AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Artificial Analysis data accessed Nov. 17, 2026, gives Claude Fable 5.2 a slight lead over Gemini 4 Carbon on its overall index, while Carbon leads on several workplace tasks and costs less per tested task. Fable answers more questions correctly in one benchmark, but Carbon has a lower measured hallucination rate when it does not know an answer. The results do not establish a single best model for every workplace.

Anthropic’s Claude Fable 5.2 scored slightly higher than Google’s Gemini 4 Carbon on Artificial Analysis’s overall Intelligence Index, but Carbon led several tests tied to workplace tasks and cost less per benchmark task, according to results accessed Nov. 17. The comparison suggests companies choosing between them will need to weigh the specific work they plan to automate against price and the risk of incorrect answers.

At the High effort setting, Fable 5.2 scored 62 on the Artificial Analysis Intelligence Index, compared with 60 for Gemini 4 Carbon. The listed cost per task was $9.08 for Fable and $5.48 for Carbon. At Fable’s Max setting, its score rose to 65, while the task cost rose to $26.54. These figures describe the index’s tested tasks; they are not estimates of what every company will spend on its own work.

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

The results varied by task. Fable 5.2 scored 49.1% on repository coding, against Carbon’s 40.6%, and 83.3% on the long-context test, against 79.3%. Carbon scored higher on terminal work, at 65.2% versus 59.6%; professional documents, at 28.4% versus 21.8%; and business software automation, at 77.8% versus 71.4%. The figures are results from the named independent benchmarks, not guarantees of performance in a particular workplace.

The comparison also found different patterns in answer reliability. Fable scored 77.2% accuracy on the cited benchmark, compared with Carbon’s 58.3%. But its measured hallucination rate was 58.0%, against 11.4% for Carbon. The source describes this measure as how often a model guesses incorrectly when it does not know an answer. Accuracy and hallucination rate capture different aspects of performance and should not be treated as interchangeable.

At a glance
reportWhen: Claude Fable 5.2 released Nov. 9, 2026;…
The developmentAnthropic and Google launched new models in November, prompting a comparison of their workplace performance, reliability and costs using Artificial Analysis benchmarks.
Bar chart titled The top score costs the most per task. Artificial Analysis Intelligence Index score and cost per index task: Claude Fable 5.2 (Max) 65 and 26.54 dollars; Claude Fable 5.2 (High) 62 and 9.08 dollars; Gemini 4 Carbon (High) 60 and 5.48 dollars; Gemini 4 Hydrogen (Xhigh) 54 and 3.16 dollars; Gemini 4 Argon (High) 50 and 4.97 dollars. A highlighted note says Carbon and Argon are at Google's introductory price and their token prices double on December 1
Index score and cost per task on the Artificial Analysis Intelligence Index. Source: Artificial Analysis model pages, accessed November 17, 2026; price change: Google Gemini API pricing page. Chart: Thorsten Meyer AI.

Choosing Models by Workplace Task

For employers, the results make model selection a trade-off among task performance, cost and reliability. A team working with software repositories or long documents may find Fable’s tested strengths relevant. Teams focused on terminal operations, document production or business software automation may want to examine Carbon’s stronger scores in those categories.

Costs may affect whether a model is used for occasional specialist work or deployed across routine workflows. Artificial Analysis lists one index task at $9.08 for Fable at High and $5.48 for Carbon. Google’s introductory token prices for Carbon and Argon were due to double on Dec. 1, according to the pricing information cited in the source. Token rates and benchmark task costs are different measures, and actual bills depend on token use and workload.

The reliability results matter wherever outputs are used to make decisions or prepare work for others. The reported numbers do not establish how often either model will make errors in a company’s specific setting. They do give buyers a reason to evaluate both correct-answer performance and the tendency to guess when uncertain, with human review suited to the consequences of each task.

Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Bose Portable Smart Speaker with Alexa and Google Assistant, Black

Bose Portable Smart Speaker with Alexa and Google Assistant, Black

Bose QuietComfort Wireless Noise-Cancelling Headphones, Sandstone

Bose QuietComfort Wireless Noise-Cancelling Headphones, Sandstone

As an affiliate, we earn on qualifying purchases.

