens and GPT-6.2 Astra 27 million, for the same score of 62. Fable writes about five times more to reach the same result. Artificial Analysis ranks it 89th out of 91 for verbosity. It does generate twice as fast, 111 tokens per second against 53, but it writes so much more that the task takes twice as long.

What a point of intelligence costs at the top

Fable 5.2 does keep a real edge: at maximum effort it climbs to 65, where Astra tops out at 62. Those three points cost $26.54 per task, against $4.00, 6.6 times as much. For a research team attacking a problem nobody has solved, three points can be worth any price. For a company running thousands of contracts or financial tables through a model, the gap in the bill matters far more than the gap in the score.

The nine Artificial Analysis evaluations are long and demanding, often run by an agent that chains steps. On a short request, summarizing an email or rewording a paragraph, Fable's verbosity weighs much less. And writing more is not always a flaw: it is also what lets Fable go further on open problems, where Epoch AI faults Astra for its caution.

What about GPT-6.2 Sol, the everyday model?

For most uses, the Astra versus Fable duel is a false problem: both labs themselves recommend a smaller model by default. Anthropic advises starting with Claude Opus 5.5 for most coding and knowledge tasks, and moving up to Fable only for the most ambitious work. OpenAI starts Work conversations on Sol.

On the Artificial Analysis index, GPT-6.2 Sol reaches 57 points for $1.01 per task at maximum effort. Claude Opus 5.5 gets the same score at "xhigh" effort for $6.05. Sol matches GPT-6 Astra, September's flagship, for about 15% of its cost per task. Fable 5.2 at medium effort (56 points) costs $3.80: nearly four times Sol, for one point less. OpenAI goes further: on FrontierCode 2.0, a test that checks whether code changes would pass review by the maintainers of a real project, Sol lands within half a point of Astra for about one fifth of the cost.

Sol does have a weak spot in coding. According to The Decoder, Artificial Analysis measures it three points below GPT-6.1 Sol on its coding agent index, with a 14-point drop on a test where it more often leaves key facts out of its answers. For a development team whose tooling is tuned for GPT-6.1 Sol, testing on its own repositories before switching seems prudent to us.

What are the known weaknesses of GPT-6.2?

The most serious one comes from OpenAI itself. In its system card, the lab measures how often a model keeps quiet about an obstacle that stopped it from finishing: a deleted file, an unavailable tool. GPT-6.2 Astra did so in 22.28% of trials, against 11.09% for GPT-6 Astra. OpenAI says it is "concerned by this regression" and suspects its efforts to make answers more concise. An instruction added in Codex reduces the problem, but it does not apply to direct API requests. An agent built on Astra can therefore report a task as done when it is not.

On Artificial Analysis's hard factual questions, Astra makes up an answer instead of saying it does not know in 59% of cases, and Sol in 55%. These questions are chosen to be very hard: the figure does not mean Astra is wrong half the time in daily use, but that it prefers an answer to an admission of ignorance. Testers with early access describe unproductive loops and tasks abandoned too early; OpenAI replies that substantial changes were made in the final days before launch.

The launch drew attention: the Hacker News thread gathered 1,291 points and 808 comments in a day. Cost per task gets praise there; the leap from benchmarks to real work, much less.

What does the "High" threshold in self-improvement mean?

OpenAI grades its models by risk level in its Preparedness Framework. GPT-6.2 Astra is OpenAI's first public model to reach the "High" threshold in AI self-improvement, that is, its ability to speed up AI research. According to OpenAI, this threshold amounts to giving each of its researchers a highly capable research engineer as an assistant, compared with 2024.

For a company, the practical consequence lies elsewhere. Astra and Sol reach the "Critical" threshold in cybersecurity, and OpenAI has tightened safeguards on vulnerability-related requests, at the risk of blocking legitimate vulnerability research. Anthropic made the same choice: Fable 5.2 blocks most cybersecurity tasks, which are reserved for its Mythos version on application. A security team counting on either model for penetration testing will have to go through these vetted programs.

Data and sovereignty: where do your requests go?

Here GPT-6.2 has a concrete advantage. In the API, Astra and Sol support data residency in the European Union, for a 10% price uplift. Fable 5.2, by contrast, requires 30-day retention of requests for safety monitoring, barring an exception, and Anthropic offers directly only two processing zones, global or United States. We covered this regime in our article on Claude Fable and zero data retention.

But residency settles the location, not the applicable law. OpenAI and Anthropic are American companies: the CLOUD Act requires them to hand over data under their control to US authorities, "regardless of whether […] located within or outside of the United States". Neither GPT-6.2 nor Fable 5.2 is a sovereign AI, whatever data center you pick. We explain the distinction on our sovereign AI page and in our analysis of Europe's digital dependence on the United States.

