Gemini 4 Carbon:
Google Takes the Lead
Google has unveiled Gemini 4 Carbon, its new frontier AI model, and it has immediately claimed the top spot on the updated Artificial Analysis Intelligence Index. The headline is striking. The details are even more interesting: stronger terminal coding, multi-day engineering work, and aggressive cybersecurity benchmarks. But a leaderboard win is not the same as universal superiority — and the model is not broadly available yet.
The announcement in context
On October 28, 2026, Google announced Gemini 4 Carbon just two weeks after making Gemini 4 Argon broadly available in the Gemini app. Google positions Carbon for difficult, sustained engineering and security workflows, rather than short conversational prompts. Its teams report running the model for days while debugging and patching complex codebases. That is an internal observation, not an independently verified uptime or autonomy guarantee.
According to Google's launch announcement, Carbon ranks first on Artificial Analysis Intelligence Index v5.0 at a rounded score of 59. This independent index combines multiple tests across coding, science, knowledge work and reasoning. It is a useful snapshot of model capability, but it should not be mistaken for a single measurement of general intelligence.
01 / Where Carbon stands against GPT and Claude
At launch, Carbon moves ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on the Artificial Analysis Intelligence Index v5.0. The lead is real in this composite ranking — but it is narrow at the top. Rounded and one-decimal scores below are displayed for context; models are evaluated at their respective published reasoning settings.
Source: Artificial Analysis Carbon vs Argon (Carbon 59, Argon approximately 50), and the contemporaneous v5 model score comparison (other models). Exact precision and settings may differ across presentations. Higher is better. These are not percentages.
A small composite-index lead. It does not prove Carbon wins every coding or enterprise task.
A substantial leap for Google's own family on the same updated index.
How to interpret this: Opus remains a formidable code-focused model, Astra can be strong on difficult reasoning, and Sol offers a very different price–capability balance. The model and the agent harness both matter when testing real repositories.
02 / Coding and autonomous engineering: a major upgrade
Independent Artificial Analysis measurements show Carbon making clear gains over Argon on terminal and agentic coding tasks. The most revealing result may be that some tasks improved dramatically while others did not: Carbon's AutomationBench-AA result is unchanged at 78%.
| Independent benchmark | Carbon (High) | Argon (High) | Interpretation |
|---|---|---|---|
| Terminal-Bench 4.0 | 65% | 57% | Terminal-based coding and tool use |
| FrontierCode 2.0 | 41% | 26% | Harder software engineering tasks |
| AutomationBench-AA | 78% | 78% | No score increase on this measure |
| Terminal-Bench-Science 0.1 | 65% | 53% | Scientific command-line tasks |
| HLE-Diamond | 69% | 57% | Hard knowledge and reasoning questions |
| AA-Omniscience v1.1 | 54 | 46 | Factual accuracy evaluation score |
Google also reports that Carbon rose by 89 Elo points to reach #4 on WebDev Arena, and that engineers have observed multi-day work on subtle bugs in large codebases. Neither claim means an API call itself runs continuously for days; persistent work requires an agent framework, tools, storage, retries, and permission controls.
The caveat: a high Terminal-Bench score is not a guaranteed result in Claude Code, Codex, Antigravity or OpenCode. Those systems differ in context management, tool handling and workflows. Until Carbon receives wider access, independent head-to-head testing in the same agent environment is limited.
03 / Cybersecurity is Carbon's biggest story
Google's headline security numbers span the whole vulnerability lifecycle — finding, assessing and fixing software weaknesses. These figures are reported by Google from named evaluations and, as Google notes, were run without guardrails. They should not be treated as independently confirmed success rates for every real-world repository.
Discovering real software vulnerabilities in the benchmark.
Checking whether a flaw is actually exploitable.
Producing changes that fix the benchmark's underlying issue.
Google-reported uplift in successful secure code generation versus Argon.
Safety improvements — promising, but scoped
Google says Carbon recorded 0.0% successful prompt-injection attacks in the evaluated Gray Swan Indirect Prompt Injection benchmark, versus 0.7% for Argon. This means no attacks in that test succeeded. It does not mean Carbon is immune to all prompt injection in the wild.
Google also reports an approximately 80% decrease in high-severity misalignment alerts from its own live coding environments, and around 2× fewer unsupported claims in internal testing. These improvements are relevant to real engineering agents, but the baselines, test population and deployment context matter.
04 / Pricing, context and release: the catch
Carbon is not a public API offering at the time of this announcement. Google is initially granting access to a subset of cybersecurity defenders and subsequently selected trusted partners through the Fairwind Program. Google says wider developer, business and consumer access will come later, without a firm public date.
| Gemini 4 Carbon | Introductory rate | Later standard rate |
|---|---|---|
| Input / 1M tokens | $2.00 | $4.00 |
| Output / 1M tokens | $10.00 | $20.00 |
| Cached input / 1M tokens | $0.10 equivalent* | $0.20 equivalent* |
*Calculated from Google's stated 95% cached-input discount. These are announced prices, not a public checkout offer today. There may be other billing conditions. Source: Google launch post and pricing footnote.
Artificial Analysis lists a 1 million token context window and roughly 100 output tokens per second for Carbon (High) in its testing. It also reports a roughly 97-second time to first token on its measured setup — a reminder that output rate and responsiveness are different metrics. Its benchmark cost figures are calculated using the later standard prices, not the introductory deal.
For now, claims about using Carbon immediately in OpenCode or other API clients should be treated skeptically unless a particular provider can demonstrate genuine model access.
05 / What Carbon changes in the frontier race
Google has reclaimed the top spot on a prominent independent intelligence index. That matters, especially given the short interval since Argon. More important than the symbolic #1 is the pattern of improvements: coding agents, scientific terminal tasks and security work are becoming increasingly central to frontier-model releases.
| Model | Why it matters | Trade-off |
|---|---|---|
| Gemini 4 Carbon | Index leader; strong terminal and security results; ambitious long-horizon positioning. | Restricted access; some performance claims are internal. |
| Claude Opus 5.5 | Near the top of the index; an important reference for complex coding and agents. | Model performance depends heavily on the agent implementation and cost. |
| GPT-6 Astra | Frontier reasoning and challenging engineering. | Higher published per-token price than lower-cost alternatives. |
| GPT-6.1 Sol | Competitive flagship-level capabilities at lower token rates. | Does not top the composite index. |
| Gemini 4 Argon | Existing publicly available Gemini 4 option. | Noticeably behind Carbon on the new index. |
Our assessment: Carbon's launch is a major win for Google, not the end of competition. The gap between Carbon and Opus is thin, the coding landscape is task-dependent, and the most attractive version of Carbon — the one ordinary developers can actually test against the same repository — has not arrived yet.
When it does, the correct comparison will be simple: same tasks, same repository, same tools, repeated runs, and total cost per successful fix. Until then, Carbon has earned the headline, but the real-world verdict remains open.
06 / Sources and methodology
Updated October 28, 2026. All links lead to original announcements, independent benchmark pages, or contemporaneous reporting. The labels “independent”, “Google-reported” and “internal” indicate different evidence quality.
- Google — Introducing Gemini 4 Carbon (Oct 28, 2026). Launch, cybersecurity figures, controlled rollout, internal reliability and prices.
- Artificial Analysis — Carbon (High) vs Argon (High). Independent v5 intelligence, coding and scientific evaluations; speed, context and token usage.
- Artificial Analysis v5.0 comparison, Oct 28. Contemporary compilation of rival model scores and pricing at published reasoning settings.
- Reuters — Google launches Gemini 4 Carbon (Oct 28, 2026). Independent reporting on the announcement and initial access restrictions.