November Launches and Model Tiers

Anthropic released Claude Fable 5.2 on Nov. 9. Google opened Gemini 4 Carbon to paying users on Nov. 10, then released Gemini 4 Hydrogen on Nov. 12 for all users, including through Google Search, according to the source. The comparison treats Gemini 4 as a family rather than a single product: Carbon is described by Google as its most capable reasoning model, while Hydrogen is positioned for faster execution.

Google also offers Gemini 4 Argon, which it describes as an efficient frontier model and recommends for many software engineering and knowledge-work tasks. The source says Carbon, Argon and Hydrogen each have a one-million-token context window; Fable 5.2 has a one-million-token context window and supports up to 128,000 output tokens. These specifications and descriptions come from vendor materials cited by the source, rather than from the benchmark results.

Pricing depends on the model and date. The source lists Fable at $8 per million input tokens and $40 per million output tokens, with higher rates for fast mode. Google’s listed Carbon and Argon rates were $2 per million input tokens and $10 per million output tokens through Nov. 30, rising to $4 and $20 on Dec. 1. Hydrogen was listed at $1.50 per million input tokens and $7.50 per million output tokens. These rates do not alone determine cost per completed task.

““the most ambitious, open-ended technical work””

— Anthropic

Limits of the Benchmark Evidence

The cited results come from Artificial Analysis benchmark runs, and the source does not report how closely those tasks match the workloads, data and review practices of a particular employer. It also does not provide results from a controlled study of workplace productivity, such as time saved, error rates in production or total cost across full projects.

The comparison is dated Nov. 17, 2026, so the ranking and prices could change. The source notes that Google’s introductory Carbon and Argon rates were scheduled to change Dec. 1. It does not establish how the models will perform after updates or how their relative results may shift as benchmark versions and settings change.

The benchmark figures cannot determine which model will be safer or more accurate for every use. The source reports different input capabilities across the models and does not show a single, matched evaluation covering every workplace task, language or deployment configuration. Results should be read as evidence about the cited tests, not as a universal ranking.

Testing Before Wider Rollout

Companies considering either model can compare them on representative tasks from their own workflows, record the time and token use, and check outputs for errors. High-impact uses may call for additional review, especially where a confident but incorrect answer could cause harm or costly rework.

Buyers tracking API costs should account for Google’s listed Dec. 1 price change for Carbon and Argon when estimating future spend. Further benchmark results, model updates and price changes may alter the comparison; the source does not give a date for another evaluation.

Key Questions

Which model scored higher overall in the cited comparison?

At High effort, Claude Fable 5.2 scored 62 on the Artificial Analysis Intelligence Index, while Gemini 4 Carbon scored 60. Fable scored 65 at Max effort, at a higher listed cost per task.

Which model cost less per benchmark task?

In the cited index results, Gemini 4 Carbon cost $5.48 per task at High, compared with $9.08 for Fable 5.2 at High. The figures apply to the benchmark task and are not a forecast for every company workload.

Which model performed better on coding and business tasks?

Fable 5.2 scored higher on the cited repository coding and long-context tests. Carbon scored higher on terminal work, professional documents and business software automation.

What do the reliability figures say?

Fable had higher reported accuracy, 77.2% versus 58.3%. Carbon had a lower reported hallucination rate when the model did not know an answer, 11.4% versus 58.0%. These are benchmark measurements, not guaranteed error rates in workplace use.

When were Gemini 4 prices set to change?

The source lists introductory Carbon and Argon API rates through Nov. 30, 2026, with higher rates scheduled from Dec. 1. Hydrogen’s listed rates were separate.

Source: Thorsten Meyer AI

You May Also Like

Beyond US And China, Mistral Large 4 Has Appeal—but Agents Are A Concern

Artificial Analysis scores Mistral Large 4 at 38.4, ahead of other models outside the US and China but behind major US and Chinese rivals.

10 Best Desktop Processors For Gaming, Productivity, And AI In 2027

A 2027 buyer’s guide compares 10 AMD desktop processors for gaming, creative work and budget builds, with platform and cooling tradeoffs.

Texas, Arizona And SpaceX’s Plans For AI Campus Development

SpaceX says it is advancing AI computing projects in Starbase and Cochise County. Both face local decisions scheduled for Dec. 9, 2026.

A Closer Look At 4 AI Automation Software Guides For 2027

A buyer’s guide compares four AI automation books by platform, learning path and technical focus. None is automation software itself.