Claude Fable 5.2 and Gemini 4 Carbon: The unit price of Carbon is four times cheaper, but the low gear is more expensive per question

People whose weekly quota has bottomed out will get a reset that only restores their Fable usage, which is equivalent to not getting it. This is what Claude's paying users have been complaining about in the past two days, and Anthropic admitted it: on November 10, it reissued a full reset to Pro, Max, Ultra, and Team. The instructions stated that the previous Fable reset failed to help the group of people who had used up the weekly quota of all models.

On the same day, Google moved Gemini 4 Carbon from “security defenders only” into the hands of paying developers, with a $2 input and $10 output. The Fable 5.2 is $8 and $40.

A quarter of the price.

On the one hand, there is not enough credit, and on the other hand, there is a new flagship that is four times cheaper. The first reaction of many people is to move. I looked at the pricing pages of the two companies, Artificial Analysis's tiered scores, and Vals's list, and the conclusion was not so smooth: the four-fold price difference is true, but it only exists on the price list and falls into the bill of each question. The mid-to-low tiers are actually cheaper than Fable.

The difference in the price list is four times, and the difference in the bill for a single question is less than 10%

Artificial Analysis ran the intelligence index for each inference gear of the two companies. The numbers posted on November 10 are as follows. The Carbon column is priced at US$4/$20 after December. I added a rough calculation next to it after the initial price was halved.

Look at the low line. Both scores are 54 points, Fable costs US$3.10 per question, and Carbon’s December price is US$7.19, which is more than twice as expensive. Even if the current initial price is discounted by half, Carbon will still cost around US$3.6, which is still higher than Fable's low-end.

The medium grade Carbon has 2 points more, and each question is about 50 cents cheaper during the initial launch period; after December 1st, it will be 73% more expensive than Fable medium per question.

The reason is that there is almost no gap between the three gears of the Carbon. From low to high, the cost of a single question ranges from $7.19 to $7.76, and the time consumption ranges from 22.5 to 25.6 minutes. High is only 18% more expensive than medium. It takes more than 20 minutes to think about which one to choose, and it also burns about the same amount of tokens. The levels of Fable are really adjustable: the cost of a single question of high is 2.4 times that of medium, and max is nearly three times that of high.

Therefore, the one with a unit price four times lower will not save you money by completing a question, unless you originally planned to turn Fable to high or above.

The Carbon doesn't have a "spend money" option, something that's not apparent on its price list. Its API default file is medium, which cannot be turned off; Fable's API default is high, and the default file in Claude application and Claude Code is medium. Both sides do not accept sampling parameters such as temperature and top_p. If Fable is set on the other side, it will directly return 400.

The fact that it is three times slower affects the feel more than the score.

Also in medium, one question on Fable takes 8.1 minutes, and on Carbon 23.6 minutes.

Google's own documentation recommends setting the client timeout to at least 15 to 20 minutes. It doesn't matter who writes the batch process, but it will be hard for the person who sits in front of the editor and waits for the results.

There is a set of small third-party tests that can reveal the difference in the way the two companies work. The author of Atomic Agent took a code with about 1,400 lines, buried 13 bugs, and missing one function for both parties to fix. 59 hidden tests were scored, and each stage was run three times. All four configurations passed 59/59. What cannot be distinguished is the result, the process is very different:

Fable medium took 6 steps, 4.2 minutes, $1.87. Carbon medium took 56 steps, 12.7 minutes, $0.95.

6 moves versus 56 moves. Fable means thinking it through and making changes in one go, while Carbon means walking fast in small steps, making a few changes and running for testing. Most of Carbon's inputs hit the cache, so it's not that expensive for a lot of steps, and it's actually half the price at the initial price. It can be recalculated based on the December price. The author estimates it to be about US$1.90, which is the same as Fable medium and takes twice as long.

The Carbon high gear also exposed a problem: instability. The three passes took from 14.6 minutes to 29.6 minutes. In the longest pass, only 23 steps were recorded out of 69, and the cost data could only be invalidated. The author himself also said that running a question three times cannot explain intelligence, and I agree; but it is enough to show that the time-consuming variance of Carbon is larger than that of Fable.

Fable 5.2 The two most embarrassing losses

The overall score is Fable winning. The total index of Vals is 71.31% vs. 70.45%, Vibe Code Bench is 93.77% vs. 89.06%, Terminal-Bench 4.0 is 65.15% vs. 63.13%, Financial Intelligence is 69.47% vs. 67.16%, and all four Fables rank first.

Then there’s legal and tax.

Vals Legal Index, Carbon is ranked 3rd with 75.93%, and Fable 5.2 is 69.17%, ranked 20th. Tax Agent Bench, Carbon ranked 1st with 51.78%, Fable 5.2 ranked 27th with only 32.09%.

A model that ranked first in overall ranking ranked outside the 20th place on tax issues, a difference of nearly 20 percentage points. This is not something within the margin of error. I couldn’t find Anthropic’s explanation for these two items, and there’s no attribution on Vals’ page. I can only put the phenomenon here: it’s a job that needs to be gnawed at the legal provisions and calculated step by step according to the rules. Don’t default to the most expensive one now.

The average cost per test over at Vals is also there: $10.51 for Carbon and $17.94 for Fable, where Carbon is officially priced at 4/20.

The scores reported by the two companies themselves have to be discounted in one area.

