$26.54 to $4.00.
This is the "how much does each question cost" calculated by Artificial Analysis after running its set of intelligence indices for the two models: Claude Fable 5.2 is the former number, and GPT-6.2 Astra is the latter. The score was 65 to 62, with Fable 3 points higher.
The strange thing is in the price list. The price tags for both APIs are exactly the same, $8 per million tokens for input and $40 for output. For the same unit price, the difference in the bill for a question is 6.6 times. This difference is not a result of discounts, but a result of the number of words written by the two models.
After OpenAI released 6.2 on November 17, I read through the third-party data I could find and the release notes and system card summaries on both sides. The conclusion is here: It is true that Astra is cheaper, but part of the money it saves is in exchange for reliability.
Within eight days, the two companies adjusted their prices to the same number.
This is rarely mentioned. Fable 5.1 is priced at $10 and $50, dropping to $8 and $40 when 5.2 goes live on November 9th. Eight days later, GPT-6.2 Astra was released. Officially, the standard input and output price was 20% lower than that of GPT-6 Astra, with the same price points of 8 and 40.
Whether it's a coincidence or a benchmark, the starting line is the same.
The fast gears of the two companies have taken two paths.
Fable's fast modes are $16 and $80, twice the standard price, and claim to type 2.5 times faster. Astra's Ultrafast tops out at 4.5x standard speeds, and third-party gateway OrcaRouter is listed at 3x the standard price at $24 and $120, and that's only until February 28, 2027, after which it becomes 6x. The team that currently budgets for Ultrafast for speed will double this item in three months.
The version line is also worth checking out. GPT-6 Astra was released in early September, and 6.1 Sol took over at the end of September. This time Astra skipped 6.1 and went directly to 6.2. Fable went from 5.1 to 5.2 and from 49 to 65 on the AA index. The first generation was up 16 points.
Astra is cheap because Fable is so good at writing
According to OpenAI, the output of 6.2 is shorter when completing difficult tasks, and the cost of a single task is less than one-third of GPT-6 Astra. AA’s own measurements are different: after running the entire set of indices, GPT-6.2 Astra outputs 65.8 million tokens, GPT-6 Astra outputs 54.3 million tokens, and the new version has about 21% more tokens.
It has not changed much from its previous generation. The cost of a single question under the AA caliber has dropped from US$6.24 to US$4.00. The 36% reduction is basically the 20% price reduction. Coupled with some changes in the reasoning process, it has little to do with "less words".
和 Fable 比就是另一回事了。 According to the usage announced by AA, Fable 5.2 uses about 275 million tokens to run the same index, which is more than four times that of Astra. The unit price is the same, but the usage is four times, and the bill will naturally increase.
Therefore, the statement "Astra is six times cheaper" requires a premise: the comparison is that both sides are driven to the highest reasoning gear. Fable can be adjusted downwards. The low level is US$3.10 and 54 points per question on AA, and the medium level is US$3.80 and 56 points. For the same price of about four yuan, the Fable medium scored 56 points and the Astra scored 62 points, a difference of 6 points. This is the real distance between the two sides under the same budget. It is much smaller than "six times", but Astra is still ahead.
| GPT-6.2 Astra | Claude Fable 5.2 | |
|---|---|---|
| List price (input/output, per million tokens) | 8 / $40 | 8 / $40 |
| cache read | $0.40 | $0.20 |
| cache writes | Not listed separately | $10 for 5 minutes, $16 for 1 hour |
| AA Smart Index (highest grade) | 62, $4.00 per question | 65, $26.54 per question |
| AA-Omniscience hallucination rate | 59.1% | No numbers of the same caliber were found |
| Single maximum output | 128,000 tokens | 128,000 tokens |
Both cars overturned
Let’s talk about Astra first. AA's Omniscience test specifically looks at whether the model will hardcode when it doesn't know the answer. The hallucination rate of GPT-6.2 Astra is 59.1%, and the previous generation GPT-6 Astra is 49.8%, which is an increase of nearly 10 points. As scores have gone up, so has the tendency to make things up. OpenAI also has one in its own system card: In the two alignment tests, the two 6.2 models more often remain silent when encountering obstacles. For example, the files required for the task do not exist or the search tool cannot be adjusted. It does not report and continues.
Epoch AI was tested before its release, and its ability index was tied with the two, both at 174. But in Epoch's own real-world research tasks, Astra fell behind Fable, with one question missing a core issue and a data center-related task skipping key estimates.
These three things point to the same problem: Giving a seemingly complete answer takes precedence over admitting that you can't do it.
Fable's car rolled over in another location. The last time we compared Fable and Gemini 4 Carbon, we compiled Vals data. Fable 5.2 ranked 27th with only 32.09% in the Tax Agent Bench; and 69.17% in the Legal Index, ranking 20th. The model that ranks first in the comprehensive list has failed very badly in tasks that require step-by-step calculations. I haven’t seen Anthropic give an explanation yet.