For data that must stay under European jurisdiction (M&A, health, HR), the alternative exists, with a capability gap to accept. At Ask This Guy, we use Qwen 3.8 27B, an open-weights model hosted by OVHcloud, the French cloud provider: it scores 20 points on the Artificial Analysis index, far from Astra's 62, but enough for document search and everyday writing.

What an IT department can take away

The price per token is no longer a good selection criterion: two models listed at the same price can cost up to 6.6 times more than each other per task. The only number that counts is the cost of your own tasks, measured on your data. A test on fifty or so real requests, recording the tokens consumed and the response time of each model, tells you more than any leaderboard.

Astra and Fable are built for the longest and hardest work: an agent changing an entire codebase, complex financial modeling, an open scientific problem. For most of the rest, GPT-6.2 Sol or Claude Opus 5.5 do as well for a fraction of the price. In ChatGPT Enterprise, that means having the administrator enable the 6.2 models, then choosing the default model.

That leaves agents. Astra's regression on hidden obstacles confirms a rule we already follow: never take an agent at its word. Automated tests, checks that the expected files exist, and human review of risky steps cost less than a task reported as done that is not. It is also why we route each request to the smallest model that will do, rather than sending everything to the biggest one: the Jevons paradox shows that a lower unit price often drives the total bill up.

Frequently asked questions about GPT-6.2 and Claude Fable 5.2

Is GPT-6.2 available in ChatGPT?

Yes, since November 17, 2026, but only in ChatGPT Work and Codex, not in the regular chat. GPT-6.2 Astra and GPT-6.2 Sol are open to Plus, Pro, Business, Enterprise and Edu plans. On Enterprise and Edu, an administrator has to enable them first. The rollout is gradual: some accounts cannot see them yet.

How much does GPT-6.2 cost?

In the API, GPT-6.2 Astra costs $8 per million input tokens, $0.40 for cached input and $40 for output, 20% less than GPT-6 Astra. GPT-6.2 Sol costs $2, $0.10 and $10. In credit-based ChatGPT plans, Astra costs 200 credits per million input tokens and 1,000 per million output tokens, Sol 50 and 250.

Is GPT-6.2 better than Claude Fable 5.2?

Not at the top. According to Artificial Analysis, Claude Fable 5.2 scores 65 at maximum effort, against 62 for GPT-6.2 Astra, and both Epoch AI and Vals AI put it ahead on real-world tasks. On the other hand, Astra reaches 62 points for $2.52 per task, while Fable spends at least $9.08 for the same score.

What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?

Astra is OpenAI's frontier model, for long and difficult tasks; Sol is the everyday model, four times cheaper per token. According to Artificial Analysis, Astra scores 62 points for $4.00 per task and Sol 57 points for $1.01. On the FrontierCode 2.0 coding test, OpenAI reports Sol within half a point of Astra for about one fifth of the cost.

Can GPT-6.2 process data in Europe?

Yes, in the API: GPT-6.2 Astra and Sol support data residency in the European Union, billed at 10% more. This guarantees where processing happens, not which law applies: OpenAI remains an American company subject to the CLOUD Act, which can compel it to hand over data stored in Europe.

What is the cost per task of an AI model?

It is the real cost of a request carried through to completion: the token price multiplied by the number of tokens the model reads and writes, reasoning included. Two models with the same price per token can have very different costs per task if one writes much more than the other, as Claude Fable 5.2 does compared with GPT-6.2 Astra.


For two years, labs competed on the price of a token. With GPT-6.2, OpenAI is putting something else forward: the number of tokens needed to finish a task, in other words doing the same work while writing less. The question to ask an AI vendor is no longer "how much does a million tokens cost?" but "how many tokens do you need to do my work?". Few invoices today can answer it.

Partager :

Related Articles

Olala mini, OVH's first French LLM: what is it really worth?AI

Olala mini, OVH's first French LLM: what is it really worth?

Olala mini 7A1B, OVH's first LLM, runs without a GPU. What 7A1B means, what the figures leave out, and what this French LLM can do for European businesses.

9 min
DeepSeek V4.1 Pro vs DeepSeek V4 Pro: What to ExpectAI

DeepSeek V4.1 Pro vs DeepSeek V4 Pro: What to Expect

DeepSeek V4.1 Pro is not out as of November 16. What is confirmed, what the leaks are worth, and DeepSeek V4 Pro today: pricing, MIT license, data location.

9 min
AI and mathematics: when checking costs more than producingConcepts

AI and mathematics: when checking costs more than producing

OpenAI released 722 math manuscripts at once and mathematicians can't keep up. Their crisis shows where the work goes: choosing, checking and understanding.

10 min

Interested in our solutions?

Discover how Ask This Guy can help you accelerate with AI

Book a Demo