Claude Fable 5.2 vs Gemini 4 Carbon

Independent benchmarks

On Artificial Analysis, Fable 5.2 leads on eight of the eleven numbers below and Carbon on three. Fable 5.2’s clearest wins are hard knowledge and science: 80.2% against 69.1% on HLE-Diamond, and 67 against 54 on AA-Omniscience, which rewards answering correctly and penalises guessing. Carbon’s wins are practical ones: multi-app automation, terminal tasks and producing documents. On the Vals Index the two are 0.86 points apart, first and third of 36 models.

The settings differ, and that matters for cost: Artificial Analysis ran Fable 5.2 at max effort with Anthropic’s default fallback switched on, and Carbon at high. Vals does not publish the effort it used.

What Anthropic and Google report

Anthropic’s launch table includes Carbon, and it shows the same split: Fable 5.2 ahead on coding, science and computer use, Carbon ahead on AutomationBench. Note that Anthropic ran Carbon at medium effort on FrontierCode 2.0. Google’s own claims for Carbon, made at its October announcement before Fable 5.2 existed, are first place on CursorBench 4.0 and on Perplexity’s WANDR, and 92% on an agentic honesty benchmark against 70% for Gemini 4 Argon. Google also says Argon remains the better choice for most software-engineering work, so Carbon is the model to escalate to, not a default.

Price and cost per task

Per token, Carbon’s list price is a quarter of Fable 5.2’s until 30 November 2026 and half of it after Google’s 1 December change. Per task the gap is wider, because Fable 5.2 writes more: Artificial Analysis counted 274,000 output tokens per task for Fable 5.2, 193,000 of them reasoning, against 123,000 and 72,200 for Carbon. At the prices Artificial Analysis used, a task cost $26.54 on Fable 5.2 and $7.76 on Carbon; Vals, which ran a different set of tasks, reports $17.94 and $10.51 per test.

On apimodels.app the discount applies to every token type alike, so those task costs scale directly: about $15.92 per task for Fable 5.2, and about $2.13 for Carbon through 30 November (Artificial Analysis priced Carbon at Google’s $4 / $20 rate), rising to about $4.27 from 1 December. Running Fable 5.2 below max effort lowers its cost; OfficeChai’s report of the Artificial Analysis results puts it at 62 at high effort, but no cost for that run has been published.

Speed

Carbon starts answering sooner and finishes tasks faster, while Fable 5.2 writes faster once it starts. Both are reasoning models that can think for minutes on hard tasks, so stream long answers and give your client a generous timeout. These numbers come from Artificial Analysis at the settings above: Fable 5.2 at max effort and Carbon at high.

Same jobs, run by developers

Developers who gave both models the same real job mostly found the same split as the benchmarks: Fable 5.2 delivers the more complete and more reliable result, and Carbon gets close on many jobs for a small fraction of the cost. Carbon won outright in a few, and Fable 5.2 was not always better. These are single runs by different people in different tools, so read them as field reports, not a benchmark.

Two coding benchmarks run by tool makers put them level. On Cursor’s CursorBench 4.0, Carbon at high scored 60.4% at $9.22 a task and Fable 5.2 at max 60.0% at $32.79 (56.9% and $9.70 at high). On Zapier’s AutomationBench, Carbon at high scored 53.88% at $2.12 a task and Fable 5.2 at max 51.45% at $4.64. On WebDev Arena, where people vote between two generated web apps, Fable 5.2 ranks first (1836) and Carbon fourth (1768, preliminary).

Carbon’s most common failure in these reports was claiming more than it delivered. In AICodeKing’s six app builds, every one of Carbon’s final messages claimed at least one feature the reviewer found broken or overstated. If you use Carbon for builds, run the result yourself, or ask it to click through the app in a browser before it reports back.

Which to use

Hard coding, research and science questions where a better answer is worth several times the cost: Claude Fable 5.2. It leads on FrontierCode, Terminal-Bench-Science, HLE-Diamond and CritPt, and on Artificial Analysis’s hallucination-penalising knowledge test.

Automation across apps, shell work and documents: Gemini 4 Carbon. It wins AutomationBench-AA, Terminal-Bench 4.0 and GDP.pdf, and Anthropic’s own AutomationBench.

Video, audio or PDF input: Gemini 4 Carbon. Fable 5.2 accepts text and images only.

Anything you run at volume, or where a person waits for the reply: Gemini 4 Carbon, at roughly an eighth of Fable 5.2’s cost per task on our prices this month and a quarter from December, with a first answer token in about a minute instead of four.

Security research: neither by default. Fable 5.2’s safeguards refuse most high-risk cyber requests, and Google first released Carbon only to vetted cyber defenders.

A common split is to let Carbon do the volume and the first pass, and send only the tasks it fails, or the ones where the answer matters most, to Fable 5.2. On apimodels.app that is one key and a change of the model string.

Caveats

The two models were not run at the same effort, and Artificial Analysis’s Fable 5.2 run had Anthropic’s fallback switched on, so a small share of blocked requests were answered by another model. The vendor figures in the second table are each company’s own and have not been reproduced. Carbon returns at most 65,536 output tokens per response by default, against 128K for Fable 5.2, and its price doubles on 1 December 2026. These benchmarks are a few days old; rankings at the top have changed every few weeks this year.