🔍 Read the full analysis: The Real-World Work Differences Between Claude Fable 5.2 And Gemini 4 on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Tests available on November 17, 2026, do not identify one best model for every job. Claude Fable 5.2 leads the Artificial Analysis overall index and several coding and long-document tests, while Gemini 4 Carbon leads on terminal work, professional documents and business workflows, with lower reported task costs. Reliability also differs: Fable answers more questions correctly, but Carbon is less likely to guess incorrectly when it does not know.
Claude Fable 5.2 and Google’s Gemini 4 models are producing different winners across workplace tests, according to evaluations and pricing data accessed November 17. Fable 5.2 leads the Artificial Analysis Intelligence Index, while Gemini 4 Carbon scores higher on several workflow tests and has a lower reported cost per index task, leaving the choice dependent on the work and tolerance for errors.
Anthropic released Claude Fable 5.2 on November 9. Google opened Gemini 4 Carbon to paying users on November 10 and made the faster Gemini 4 Hydrogen available on November 12. Gemini 4 Argon is the family’s more efficient model. The comparison by ThorstenMeyerAI.com draws on Artificial Analysis model pages accessed November 17 and vendor-published information; the article says neither company nor Artificial Analysis reviewed or sponsored it.
On the Artificial Analysis Intelligence Index, Fable 5.2 scores 65 at Max effort and 62 at High. Carbon scores 60 at High, Hydrogen 54 at its highest setting, and Argon 50 at High. But the index’s reported cost per task is $26.54 for Fable at Max and $9.08 at High, compared with $5.48 for Carbon, $4.97 for Argon and $3.16 for Hydrogen. Google’s Carbon and Argon prices used for those figures are introductory rates that double on December 1; task costs and token prices are not interchangeable.
Individual test results vary. On FrontierCode-AA 2.0, Fable scores 49.1% to Carbon’s 40.6% for coding in open-source repositories. Carbon leads on Terminal-Bench 4.0, 65.2% to 59.6%; GDP.pdf professional-document reasoning, 28.4% to 21.8%; and AutomationBench-AA business workflows, 77.8% to 71.4%. Fable leads on the AA-LCR long-document test, 83.3% to 79.3%, and on the AA-Briefcase office-style knowledge-work Elo rating, 1,677 to 1,611. A higher benchmark score does not establish that a model will perform equally well in every company’s systems.
Choosing Models by Workload
The results make model selection a task-and-cost decision, not a simple ranking. Teams doing repository coding or long-document analysis may find Fable’s test leads relevant. Teams automating business software or handling terminal tasks may give more weight to Carbon’s results. These are benchmark signals, not guarantees about performance on a particular team’s data, tools or review process.
Reported cost differences could affect how often organizations can run a model, especially for repeated or agentic workflows. But the listed task-cost figures reflect a specific independent evaluation, while API bills depend on usage, token volume and model settings. Google’s announced December 1 price increase for Carbon and Argon also means the current comparison is time-sensitive.
Reliability presents a separate trade-off. In the reported evaluation, Fable answered more questions accurately, but Carbon was less prone to confidently supplying a wrong answer when it lacked one. For work where errors are costly, teams may need to assess abstention, verification and human review alongside raw accuracy and price.
Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones (Black)
As an affiliate, we earn on qualifying purchases.
Three Gemini Models, One Fable
Fable 5.2 is Anthropic’s model for demanding, open-ended technical work. It offers five effort settings, a stated one-million-token context window and up to 128,000 output tokens. The source says it accepts text and images. Google presents Gemini 4 as a family: Carbon as its most capable reasoning model, Argon as the efficient option it recommends for most work, and Hydrogen as a model for rapid execution across agentic workflows. All three are described as having a one-million-token context window.
The source reports input and output prices per million tokens of $8 and $40 for Fable 5.2, or $16 and $80 in fast mode. Carbon and Argon are listed at $2 and $10 through November 30, then $4 and $20 from December 1; Hydrogen is listed at $1.50 and $7.50. Carbon and Argon are not available through the Gemini API free tier, while Hydrogen is. These vendor price listings are distinct from the independent cost-per-task estimates.
The comparison also notes different evidence sources. Artificial Analysis runs models through its own tests; company launch tables are vendor results and may not be independently repeatable. In a separate Epoch AI Capabilities Index, Fable scores 173.44 and Carbon 169.76, but their reported uncertainty ranges overlap. That result supports caution about treating a small aggregate lead as decisive.
“the most ambitious, open-ended technical work”
— Anthropic, describing Claude Fable 5.2
Limits of the Comparison
The available figures do not establish which model performs best in a specific workplace. Benchmark scores depend on test design and settings, and results may not transfer to a company’s software, documents or approval processes. The source says the Epoch AI uncertainty ranges for Fable and Carbon overlap, so that index does not resolve their relative standing with confidence.
The reported reliability figures also need careful interpretation. Fable’s accuracy is 77.2%, compared with 58.3% for Carbon, while the reported hallucination rates are 58.0% and 11.4%, respectively. The source excerpt does not provide enough detail here about the test’s full method or how its hallucination rate is defined, so the values should not be treated as universal error rates. Pricing may change, and the source does not give comparable task-cost figures for every model setting or workload.
Prices Change December 1
The next stated pricing milestone is December 1, when Google’s introductory Carbon and Argon API rates are scheduled to double, according to the pricing page cited in the report. Buyers comparing costs will need to recalculate estimates using their expected token use and workload rather than assume the current per-task figures will hold.
For now, the comparison’s practical next step is to test the candidate models on representative tasks, with the same inputs and review standards, while tracking cost and incorrect answers. Further independent evaluations or updated vendor pricing could change the picture; the source does not identify a single model as the right choice for all work.
Key Questions
Which model leads the overall ranking?
Claude Fable 5.2 scores highest on the Artificial Analysis Intelligence Index figures reported on November 17, 2026: 65 at Max effort and 62 at High. The ranking does not mean it wins every individual test.
Which model is cheaper per task?
In the reported index evaluation, Gemini 4 Hydrogen costs $3.16 per task, Carbon $5.48 and Fable 5.2 at High $9.08. These are test-specific estimates, not a promise about every user’s bill; Carbon’s figure uses an introductory price scheduled to rise December 1.
Which model performs better for coding?
Fable 5.2 leads Carbon on the cited FrontierCode-AA 2.0 open-source repository test, with scores of 49.1% and 40.6%. Carbon leads on the separate Terminal-Bench 4.0 terminal-work test, so the answer depends on the kind of coding task.
What does the reliability comparison show?
The source reports higher answer accuracy for Fable, 77.2% versus 58.3%, and a lower reported hallucination rate for Carbon, 11.4% versus 58.0%. Those figures come from a particular evaluation and should not be read as general error rates without its full methodology.
Source: ThorstenMeyerAI.com
