Gemini 4 Carbon vs Claude Fable 5.2: coding scores and cost per task
Updated November 10, 2026 · first published November 10, 2026
Gemini 4 Carbon Medium scored 60.1% at $8.49 per task on CursorBench 4.0. Claude Fable 5.2 Max scored 60.0% at $32.79. At lower budgets, Fable Medium scored 54.8% at $4.66 and Opus 5.5 High scored 56.0% at $3.97. The effort setting can change the buying decision. CursorBench leaderboard.
What did the coding benchmark actually cost?
Selected CursorBench 4.0 configurations, November 10, 2026. Source and methodology.
| Model | Effort | Score | Average cost per task |
|---|---|---|---|
| Gemini 4 Carbon | High | 60.4% | $9.22 |
| Gemini 4 Carbon | Medium | 60.1% | $8.49 |
| Claude Fable 5.2 | Max | 60.0% | $32.79 |
| Claude Fable 5.2 | High | 56.9% | $9.70 |
| Claude Fable 5.2 | Medium | 54.8% | $4.66 |
| Claude Opus 5.5 | High | 56.0% | $3.97 |
| Gemini 4 Argon | Medium | 52.8% | $4.39 |
Carbon Medium cost about 74% less than Fable Max at nearly the same reported score. Within Carbon, moving from Medium to High added $0.73 per task for a reported 0.3 percentage points. Cursor cautions that small differences can be noise. Its dollar figures apply published token prices to observed usage in its own harness. Keep that harness and rate basis attached to the figures when comparing them with your own runs. CursorBench methodology.
At the cheaper end, Carbon Medium cost $4.52 more than Opus High for 4.1 percentage points on this snapshot. That makes Opus High a useful control for a pilot, especially if the work is already going through Claude. Use the table to pick configurations to test on the same repository issues with the same tools and acceptance rules.
How does Carbon’s December price change the comparison?
Carbon and Argon share the same public Standard Gemini API rates. On December 1, each moves from $2 to $4 per million uncached input tokens and $10 to $20 for output, including thinking. Implicit cache reads rise from $0.10 to $0.20. Fable 5.2 lists $8 input and $40 output, and GPT-6.1 Sol lists $2 and $10. These are first-party API prices checked November 10, 2026. Google rates · Anthropic rates · OpenAI rates.
| Standard API model | Input / 1M | Cached input / 1M | Output / 1M | Example job |
|---|---|---|---|---|
| Carbon / Argon through Nov 30 | $2 | $0.10 | $10 | $0.58 |
| Carbon / Argon from Dec 1 | $4 | $0.20 | $20 | $1.16 |
| Claude Fable 5.2 | $8 | $0.20 | $40 | $2.24 |
| Claude Opus 5.5 | $4 | $0.20 | $20 | $1.16 |
| GPT-6.1 Sol | $2 | $0.10 | $10 | $0.58 |
The example holds usage fixed at 120,000 uncached input, 400,000 cached input and 30,000 billed output tokens per job. December Carbon costs 0.12 × $4 + 0.40 × $0.20 + 0.03 × $20 = $1.16. The scenario assumes the same cache hits on every model and excludes cache writes, tool fees and reviewer time. On Anthropic, add any cache writes at the applicable rate. The equal-token scenario isolates prices. A model that takes more steps, generates more thinking tokens or gets fewer cache hits can cost more on the same issue.
How to run a model comparison before changing your routing
- Choose actual backlog items across routine edits, ambiguous bugs and multi-file changes. Freeze the repo snapshot, tool access, time limit and acceptance tests for each issue.
- Start with Carbon Medium, Fable 5.2 Medium and one incumbent such as Opus High or GPT-6.1 Sol. Add Fable Max on the hardest items to see whether its extra spend clears an acceptance threshold the cheaper settings miss.
- Count accepted patches, then divide total cost, including failed attempts and retries, by accepted patches. Track reviewer minutes and elapsed time separately so you can see where API savings shift work onto people.
- Price the observed Carbon token buckets at December rates before making a decision. Keep that estimate beside the billed pilot cost. Use the same approach for any paid tools or caching charges in your own runs.
Frequently asked questions
Is Gemini 4 Carbon cheaper than Claude Fable 5.2?
Carbon has lower first-party uncached input and output rates. On CursorBench 4.0, Carbon Medium was cheaper than Fable Max, while Fable Medium was cheaper than Carbon Medium. Compare at the quality threshold your work requires.
Does the CursorBench comparison show how Carbon performs in Antigravity?
The published figures are from Cursor’s evaluation harness. Antigravity has its own tools, model settings and quota or billing terms. Run an Antigravity pilot to estimate those costs.
When does Gemini 4 Carbon’s API price go up?
Google lists introductory Standard pricing through November 30, 2026. From December 1, Standard is $4 per million uncached input tokens, $0.20 per million implicit cached input tokens and $20 per million output tokens.
Related
Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →