What happened
Google describes Gemini 4 Carbon, model ID gemini-4-carbon, as its strongest reasoning model, built for sustained agentic workflows in software engineering, multimodal knowledge work, scientific reasoning and cybersecurity defense. The documentation says it spends more thinking tokens on hard problems to explore hypotheses, verify work and check edge cases. The model is available on the paid tier of the Gemini API and in Google AI Studio; it is not available on the Free tier.
Carbon has a unified 1,048,576-token context window that includes input, thinking and output. A response can exceed 64,000 tokens, up to one million tokens minus the input. Standard synchronous responses stop at 65,536 tokens. For longer output, Google says to set background execution to true and raise max_output_tokens above 65,536. The server continues generation in 64K-token hops; the client should collect each model_output step. Google cautions that background execution without an explicitly larger output limit still stops after 64K tokens.
The default thinking level is medium; developers can set low, medium or high. Google frames these as a quality, latency and token-use trade-off, and recommends evaluating all three levels on a workload before selecting the lowest one that meets its quality bar. Carbon shares its context window, built-in tools and per-token pricing with Gemini 4 Argon, but spends more thinking tokens per turn and has higher latency. Google positions Argon for faster interactive coding and routine feature work, and Carbon for harder debugging, architectural design, dense multimodal synthesis, scientific proofs and security audits.
Introductory pricing runs through November 30, 2026: $2 per million input tokens and $10 per million output tokens. From December 1, input costs $4 per million and output costs $20. Output pricing includes thinking tokens. Google says implicit caching is enabled by default for prompts of at least 4,096 tokens and discounts cached input by 95%, from $0.10 to $0.20 per million tokens after the introductory period. Tokens processed for internal tool instructions and function declarations are billed, and Google Search grounding is billed per query. Explicit context caching is unsupported. Sampling parameters temperature, top_p and top_k are ignored.
What the archive adds
New Runtime’s earlier coverage on Gemini 4 Carbon’s general availability (November 11, 2026) described API access and configuration in Agent Studio. This documentation adds operational detail: long-decode mechanics, migration constraints, thinking levels and token prices. It is a new fact on a known story, not a restatement.
Trend signal
This release supports the shift toward capacity-linked API pricing described in Compute-backed AI contracts: Carbon’s effective cost depends not just on token rates but also on thinking effort, caching and grounding. The higher price after November 30 makes workload-level evaluation important. This does not establish that task-based pricing is replacing token billing; Google still bills tokens, including internal reasoning tokens.
Why now
Google’s introductory rates expire November 30, 2026, with higher input and output rates beginning December 1. That gives teams a concrete date to compare expected workload costs under the introductory and subsequent prices.
Why it matters
A million-token context and long decode can make Carbon a candidate for tasks that need extensive source material or unusually long outputs. But context capacity is not the same as economical use: every processed token, including thinking tokens, counts toward the bill, and Carbon trades additional reasoning for latency and token spend. Teams should compare it with Argon and cheaper alternatives on representative tasks, measuring answer quality, turnaround time and total billed tokens rather than choosing by context-window size alone.
The migration details also affect implementation. Existing Carbon clients cannot rely on explicit caching or sampling parameters, and applications with tight output caps or request timeouts may truncate work or fail on long reasoning passes. Background execution, appropriate output limits and server-side conversation state are part of the operating model for sustained tasks, not merely optional tuning.
Limits and uncertainty
Google characterizes Carbon’s capabilities and relative strengths, but the documentation does not provide independent benchmark results, workload-specific latency figures or measured accuracy gains. Actual token use and cost will depend on prompts, thinking level, tool use and output length. The stated cache discount applies to cached input on cache hits; it does not mean all long prompts receive that rate. The documentation also gives no guarantee that a particular task will fit within the context or output limits, or that a higher thinking level will improve its result. Teams should validate performance and billing against their own tasks.
Skill to build now
Build a small evaluation set from representative hard tasks and run each at low, medium and high thinking levels. Record success against your existing quality checks, latency and billed input and output tokens; include cache hits and any grounding queries. Choose the lowest-cost setting that meets your quality bar, and test background execution and timeout behavior if outputs could exceed 65,536 tokens.
Reader takeaway
Gemini 4 Carbon pairs a million-token context with deeper reasoning and long decode, but those capabilities carry explicit cost and latency trade-offs. Before adopting it, benchmark your own tasks at each thinking level and budget for the December 1 price increase.