Gemini 4 Carbon (high): benchmark scores, price and ranking
Gemini 4 Carbon (high) from Google DeepMind ranks #3 of 827 LLMs on the LLMs Tiger aggregate benchmark score (62.4), based on 17 benchmark results captured up to 2026-11-18.
- Organisation
- Google DeepMind
- Type
- Frontier model
- Open weights
- No
- Released
- 2026-10-28 (source: epoch)
- Cheapest API price
- $2.00 in / $10.00 out per million tokens (via gemini, observed 2026-11-18T15:47Z)
- Aggregate rank
- #3 of 827 (score 62.4)
Relative to other models, Gemini 4 Carbon (high) is strongest on GPQA Diamond (#2 of 302) and weakest on PostTrainBench (#5 of 13).
Across the 17 benchmarks where at least ten models are scored, Gemini 4 Carbon (high) is in the top 10% on 13 and in the top half on 17.
Its cheapest listed input price, $2.00 per million tokens, is lower than that of 18% of the 443 models with a listed price.
Gemini 4 Carbon (high) benchmark results
About Gemini 4 Carbon (high)
Gemini 4 Carbon (high) is a frontier model from Google DeepMind. It was released on 2026-10-28, according to Epoch AI. The cheapest API listing we track charges $2.00 per million input tokens and $10.00 per million output tokens, through gemini. We hold 17 benchmark results for it across 4 areas, and 13 are in the top tenth of their boards.
Where Gemini 4 Carbon (high) stands
On our aggregate score, Gemini 4 Carbon (high) is #3 of 827 ranked models, in the top ten. Its score is 62.4; the scale is centred on 50, so it sits 12.4 points above the typical model. That beats 100% of the ranked field. Among the 44 ranked Google DeepMind models it is #1. The next one down is Gemini 4 Hydrogen (xhigh) at 54.7.
Gemini 4 Carbon (high), benchmark by benchmark
On GPQA Diamond, Gemini 4 Carbon (high) scores 96.5% and ranks #2 of 302. Only one model scores higher, led by GPT-6.2 Astra (max) at 96.8%, which is 0.3 points ahead. It covers expert-level science questions. The board median is 75.1%, so Gemini 4 Carbon (high) is 21.4 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
CritPt is where Gemini 4 Carbon (high) does best: 36.0%, first of 138. The next result is Claude Fable 5.2 (max) at 35.4%, 0.6 points behind. It covers unpublished research-level physics problems. The board median is 2.0%, so Gemini 4 Carbon (high) is 34.0 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
On OTIS Mock AIME, Gemini 4 Carbon (high) scores 100.0% and ranks #4 of 206. Only 3 models score higher, led by Claude Fable 5.1 (max) at 100.0%, which is 0.0 points ahead. It covers competition maths (AIME 2024-25 style). The board median is 66.4%, so Gemini 4 Carbon (high) is 33.6 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
On Chess Puzzles, Gemini 4 Carbon (high) scores 78.0% and ranks #2 of 84. Only one model scores higher, led by Claude Fable 5.2 (max) at 79.0%, which is 1.0 points ahead. It covers chess tactics puzzles. The board median is 33.0%, so Gemini 4 Carbon (high) is 45.0 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
On SciCode, Gemini 4 Carbon (high) scores 64.2% and ranks #3 of 113. Only 2 models score higher, led by Claude Fable 5.2 (max) at 67.7%, which is 3.5 points ahead. It covers scientific research code. The board median is 46.1%, so Gemini 4 Carbon (high) is 18.2 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
GDP.pdf is where Gemini 4 Carbon (high) does best: 56.4%, first of 37. The next result is Muse Wonder (max) at 54.8%, 1.6 points behind. It covers economically valuable document tasks. The board median is 32.8%, so Gemini 4 Carbon (high) is 23.6 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
