TL;DR
Anthropic’s Claude Fable 5.2 scored highest on the Artificial Analysis Intelligence Index, while Google’s Gemini 4 Carbon led several individual work tests at a lower reported cost per task. The comparison, accessed November 17, 2026, shows trade-offs in coding, documents, price and how often each model guesses incorrectly; results may vary by task and setting.
At the same High effort setting, Fable 5.2 scored 62 and Gemini 4 Carbon scored 60. Artificial Analysis reported a cost of $9.08 per index task for Fable and $5.48 for Carbon. These figures describe the index task, not a fixed price for every business workload.
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Fable led Carbon in the comparison’s open-source repository coding test, scoring 49.1% to Carbon’s 40.6%, and in long-context testing, 83.3% to 79.3%. Carbon scored higher in terminal work, professional documents and business software automation: 65.2% to 59.6%, 28.4% to 21.8%, and 77.8% to 71.4%, respectively.
The reported reliability measures also diverged. Fable answered more questions correctly in the cited test, with 77.2% accuracy versus Carbon’s 58.3%. But its reported hallucination rate was 58.0%, compared with 11.4% for Carbon. Those metrics come from the cited benchmark suite and should not be treated as a guarantee of performance in a particular company’s use case.

Benchmarks Point to Different Workloads
For teams selecting a model for coding, research or document-heavy work, the overall index is only one input. The results suggest Fable may suit repository coding and long documents, while Carbon performed better on the measured terminal, document and business-software tasks. A team’s own prompts, integrations and review needs may produce different results.
Price can also shape deployment choices. In the comparison, Carbon’s index task cost was lower than Fable’s at High effort. Google’s listed introductory prices for Carbon and Argon were due to rise on December 1, 2026, from $2 per million input tokens and $10 per million output tokens to $4 and $20. Token rates alone do not establish total workload cost, since usage and model reasoning affect how many tokens a task consumes.
The reliability figures matter when outputs can trigger decisions or actions. The benchmark indicates a difference between answering correctly when the model responds and recognizing when it lacks an answer. Businesses should assess both behavior and the cost of checking results before assigning consequential work.
Products Worth Considering

Samsung Galaxy Tab A11+ Tablet (6GB RAM, 128GB Storage, Gray)

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones (Black)

DJI Osmo 360 Road Cycling First-Person Combo Action Camera
As an affiliate, we earn on qualifying purchases.
Two Launches, Several Model Tiers
Anthropic released Claude Fable 5.2 on November 9. Google opened Gemini 4 Carbon to paying users on November 10, then released Gemini 4 Hydrogen on November 12. Gemini 4 is a family: Carbon is positioned as its most capable reasoning model, while Argon is described by Google as an efficient option for common work. Hydrogen is designed for faster agentic workflows and is also used in Google Search, according to the source material.
The comparison includes more than the two top models. At their reported highest settings, Fable scored 65 at Max and 62 at High; Carbon scored 60 at High; Argon scored 50 at High; and Hydrogen scored 54. Their reported index-task costs ranged from $3.16 for Hydrogen to $26.54 for Fable at Max. Differences in effort settings mean these figures are not all like-for-like comparisons.
Both vendors recommend models below their top tier for many users. Anthropic recommends starting with Claude Opus 5.5 for most work, while Google calls Argon “the ideal choice for the majority of software engineering and knowledge work.” Those are vendor recommendations, distinct from the independent benchmark results.
““the ideal choice for the majority of software engineering and knowledge work””
Limits of the Published Comparison
The source material does not establish how these benchmark scores translate to a particular company’s applications, data, prompts or review process. It also does not provide a full account of benchmark uncertainty or explain how much performance changes across repeated runs. The figures are a snapshot accessed November 17, 2026, and model versions, pricing and availability can change.
It remains unclear which model would produce the lowest total cost for a given team. The per-task figures cover a benchmark task, while actual costs depend on workload and token use. The cited accuracy and hallucination results also do not establish how either model performs across all kinds of questions or business decisions.
Test Against Your Own Work
Teams weighing the models can compare them on representative tasks from their own workflows, tracking output quality, error rates, review time and total usage cost. They should account for effort settings and check current prices, including Google’s stated December 1 price change for Carbon and Argon.
The source material does not identify a scheduled follow-up benchmark or a next product milestone. The practical next step is to watch for updated model and pricing information, then reassess results against each team’s requirements.
Key Questions
Which model scored higher overall?
Claude Fable 5.2 scored 62 at High effort on the Artificial Analysis Intelligence Index, compared with 60 for Gemini 4 Carbon at High, in results accessed November 17, 2026.
Which model was cheaper in the comparison?
Gemini 4 Carbon’s reported cost was $5.48 per index task, compared with $9.08 for Fable 5.2 at High effort. This is a benchmark-specific figure, not a universal cost for business tasks.
Which model performed better on coding?
Fable 5.2 scored 49.1% versus Carbon’s 40.6% on the cited open-source repository coding test. Carbon scored higher on the separate terminal-work test.
What did the hallucination figures show?
The cited benchmark reported a 58.0% hallucination rate for Fable 5.2 and 11.4% for Carbon. The source also reported higher accuracy for Fable, so the measures reflect different aspects of performance and do not guarantee outcomes on a company’s own tasks.
Source: Thorsten Meyer AI
