For long-horizon coding, Fable 5.2 leads the listed hard-code evaluations. Carbon scores higher on the listed Terminal-Bench 4.0 and AutomationBench-AA rows, and its introductory token rate is lower.
Artificial Analysis's November 11 snapshot puts Fable High at 61.82 (#3) and Carbon High at 60.03 (#4). Fable High costs $9.08 per task and takes 17.3 minutes; Carbon High costs $7.76 and takes 25.6 minutes. At Medium, Carbon's AA index is higher (58.26 vs 56.41), but Fable is cheaper and faster: $3.80 and 7.8 minutes versus Carbon's $6.59 and 23.3.
| Evaluation | Carbon High | Fable High |
|---|---|---|
| Terminal-Bench 4.0 | 65.15 | 59.6 |
| FrontierCode 2.0 | 40.56 | 49.13 |
| AutomationBench-AA | 77.82 | 71.36 |
| HLE-Diamond | 69.1 | 73.8 |
| Non-hallucination | 88.58 | 42.02 |
| Accuracy | 58.27 | 77.23 |
Which model leads on each test at High effort? Subtracting Fable from Carbon makes the direction visible; the table above keeps the absolute scores.
High effort · Carbon minus Fable, percentage points. Source: Artificial Analysis, 11 November 2026.
Positive values favour Carbon; negative values favour Fable. These independent tests do not form a combined quality score: a larger gap on one benchmark does not imply a larger gain in general ability. In particular, non-hallucination and accuracy measure different outcomes.
Anthropic's own table ran Fable at Xhigh and Carbon at Medium: Fable led FrontierCode 49.1 to 41.1 and OSWorld 89.3 to 81.5; Carbon led AutomationBench 53.9 to 51.4. These are vendor-reported results on a different setup.
Deployment differences
Carbon's Google-native context is 1,048,576 tokens, with low/medium/high thinking and a 65,536 standard output cap. Google-native long decode can reach up to 1,048,576 minus input, but AIHubMix currently documents only text input and Chat Completions for Carbon; its larger output path, native context and tool support are unconfirmed on the gateway. Fable has 1M context, adaptive thinking that cannot be disabled, 128,000 output, text, image and PDF input, and a 30-day retention requirement with no ZDR unless authorized. Forced tool use (tool_choice any/tool) returns 400 on Fable.
Fable's official cache-read price is $0.20 per million tokens, alongside $8/$40 input/output. Carbon is $2/$10 until November 30, then $4/$20. AIHubMix lists Carbon at Google's introductory rate; the per-task numbers above use AA's standard prices.
One key, two IDs on AIHubMix
AIHubMix documents Chat Completions for Carbon and Chat Completions, Responses and Messages for Fable. Carbon's documented input is text. The gateway may retry another provider, so provider behaviour can vary. No live smoke test was run for this post; the sample only shows the request shape.
for model in ("gemini-4-carbon", "claude-fable-5-2"):
response = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": "Reply with OK."}],
max_tokens=4096,
)
Use Fable for hard code and image input when its retention policy is acceptable. Use Carbon for the listed terminal and automation benchmarks, higher non-hallucination score and the lower High-effort task cost before December 1. Neither is a chat-speed model at High effort.
Keep reading: the Gemini 4 Carbon series
- Sibling trade-off: Gemini 4 Carbon vs Argon: Same Price, When Is Carbon Worth It?
- Migration details: Gemini 4 Carbon API Migration: 400 Errors, Timeouts and Output Limits
Top comments (0)