EBR-Bench puts Gemini 4 Carbon (high) at 87.6%, #3 of 27. That beats 89% of the models tested, but 4.8 points short of GPT-6.2 Astra (max) (92.4%). The results either side are Claude Fable 5.2 (max) (90.5%) and GPT-6.1 Terra (max) (78.1%). It covers evidence-based reasoning. The board median is 39.5%, so Gemini 4 Carbon (high) is 48.1 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
On Furniture Assembly, Gemini 4 Carbon (high) reaches 77.5%. It ranks #7 of 36, ahead of 81% of the field, and 12.5 points behind Claude Fable 5.2 (max) at 90.0%. The results either side are GPT-6.1 Terra (max) (79.2%) and Claude Sonnet 5.5 (max) (75.0%). It covers planning a physical assembly. The board median is 44.2%, so Gemini 4 Carbon (high) is 33.3 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
Gemini 4 Carbon (high) manages 61.2% on FrontierSWE, good for #5 of 24. Claude Fable 5.2 (max) leads at 69.9%, 8.7 points further up, and 79% of models sit below. The results either side are Claude Sonnet 5.5 (max) (61.9%) and Claude Fable 5.1 (max) (56.3%). It covers hard multi-hour software engineering. The board median is 29.8%, so Gemini 4 Carbon (high) is 31.4 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
PostTrainBench puts Gemini 4 Carbon (high) at 41.0%, #5 of 13. That beats 62% of the models tested, but 3.7 points short of Claude Fable 5.1 (max) (44.6%). The results either side are GPT-6 Astra (max) (41.9%) and gemini-4-argon-high (40.4%). It covers post-training a smaller model. The board median is 38.8%, so Gemini 4 Carbon (high) is 2.2 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-11-18.
The full table above also lists the other 7 results, including boards with fewer than ten models, where a rank says little.
Strengths and weaknesses
Software engineering: strong, ahead of 92% of the field on average across 4 benchmarks. Its best board here is SciCode (#3 of 113) and its weakest is FrontierSWE (#5 of 24). Reasoning, maths and science: strong, ahead of 96% of the field on average across 6 benchmarks. Its best board here is GPQA Diamond (#2 of 302) and its weakest is FrontierMath T1-3 (#7 of 78). General knowledge and chat: strong, ahead of 86% of the field on average across 6 benchmarks. Its best board here is Chess Puzzles (#2 of 84) and its weakest is PostTrainBench (#5 of 13).
Price and value
The median input price across the 443 priced models is $0.482 per million tokens, so Gemini 4 Carbon (high) costs more than the typical listing. We found no cheaper model with a similar or better aggregate score, so Gemini 4 Carbon (high) is the lowest-priced option at its level. Similar scores cost more on Claude Opus 5.5 (max) ($4.00) and Claude Fable 5.2 (max) ($8.00).
Release history
The Google DeepMind model released before it was Gemini 3.8 Flash (high) on 2026-09-02, which scores 53.4. The next one was Gemini 4 Hydrogen (xhigh) on 2026-11-12, at 54.7.
Reading these numbers
The 17 scores on this page come from 2 sources: Epoch AI Benchmarking Hub (16), ARC Prize official leaderboard (1). 1 of them come from an official leaderboard and 16 from datasets or third-party runs. The leaderboard also lists 4 other reasoning-effort settings of this model, which redirect here. This page shows the highest.
Gemini 4 Carbon (high) is also listed at 4 other reasoning-effort settings; this page shows the highest-effort row, which is the one the leaderboard keeps by default.
Models ranked nearby
- #1 Claude Fable 5.2 (max) (63.4)
- #2 Claude Opus 5.5 (max) (63.1)
- #4 GPT-6 Astra (max) (62.3)
- #5 Claude Fable 5.1 (max) (61.4)
- #6 Claude Sonnet 5.5 (max) (60.7)
More from Google DeepMind
- Gemini 4 Hydrogen (xhigh) (#25, 54.7)
- Gemini 3.8 Flash (high) (#38, 53.4)
- Gemini 3 Pro Preview (#41, 53.3)
- Gemini 3.7 Flash (high) (#66, 52.6)
- Gemini 3 Deep Think (#78, 52.4)
- AI co-mathematician (#107, 52.0)
Compare it on the live leaderboard → · How the aggregate score works