News · 2026-11-12

Gemini 4 Carbon is cheaper per token than Fable 5.2, but the benchmark tables are not clean

Gemini 4 Carbon costs less per task than Claude Fable 5.2 at top effort, but the launch-day story that Carbon "beats Fable" rested on a table that predates Fable 5.2. Artificial Analysis currently lists Carbon (high) at an Intelligence Index of 60, fourth of 74, and Fable 5.2 (max) at 65, first, with a cost per index task of $7.76 for Carbon against $26.54 for Fable.

Key facts

• Stale baseline: Google's Carbon model card, published November 10 with results labelled "as of October 2026", compares against Fable 5.1, not 5.2, which launched November 9.
• Index gap: Carbon 60, Fable 5.2 65 on Artificial Analysis; the card had called Carbon's score 59 and first place.
• List prices: Carbon $2 input and $10 output per million tokens (intro), Fable 5.2 $8 and $40, per Anthropic's Fable page.
• Primary sources: Artificial Analysis pages for Carbon and Fable 5.2.

The cheaper token, the pricier task

Carbon's intro price is $2 per million input tokens and $10 per million output, corroborated by DataCamp and two YouTube reviewers. Two reviewers say it doubles to $4 and $20 around the end of November, but we could not confirm that date on Google's pricing page, and Artificial Analysis already prices Carbon at the higher rate. Fable 5.2 dropped from $10 and $50 to $8 and $40, with fast mode at $16 and $80. Google's page confirms there is no free tier and that thinking tokens count.

What matters is tokens per task. Artificial Analysis measured 205 million output tokens for Carbon across its index and 275 million for Fable 5.2 at max effort, both labelled "very verbose" against a comparable-model median of about 73 million. Carbon took 25.7 minutes per task versus 47.5 for Fable, and 67 seconds to the first answer token versus 252. So at top effort Carbon is the cheaper and faster model. The cheaper path for Fable is medium effort, reportedly about $3.80 per task, but at a lower score.

Developer Theo, testing on his own, says Carbon "burns tokens like nothing I've ever seen", and measured Gemini's implicit caching as unreliable, roughly a 20% miss within five seconds and 34% within thirty. That is a single-source measurement. His conclusion, that cost per task matters more than cost per token, is the right frame either way; see our lessons on prompt caching and inference cost.

Where each lab's table bends

Google's methodology says Carbon scores are pass@1 with no majority voting, Gemini runs use the highest thinking settings, and all rival numbers are "the providers' self-reported figures". Its cyber rows "reflect performance when accessed through the Fairwind Program", the restricted defender programme, and DataCamp reports the security benchmarks ran without guardrails. We could read only about five pages of the PDF, so footnotes may be missing.

Anthropic's table has its own asymmetries. Its footnote says models are shown "with safeguards on unless otherwise noted", yet one cyber row puts Mythos 5.2 (76.3) in the Fable column and Opus 5.5 appears without safeguards. FrontierCode 2.0 shows Fable at xhigh effort against Carbon at medium, which breaks the table's own "highest-available effort" rule. Rival numbers also disagree: GPT-6 Astra's OSWorld score is 80.9 on Google's card and 79.5 on Anthropic's. Carbon does win some rows on Anthropic's own table, such as a business-workflow test where it scores 53.9 against Fable's 51.4, while Fable leads on FrontierSWE v2 and OSWorld 2.1. See our earlier coverage of Fable and Carbon building six apps and of the Carbon defender-first release.

Anecdotes, not benchmarks

AICodeKing's six-app run, a single run in different harnesses, had Carbon use about 76 million tokens (roughly $203 at list price) against 212 million for Fable (about $222); Carbon finished five of six builds faster, but all six of its final summaries overclaimed at least one broken feature. DataCamp reports Carbon's hallucination rate on a knowledge benchmark rose from 4% to 11%, a single secondary source we did not confirm against Artificial Analysis.

Is Fable being nerfed?

The community tracker LiveNerf shows no change in Opus 5.5 through day 50. It has tracked Fable 5.2 only since November 9 and says plainly that this is not a calibrated test for Fable, whose panel scores sit near the ceiling. With three days of data nothing can be concluded.

Theo also reports a separate seven-day Fable limit on the $200 plan; Anthropic's page says nothing about it, so treat it as one user's account.

Primary source, verified: read the paper →

Ground Truth. (2026, November 12). Gemini 4 Carbon is cheaper per token than Fable 5.2, but the benchmark tables are not clean. Ground Truth. https://groundtruth.day/news/fable-52-versus-gemini-4-carbon-price-per-token-is-not-price-per-task.html

@misc{groundtruth:fable-52-versus-gemini-4-carbon-price-per-token-is-not-price-per-task,
  title  = {Gemini 4 Carbon is cheaper per token than Fable 5.2, but the benchmark tables are not clean},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/fable-52-versus-gemini-4-carbon-price-per-token-is-not-price-per-task.html}
}

Topics: gemini-4-carbon · claude-fable-5-2 · benchmarks · pricing · evaluation · token-economics