Skip to content
DrCompsSignal / computer news
Menu

Google makes Gemini 4 Carbon generally available with million-token output

Gemini 4 Carbon is generally available with a unified 1,048,576-token input-and-output budget, continuation tokens, durable caching, and configurable thinking.

Google has made Gemini 4 Carbon generally available through its enterprise agent platform. The model supports a unified 1,048,576-token input-and-output budget and can continue very long responses across requests with a continuation token.

What changed

The generally available release adds configurable low, medium, or high thinking levels, implicit and durable context caching, Office and OpenDocument file inputs, and Provisioned Throughput support. Google also lists a gemini-4-carbon-cyber-permissive variant for verified defensive-security users admitted through its Fairwind program; that variant is not generally open to every customer.

The unusually large combined token budget changes the shape of long-running agent and document workflows. A system can provide a large working context, request a long generated artifact, or divide the budget between the two. Outputs approaching the limit require continuation-token handling, so existing clients should not assume one response object will contain the entire result.

Practical consequence

Teams evaluating Carbon can consolidate some multi-stage analysis, code, or document tasks that previously required manual chunking. They still need application-level tests for continuation, caching, timeouts, retry behavior, latency, and the cost of high-thinking or very long generations. Google lists no retirement date for the initial Carbon version, but channel availability and access terms vary.

Benchmarks and limitations

Google’s model card reports 61.2% on FrontierSWE2, 81.5% on OSWorld 2.1, and 96.2% on GraphWalks. These are Google-published results, not independently reproduced DrComps benchmarks, and the comparison table does not show Carbon leading every competing model on every test. They should be treated as evaluation inputs rather than a universal performance ranking.

The model card gives Carbon a March 2026 knowledge cutoff and warns that it can hallucinate, occasionally respond slowly or time out, and consume more tokens at higher thinking levels. The release notes and model card do not provide a complete workload-level price comparison, so the million-token ceiling should not be interpreted as evidence that using the full budget is economical for every application.