As of 2026-11-16, Epoch AI lists Gemini 4 Carbon with an ECI of 170 and a current rank of #2/240. It ranked #1/238 at launch. The component results show strong scores on the listed board-game and game tasks, math benchmarks, and SimpleQA Verified. However, each benchmark measures different capabilities, while ECI is a composite scale fitted across more than 50 benchmarks. It cannot replace individual results or represent performance on every task.
Suitable for: Reviewing Epoch AI's current model profile and in-house benchmark results, and comparing Carbon with the models selected on the page using the same page-level criteria.
Not suitable for extrapolating to: Inferring performance on a specific task from ECI alone, generalizing a few benchmarks or a high score on one benchmark to broad capabilities, or treating results from different evaluation setups as direct replications under identical conditions.
Applicable model version: The model Epoch AI identifies as Gemini 4 Carbon. The page does not specify a particular model revision.
Test environment or client: Epoch AI's model profile and individual benchmark records. The model page does not provide a unified list of specific runtime environments or complete harnesses.
Reasoning tier and parameters: Not specified on the model page.
The Epoch Capabilities Index (ECI) puts results from more than 50 benchmarks on a shared general-capability scale. Epoch AI estimates the relative difficulty of each benchmark from models' overlapping results across multiple benchmarks, then fits model capabilities. ECI is a composite metric, not a raw score from a single test. Epoch AI explains that the ECI score itself is an abstract value. Its benchmark curves show how it corresponds to scores on specific benchmarks. See the ECI documentation and the papers it cites for the detailed methodology.
The model summary lists Carbon's current ECI as 170 and its rank as #2/240, and notes that it ranked #1/238 at launch. The full results page also shows “best across settings” in parentheses, for example, 91% (95%) for FrontierMath Tiers 1–3. The direct benchmark scores in the table and the best scores in parentheses should therefore be read separately.
Some Epoch AI benchmarks are developed in-house, while some data comes from benchmark providers. Their samples, scoring methods, and task formats differ. Cross-model comparisons in the table should be limited to the same benchmark and the result criteria shown on the page.
The model page selects Gemini 4 Argon, GPT-6 Astra, and Claude Opus 5.5 for comparison:
| Category | Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Composite | ECI rank | #2 | #9 | #3 | #4 |
| Games | Chess Puzzles | 78% | 59% | 76% | 67% |
| Games | Mystery Game Puzzles | 73% | 64% | 84% | 71% |
| Games | Earthborne Rangers (EBR-bench) | 88% | 65% | 76% | 71% |
| Math | FrontierMath Erdős | 3% | 1% | 3% | 3% |
| Math | FrontierMath Tiers 1–3 (v2) | 91% | 86% | 94% | 91% |
| Math | FrontierMath Tier 4 (v2) | 95% | 78% | 98% | 95% |
| World knowledge | SimpleQA Verified | 78% | 76% | 76% | 72% |
Carbon scores higher than Argon and Opus 5.5 on all three listed game benchmarks and on SimpleQA Verified. Its math results vary by benchmark: it scores below Astra on Tiers 1–3 and Tier 4, and ties Astra and Opus on Erdős. These are component-level comparisons among the models selected on the page, not an overall ranking across other tasks.
The full results page also shows Carbon's best scores in parentheses: 79% on Chess Puzzles, 90% on Mystery Game Puzzles, 90% on EBR-bench, 4% on FrontierMath Erdős, 95% on Tiers 1–3, 100% on Tier 4, and 94% on SimpleQA Verified. The page identifies these as best scores across settings. They should not be substituted for the component scores in the table.
ECI: 170. The value in parentheses on the full results page is 173. Current rank: #2/240. Rank at model release: #1/238.
Chess Puzzles: 78% (best: 79%). Epoch AI generates 100 new chess-position puzzles. Stockfish determines one best next move for each puzzle, and the model is scored on whether it selects that move. The benchmark is mainly a lightweight proxy for spatial reasoning and planning, not a measure of broad game-playing ability.
Mystery Game Puzzles: 73% (best: 90%). The 100 puzzles ask the model to select the best next move, but the game identities are not disclosed to reduce the risk of benchmark-specific preparation. The score still reflects only this set of puzzle tasks.
Earthborne Rangers (EBR-bench): 88% (best: 90%). This evaluation tests whether a system improves after repeatedly playing an unfamiliar, complex board game and recording its experience. By default, each sample includes 8 learning games and 2 scoring games, and the page reports the higher score from the two scoring games. A single topline score therefore does not reveal the learning curve. Epoch AI also notes that recent results use the Card ban setting, so older results may differ in setup.
FrontierMath: 91% on Tiers 1–3 (best: 95%), 95% on Tier 4 (best: 100%), and 3% on Erdős (best: 4%). Tiers 1–3 (v2) consists of difficult math problems written and reviewed by experts. The version notes that the dataset has 338 problems: 295 in Tiers 1–3 and 43 in Tier 4. Few problems are public, so leaderboard numbers usually correspond to private test sets. In a June 2026 update, Epoch AI corrected errors in 42% of the problems. The model page does not detail the sample or scoring methods for FrontierMath Erdős.
SimpleQA Verified: 78% (best: 94%). This set contains 1,000 factual question-answering questions across several everyday knowledge domains. Epoch AI says it was created by Google based on earlier work, with improvements including filtering similar questions, balancing topics, and verifying answers. It measures factual question-answering on this set, not all knowledge or factual reliability.
ECI and component scores are different levels of measurement. ECI aggregates more than 50 benchmarks to support comparisons on a general-capability scale. The percentages in the table are results on specific benchmarks. ECI cannot replace task-level scores, and an individual score cannot represent composite capability.
Component results vary within the same broad category. Carbon leads the listed Argon and Opus 5.5 models on the two listed game scores and SimpleQA Verified. Its math performance is not uniformly ahead: it scores below Astra on Tiers 1–3 and Tier 4, and ties Astra and Opus on Erdős.
Best scores across settings do not guarantee results under uniform conditions. The page labels the parenthetical values as best scores across settings, but does not list the settings, number of runs, or intervals for each result. Comparisons should prioritize consistent component-level criteria from the page while accounting for this information gap.
Rankings are time-sensitive snapshots. #2/240 is the value shown on the page on the collection date. Epoch AI also records #1/238 at launch. New models or data updates can change rankings, so the current rank is not a fixed conclusion.
Evaluation coverage is limited. The seven results shown on the page cover specific datasets and task designs. Some are developed by Epoch AI and others by different organizations. Because sources and harnesses are not fully consistent, percentages across benchmarks should not be compared directly.
Open the Epoch AI Gemini 4 Carbon model page and record the ECI summary, current rank, rank at release, comparison table, and the best scores in parentheses under “Full results.” Then open the relevant benchmark documentation to review sample design and scoring limitations. Access date: 2026-11-16. The model page does not provide a unified prompt, parameters, complete logs, or model revision information sufficient to independently rerun all results.
Gemini 4 Carbon