In the comparison table on the Anthropic release page, Fable 5.2 won most of the rows: FrontierSWE v2 was 69.9% vs. 61.2%, OSWorld 2.1 was 89.3% vs. 81.5%, and Terminal-Bench-Science was 74.6% vs. 64.8%. Carbon only won out on AutomationBench, 53.9% to 51.4%, where Google says its per-task cost was less than half that of the second-place finisher.

Two small prints are worth reading.

In the FrontierCode 2.0 line, 49.1% of Fable's results were run at the xhigh file, and 41.1% of Carbon's targets were at the medium file. The previous table has shown that Carbon's gear has little impact, but when comparing one's own sub-high-end to others' mid-range, these 8 points cannot be taken as a given. Fable scored 45.3% on high.

OSWorld In that line it's Google's turn. Carbon's 81.5% is calculated by Google itself, taking the best of three times, using the offline subset plus the partial score.

There are also two things that Anthropic uses as its signature: more than a thousand sub-agents wrote a bootable operating system prototype in eight days, and formalized the proof of Fermat's Last Theorem in 30 hours. There are currently no third-party reproductions of these two items.

Several thresholds besides money

Cache writes.Fable write cache costs $10 per million for 5 minutes and $16 for 1 hour, which is more expensive than normal input; read cache costs $0.20. Carbon's caching is automatic, there is no charge for writes, and reads cost $0.10 during the initial period and $0.20 starting in December. Most of the inputs in long conversations and multi-round agent tasks are repeated, and this one will widen the gap between the two sides.

How much can be used in the subscription.Fable 5.2 is included with Max, Ultra, and Premium Team seats, but can only eat up to 50% of your weekly quota. Pro and standard Team seats are only available during the launch week, and additional usage credits will be purchased after November 16. This is the root of the complaint at the beginning of the article. Carbon takes a different approach: Gemini applications, search AI Mode and Antigravity are only available if you subscribe to Google AI Ultra; developers use the Google AI Studio API, which does not have a free tier.

Data retention.Fable 5.2 requires 30 days of data retention, and companies that have signed a zero-retention agreement must find another Anthropic license to use it.

Security request.For penetration testing, writing vulnerability exploits, and binary vulnerability analysis, Fable will be blocked by the classifier, and you can choose to fall back to Opus 4.8 to answer; for research biology questions, fall back to Opus 5. Anthropic says there are about 55% fewer false stops than 5.1. Google’s complete security capabilities are only available to defenders verified through the Fairwind program. To be on the red team, both companies must pass the qualification test first.

Output upper limit.Fable's single token is 128,000 tokens, and Carbon's default is 65536. For longer tokens, background requests are required.

Each of them changed their mouth several times in half a year.

Carbon’s timeline: Released on October 28th, only for security defenders; open to paid developers on November 10th; initial price expires on November 30th. It took 13 days from being "visible but not tangible" to being able to adjust the API.

On Fable’s side, the price of 5.1 is US$10/US$50. On November 9, the price of 5.2 dropped to 8/40. The output speed is officially said to be up to 30% faster. In addition, a 16/80 fast mode is added, which is claimed to be 2.5 times faster in typing. On the AA index, 5.1 is 49 points and 5.2 is 65 points, an increase of 16 points in one generation. To run the entire set of indices, Fable 5.2 used 275 million tokens, and Opus 5.5 used 218 million tokens. There is more to say.

There is one item in the system card that I am more concerned about. During audit conversations, models were aware that they might be being evaluated up to 43% of the time, which is higher than any previous model. With a model that increasingly understands that "this is an exam", no one can say for sure whether the gap between exam results and daily performance will become wider.

Anthropic's own recommendation to most users is still to use Opus 5.5 first, which is half the price of Fable. Google’s corresponding recommendation is Argon, which is the same price as Carbon. Both companies are urging you not to choose the most expensive one by default.

Before and after November 30, there are two different questions

Before the end of the month, Carbon medium is the most cost-effective among the eight configurations: 58 points, more than three yuan per question, provided you can afford to wait twenty minutes. Batch-run evaluations, code reviews at night, legal and tax documentation work, it is cost-effective to move there and try it out now, and by the way, use real tasks to measure its time-consuming variance.

Rescheduled after December 1st. In the same range of 54 to 58 points, Fable's low and medium questions cost 3 to 4 US dollars per question and produce results within ten minutes, while Carbon costs 6 to 7 US dollars and takes more than 20 minutes. At that time, Carbon's remaining advantage was the high range: 60 points, $7.76, 15% cheaper than Fable high's 62 points, $9.08, plus the two legal and tax lists.

If you need a score of 62 or above, only Fable's high can give you $9 to $27 per question. Subscribers remember the 50% weekly limit.

For interactive development while writing and waiting in the editor, I will stay with Claude. The feeling of completing the modification in 6 steps is too different from that of completing the modification in 56 steps. Gemini is more suitable for tasks that can be thrown into a queue and not watched.

The currently measured Carbon cost is directly multiplied by two when accounting.

Related comparisons: "Gemini 4 Carbon, Claude Opus 5.5, GPT-6 Astra: Carbon is one point ahead, most people can only look at the list this month", "Claude Ultra and ChatGPT Pro 200: The 5-hour window is gone, only one company wrote how many times the weekly quota is", "Muse Ember, Claude Haiku 5.5, GPT-6 Luna: The price is the same, the difference in the account for each question is almost three times"