Google: Gemini 4 Carbon
Google frontier model announced October 28, 2026 for long-horizon engineering, agentic tasks and cybersecurity. Google plans initial access for a subset of cyber defenders, followed by trusted Fairwind partners; developer, business and consumer availability is planned. Announced introductory price: $2 input / $10 output per million tokens, $0.10 cached input; $4 / $20 after the introductory period, whose end date is unspecified. API model ID, context and output limits have not been published; the catalog ID is a placeholder.
Sources
Same model, different sources. Our first-party rate is listed first, then router and aggregator rows stay separate so you can see the spread instead of a single blended number.
| Source | Provider | Input | Output | Cache read | Context |
|---|---|---|---|---|---|
| tokenstatcheapestgoogle/gemini-4-carbon | $2 | $10 | $0.1 | N/A | |
| LiteLLMgoogle/gemini-4-carbon | N/A | $2 | $10 | $0.1 | N/A |
| LiteLLMgemini-4-carbon | N/A | $2 | $10 | $0.1 | N/A |
Scores
Vendor launch report. Google launch table; Gemini results use highest thinking and pass@1 unless otherwise noted. OSWorld: offline subset, partial score, best of three runs; AutomationBench: private set. Cybersecurity results were evaluated without guardrails. AA Index v5.0 is as cited by Google, not a live leaderboard result. See Google’s evaluation methodology for full setups.
Sparse coverage on purpose. Missing a score means we have not linked one yet, not that the model is unranked. Hover or tap the i next to any score for its source. Full boards on /benchmarks.
- AA Intelligence Index v5.0 (Google-reported)59i
- Vals Index v2.2 (%, Google-reported)70.4i
- SWE-sweep (subtask-weighted reward, %, Google-reported)9.6i
- FrontierSWE 2 (%, Google-reported)61.2i
- Terminal-Bench 4.0 (AA, %, Google-reported)65.2i
- Agents’ Last Exam v1.1 (score, %, Google-reported)77.4i
- OSWorld 2.1 (offline partial, best of 3, %, Google-reported)81.5i
- AutomationBench (Zapier private set, %, Google-reported)53.9i
- Vending-Bench 2 (bank balance, USD, Google-reported)16484i
- Chartography (no tools, %, Google-reported)72.5i
- Blueprint-Bench 2 (similarity, %, Google-reported)58.6i
- GraphWalks (256k–1M, BFS F1, %, Google-reported)96.2i
- RiemannBench (%, Google-reported)80i
- Terminal-Bench-Science 0.1 (AA, %, Google-reported)64.8i
- CWE-bench v1 (%, Google-reported)71i
- ExploitGym (%, Google-reported)54.8i
- SEC-bench Pro (JS engines, pass@1, %, Google-reported)88.2i
Arena rows are community-vote Elo from Arena AI, Artificial Analysis indices arrive via OpenRouter, and launch-eval figures are collected from vendor announcements, analyst writeups, and community benchmarks. Full boards and credits on /benchmarks.
See what this model costs you
The app reads the sessions your tools already wrote and prices this model's tokens at the published API rate. Free on your computer.