Claude Fable 5.2 vs Gemini 4 Carbon
| Claude Fable 5.2 | Gemini 4 Carbon | |
|---|---|---|
| Announced | ||
| Released | ||
| Input / 1M tokens | $8 | $2$4 from 1 Dec 2026 |
| Cached input / 1M tokens | $0.20 | $0.10$0.20 from 1 Dec 2026 |
| Output / 1M tokens | $40 | $10$20 from 1 Dec 2026 |
| Context window | 1M | 1M |
| Output limit | 128K | 65K |
| Benchmarks | ||
| Terminal-Bench 4.0 | 67.6%1 | 65.2%2 |
| Terminal-Bench Science 0.1 | 74.6%3 | 64.8%4 |
| AutomationBench 1.0.6 Private set | 51.4%5 | 53.9%6 |
| OSWorld 2.1 Offline set (82 tasks), partial score | 89.3%7 | 81.5%8 |
| FrontierSWE v2 | 69.9%9 | 61.2%10 |
| Chartography No tools | 73.4%11 | 72.5%12 |
| Claude Fable 5.2 | Gemini 4 Carbon | |
|---|---|---|
| Other benchmarks | ||
| Humanity's Last Exam No tools, full set (2,500 questions) | 73.2%13 | No data |
| Humanity's Last Exam With tools, full set (2,500 questions) | 74.2%14 | No data |
| SWE-bench Pro Set not stated | 93.0%15 | No data |
| OSWorld 2.1 Offline set (82 tasks), strict pass rate | 54.6%16 | No data |
| DeepSWE v1.1 | 74.3%17 | No data |
| ARC-AGI-3 Semi-Private set (55 environments) | 79.26%18 | No data |
| Chartography With tools | 91.6%19 | No data |
| SEC-bench Pro 260505 | No data | 88.2%20 |
| Agents' Last Exam v1.1 Score, all tasks (122) | No data | 77.4%21 |
| GraphWalks BFS, 256K + 1M problems | No data | 96.2% |
Notes
- Terminal-Bench 4.0, Claude Fable 5.2: Claude Code in --bare mode · max effort · 10 trials per task · no internet access · safeguards on
- Terminal-Bench 4.0, Gemini 4 Carbon: Run by Artificial Analysis · mini-swe-agent harness · high thinking · average of 3 runs
- Terminal-Bench Science 0.1, Claude Fable 5.2: Claude Code in --bare mode · highest available effort · averaged over 10 trials per task, per the system card · no internet access, per the system card · safeguards on, flagged requests answered by a fallback model (1.3% of requests), per the system card · standard error ±4.3 points, per the system card
- Terminal-Bench Science 0.1, Gemini 4 Carbon: Run by Artificial Analysis · mini-swe-agent harness · high thinking · average of 3 runs
- AutomationBench 1.0.6, Private set, Claude Fable 5.2: Run by Zapier · highest available effort · safeguards on · set and release 1.0.6 named in the system card, which prints the same figure
- AutomationBench 1.0.6, Private set, Gemini 4 Carbon: Run by Zapier · set named in the evaluation methodology the post cites
- OSWorld 2.1, Offline set (82 tasks), partial score, Claude Fable 5.2: highest available effort · safeguards on
- OSWorld 2.1, Offline set (82 tasks), partial score, Gemini 4 Carbon: Gemini CUA harness · highest thinking · max of 3 runs, one attempt per run
- FrontierSWE v2, Claude Fable 5.2: Run by Proximal · Proximal's agent harness · highest available effort · mean of 5 trials per task, per the system card · 20-hour budget · safeguards on
- FrontierSWE v2, Gemini 4 Carbon: Run by Proximal · Proximal's public leaderboard
- Chartography, No tools, Claude Fable 5.2: highest available effort · safeguards on
- Chartography, No tools, Gemini 4 Carbon: Run by Surge AI · Surge AI's public leaderboard
- Humanity's Last Exam, No tools, full set (2,500 questions), Claude Fable 5.2: max effort · thinking set to auto · average of 5 trials · 980k-token task budget per attempt, no context compaction · Claude Opus 4.6 as model grader
- Humanity's Last Exam, With tools, full set (2,500 questions), Claude Fable 5.2: highest available effort, max effort, per the system card · thinking set to auto, per the system card · average of 5 trials, per the system card · web search, web fetch, programmatic tool calling and code execution, HLE sources blocklisted, per the system card · 980k-token task budget per attempt, no context compaction, per the system card · Claude Opus 4.6 as model grader, flagged transcripts reviewed by Claude Opus 5, confirmed cheating graded incorrect, per the system card · safeguards on
- SWE-bench Pro, Set not stated, Claude Fable 5.2: max effort · adaptive thinking · average of 5 trials
- OSWorld 2.1, Offline set (82 tasks), strict pass rate, Claude Fable 5.2: max effort · mean of 5 attempts
- DeepSWE v1.1, Claude Fable 5.2: average of 5 trials
- ARC-AGI-3, Semi-Private set (55 environments), Claude Fable 5.2: Run by ARC Prize Foundation · Standard harness · high effort
- Chartography, With tools, Claude Fable 5.2: max effort · adaptive thinking · average of 5 runs · container and image-cropping tool · Gemini 3.5 Flash grader
- SEC-bench Pro 260505, Gemini 4 Carbon: Antigravity harness · pass@1, mean of 3 attempts · 183-task JavaScript engine split (May 2026) · LLM judge for crashes in both builds
- Agents' Last Exam v1.1, Score, all tasks (122), Gemini 4 Carbon: ALE-Claw harness · 5-hour window · October 17, 2026 task files · safety filters on