The short answer
● Cheapest frontier answer today — Gemini 4 Carbon at $2 / $10 per million, through 30 November.
● Best scores on coding and science — Claude Fable 5.2, by clear margins on FrontierCode, FrontierSWE and Terminal-Bench-Science.
● Best cost per finished task — Carbon, at roughly $7.76 against Fable at roughly $26.54.
● If you are already on Opus 5.5 — Stay. It is $4 / $20 and Anthropic still recommends starting there.
Google released Gemini 4 Carbon on 10 November, one day after Claude Fable 5.2, and priced it at a quarter. That is the whole story in one line, and it is also misleading, because the two models win different work.
Price
|
Gemini 4 Carbon |
Claude Fable 5.2 |
| Input / MTok |
$2 to 30 Nov, then $4 |
$8 |
| Output / MTok |
$10 to 30 Nov, then $20 |
$40 |
| Fast mode |
Not offered |
$16 / $80, up to 2.5× faster |
| Context |
1M |
1M |
| Max output |
Not published |
128K |
| Inputs |
Text, images, speech, video |
Text, images |
| Cost per task Artificial Analysis |
~$7.76 |
~$26.54 |
Fable prices are from Anthropic’s model page. Carbon’s are as reported at launch; Google has not yet added Carbon to its public AI Studio pricing page.
The multiplier changes on 1 December. Until then Carbon is a quarter of Fable; after, half. Either way the cost-per-task figure is the one with teeth: roughly a third, because cheaper tokens compound across an agentic run.
Benchmarks
| Benchmark |
Carbon |
Fable 5.2 |
Winner |
| Artificial Analysis Intelligence Index |
60 |
65 |
Fable |
| Zapier AutomationBench |
53.88% |
51.45% |
Carbon |
| Terminal-Bench 4.0 |
65.2% |
61.6% |
Carbon |
| FrontierCode 2.0 |
40.6% |
49.1% |
Fable |
| FrontierSWE v2 |
61.2% |
69.9% |
Fable |
| Terminal-Bench-Science 0.1 |
64.8% |
74.6% |
Fable |
| Cost per task |
~$7.76 |
~$26.54 |
Carbon |
Opus 5.5 scores 59 on the Intelligence Index and GPT-6 Astra 57. FrontierSWE and Terminal-Bench-Science figures are Anthropic-reported; AutomationBench and Terminal-Bench 4.0 favour Carbon and were cited at Google’s launch. Both vendors publish the benchmarks they win, which is why the split below matters more than any single row.
The pattern is consistent and useful. Carbon wins process work — business workflow automation, terminal tasks, long multi-step runs. Fable wins hard reasoning — writing code, software engineering, scientific research. That is not a quality gap, it is a specialisation.
Which to use
- Automating business processes at volume → Carbon. It wins the benchmark and costs a third per task.
- Writing or fixing code → Fable 5.2. Nine points on FrontierCode and nearly nine on FrontierSWE is not close.
- Scientific or research work → Fable 5.2, by ten points on Terminal-Bench-Science.
- Long agentic runs where cost compounds → Carbon, unless the task is one of the two above.
- Audio or video input → Carbon. Fable takes text and images only.
- You have not measured anything yet → Opus 5.5 at $4 / $20. It scores 59 against Carbon’s 60 and costs the same as Carbon will from December.
The timing nobody is mentioning
Carbon’s $2 / $10 ends on 30 November. Gemini 4 Argon doubles on 1 December, from $2 / $10 to $4 / $20. If you are building a cost model on Gemini this month, you are modelling on numbers with weeks left on them. Fable’s $8 / $40 has no announced end date, and is itself 20% below Fable 5.1.
Frequently asked questions
Is Gemini 4 Carbon better than Claude Fable 5.2?
Not overall. Fable scores 65 on the Artificial Analysis Intelligence Index against Carbon’s 60, and wins FrontierCode, FrontierSWE and Terminal-Bench-Science. Carbon wins Zapier AutomationBench and Terminal-Bench 4.0, and costs roughly a third per finished task.
Which is cheaper?
Carbon, clearly. $2 / $10 per million through 30 November against Fable’s $8 / $40, becoming $4 / $20 in December. On Artificial Analysis cost per task it is about $7.76 against about $26.54.
Should I switch from Fable 5.2 to Carbon?
For business process automation and long agentic runs, price it out — the saving is large and Carbon wins the relevant benchmarks. For coding or research, no; the gap runs the other way and the cheaper tokens will not recover it.
Does Gemini 4 Carbon have a free tier?
It is available on the paid API tier in Google AI Studio and Android Studio, and to Google AI Ultra subscribers in AI Studio and Google Antigravity. No free API tier has been announced.