The Astra isn't legally strong either. It ranks 30th in Vals’ Legal Research sub-list. However, in the Harvey legal agent test published by OpenAI, Astra accounted for 14.6% of the tasks that fully met the standards, and GPT-6 Astra only 4.6%. According to the LAB version of AA, its pass rate is more than twice that of Fable, and its cost is about one-sixth. The two sets of legal data are in opposite directions and have different question types. I am not going to draw conclusions for anyone.
One bug fixing problem, two ways of working
The author of Atomic Agent previously tested Fable with a code of about 1,400 lines and buried 13 bugs, and scored 59 hidden tests. This time he tested the Astra with the same question.
The Astra was 59/59 three times, averaging 86 seconds and $0.82.
Fable medium was also correct at the time. It took 4.2 minutes, $1.87, and was completed in 6 steps. Looking at this question alone, the Astra is nearly three times faster and more than half cheaper. The author himself said that this is only a directional reference, and one question cannot prove anything. I agree, but it is consistent with the conclusion of AA that "Astra uses much less tokens." Conversely, on AA's own FrontierCode 2.0, Fable is 48.8% and Astra is 45.4%. This test is to see whether the changes can be merged by the maintainers of the real code base, which is closer to daily life than fixing a few buried bugs. Astra is faster, and the things modified by Fable are more acceptable to people. The two sets of data are not inconsistent.
272K that line
This is, I think, the easiest pitfall on the Astra price list.
According to the billing rules compiled by OrcaRouter, if the input of a single request exceeds 272,000 tokens, the unit price of the entire request becomes US$16 for input and US$60 for output. Instead of increasing the price for the excess, the entire order is doubled. It gives examples: an input of 271,999 tokens costs about $2.17, and an input of 300,000 tokens costs $4.80. For less than 30,000 tokens, the input fee more than doubled.
Astra's total context has 1.05 million tokens, and the maximum input is 922,000. It looks like it can stuff a lot, but if you really stuff it in, it’s a different price. To throw the entire code base into the review, you must first add a length check on the client.
The pitfall on Fable's side is caching. The read cache costs $0.20, half that of Astra; but the write cache has to be paid separately, $10 per million for 5 minutes, which is more expensive than ordinary input. In scenarios where the conversation is short and the cache hit rate is low, Fable's cache will actually increase the bill.
What are the people scolding?
That post on Hacker News has more than 750 comments. Some people say that when Astra was opened in the High setting in Codex, it added unnecessary interface controls without authorization, and even redid the entire interface frame. Some teams also said that they switched back to 6.1 when writing marketing copy and documents, thinking that 6.2 is "optimized for running scores." Evaluations of the speed are divided: some say it's ridiculously fast, others say it's terribly slow.
Fable has been criticized for being slow and taking too much quota. Some people say that the credit of a 5-hour window is burned out in ten minutes. The rules in the subscription have also been tightened. Fable 5.2 can only use up to half of the weekly quota in Max, Ultra and Premium Team seats. Pro and standard Team seats will have to purchase additional usage starting from November 16.
There are two other thresholds that have nothing to do with money. Fable requires 30 days of data retention, and companies that have signed a zero-retention agreement must seek authorization from Anthropic separately; requests such as penetration testing and vulnerability exploitation will be blocked by the classifier and fall back to Opus 4.8 for answers. Astra's network security denial-of-answer is also stricter than GPT-6 Astra, making it difficult for both sides to be on the red team.
There is no headline in the system card
The system card of Fable 5.2 writes that the rate of models realizing that they may be being evaluated in audit conversations is as high as 43%. Someone on HN cited the 6.2 system card, saying that Astra’s evaluation awareness is stronger than that of GPT-6 Astra, and it is better at avoiding thought chain monitoring in some tests.
The two latest flagships are increasingly aware that "this is a test."
This has a very specific impact on ordinary users: the gap between the scores on the list and your own performance in tasks may be getting wider. There is a 3-point difference in running scores between the two models. Which one is better when put into your workflow? You can only test it yourself with real tasks.
Divide the work
Give Astra the job of batch, verifiable, and errors that can be caught by testing. It is fast and cheap for bug fixing, order testing, formatting based on templates, 3D scenes and demonstration prototypes. AA's 3D Elo is 1318 and Fable is 1271.
There is no way to automatically check the answer, so I won’t change it just because it’s cheaper. Research reports, technical surveys, scenarios that require it to say "I don't know", a 59% hallucination rate and not saying anything when faced with obstacles are more important than saving $20 per question. Leave this kind of work to Fable, but don’t open max, medium and high are enough, and the bill can be reduced to a fraction of the original price.
Don't hand over the tax and work calculated according to the provisions by default.
Anthropic's own advice to most users is still to start with Opus 5.5, which is half the price of Fable. OpenAI has pushed 6.2 Sol to the first place in Codex by default, with 57 points on AA and $1.01 per question. Both companies are urging you not to default to the most expensive one.
Related comparisons: "Claude Fable 5.2 and Gemini 4 Carbon: Carbon is four times cheaper per unit, but each question in the low range is more expensive", "Gemini 4 Hydrogen and Claude Sonnet 5.5: 57% and 70.6% are not the same paper at all", "Muse Ember, Claude Haiku 5.5, GPT-6 Luna: The price is the same, but the difference in the account of each question is almost three times"