Google DeepMind's model card reports Carbon's results on several knowledge work, chart and blueprint understanding, long-context, and math and science benchmarks. These are vendor-reported benchmark snapshots, and the card does not provide enough detail to independently reproduce them or assess statistical uncertainty.
Tasks these results can inform: Understanding vendor-reported results for knowledge work, data analysis, chart and spatial understanding, long-context retrieval, and mathematical and scientific reasoning, as well as the model's stated intended uses.
Tasks these results should not be generalized to: They do not support inferences about success rates across all real-world workflows, relative rankings, or costs. The card provides no sample sizes, confidence intervals, or complete test configurations, and presents no independent user-facing retests.
Model version covered: Gemini 4 Carbon. The card says it is based on Gemini 4 Argon and uses more thinking tokens than Argon. It does not specify an API model ID or revision number.
Test environment or client: A summary in the official model card. The specific environment, model invocation version, and harness for most benchmarks are not specified. AutomationBench is labeled Zapier, and OSWorld-2.1 uses an offline subset.
Reasoning tier and parameters: No specific tier, sampling parameters, or token configuration is specified. The card says thinking levels are configurable to balance quality, latency, and cost.
The card says the evaluation covers knowledge work, multimodal, long-context, science, and math categories, with results dated as of October 2026. It links to a separate Gemini 4 Carbon evaluation methodology page, but this card does not list the sample count, prompts, number of runs, statistical intervals, or complete harness for each test. The scores therefore cannot be reproduced from this page alone, nor can it confirm that testing conditions were fully consistent across models.
The table below includes only knowledge work, multimodal, long-context, and math and science results. All values are reported by Google DeepMind in the model card and are not independent measurements. The original card labels GraphWalks as 256k to 1M, BFS (F1). Notes in the Chartography and Blueprint-Bench 2 table headers specify the score type and similarity score, respectively.
| Category | Benchmark and card note | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|---|
| General capability | AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| General capability | Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| Knowledge work | AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Knowledge work | Vending-Bench 2 (bank balance, USD) | $16,484 | $13,718 | $15,515 | $9,235 | $5,422 |
| Multimodal understanding | Chartography (Score) | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Multimodal understanding | Blueprint-Bench 2 (Similarity score) | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| Long context | GraphWalks (256k–1M, BFS, F1) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| Science and math | RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Science and math | Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
The card also says the model is suited to high-precision knowledge work, artifact generation, and data analysis, as well as engineering tasks that require multi-agent collaboration. These stated uses do not mean the benchmarks in the table validate every such work scenario.
The date associated with the official results page: October 2026.
Inputs include text, images, audio, and video. The combined input and output context limit is 1M tokens. Outputs are text, and the card also states the combined input and output limit as 1M tokens.
Knowledge cutoff: March 2026. The card notes that knowledge in some fields may be older.
Details such as training data, training data processing, architecture, hardware, software, acceptable use, and evaluation methods are often cross-referenced to the Gemini 4 Argon model card and are not expanded on this page.
On the two knowledge work benchmarks listed in the card, Carbon scores higher than the other models in the table. This describes only the results listed in this card and does not represent success rates on general office tasks or the quality of real-world output.
The multimodal understanding results include 72.5% on Chartography and a similarity score of 58.6 on Blueprint-Bench 2. The benchmark names and metrics differ, so the scores should not be compared with each other. The card does not specify the image sample composition, scoring criteria, or error range.
The GraphWalks F1 score is 96.2%, but this result applies only to the 256k to 1M, BFS setup specified in the card. It cannot be directly generalized to all long-document question answering or context lengths.
The math and science results are for specific benchmarks. The model card does not give the test set size, number of repeated runs, or verifiable item-by-item answers, so they cannot establish reliability in open-ended research.
The card lists general limitations including hallucinations, jailbreak resistance that is still improving, occasional slow responses or timeouts, and potentially higher token consumption at high thinking levels. Details on specific limitations are partly cross-referenced to the Argon model card.
The scores lack publicly available sample sizes, confidence intervals, per-benchmark parameters, and complete harnesses. Benchmark coverage and metrics also vary. Cross-model differences should be treated as reported values in the vendor's table, not as validated general advantages.
This page does not provide enough information to independently reproduce the reported results. Retesting would also require the relevant benchmark versions and data splits, prompts and harnesses, model/API revision, thinking configuration, sampling parameters, number of runs, and scoring implementation. Google DeepMind links to a separate methodology page in the card. For reproduction, that page and the original benchmark materials should be checked, and any configuration differences from this card should be recorded.
Gemini 4 Carbon