What Google shipped

Google DeepMind opened Gemini 4 Carbon to paying API customers on 10 November 2026. Koray Kavukcuoglu, a senior vice president at Google DeepMind, wrote the announcement. It calls Carbon Google's strongest model and aims it at long, hard jobs such as multi-step coding and debugging, full-stack builds and open-ended scientific research. Cybersecurity defenders got the model two weeks earlier through Google's Fairwind Program. That's why Vals AI lists 28 October as its release date.

Google says plainly that Carbon uses more tokens than Gemini 4 Argon. It still recommends Argon for most everyday engineering and knowledge work. We think that's the right starting point for any team deciding whether to switch.

Who can use it today

Builders can reach Carbon through Google AI Studio, Android Studio and the Gemini Enterprise Agent Platform on Google Cloud, all on paid tiers. There's no free-tier access. Google AI Ultra subscribers are getting it in the Gemini app, in AI Mode in Search and in Antigravity, Google's agent development environment. Antigravity also has a preview feature called /swarm that coordinates teams of agents. Scientists can apply for six months of free access to Argon and Carbon through Antigravity's highest usage tier, subject to eligibility and capacity.

There are two model IDs. gemini-4-carbon is the general model. gemini-4-carbon-cyber-permissive is only for approved Fairwind security work. It defaults to the HIGH thinking level and doesn't support computer use.

What changes in your code

Google Cloud's migration guide is written for teams moving from gemini-3.1-pro-preview, and several old habits now break.

  • You set reasoning depth with thinking_level: LOW, MEDIUM (the default) or HIGH. It's uppercase on GenerateContent and lowercase on the Interactions API. Sending the older thinking_budget or the MINIMAL level returns an HTTP 400.
  • Temperature, top_p and top_k are fixed, and the API quietly ignores custom values. Anything you tuned through sampling now needs prompt changes or a response_schema.
  • The API rejects frequency and presence penalties, and any candidate count above one.
  • Explicit context caching is gone. Implicit caching applies to shared prompt prefixes of 4,096 tokens or more. Put system prompts, tool definitions and reference documents at the top, and keep them identical between calls.
  • Chat history can't end on a model turn, so you can't prefill the model's reply.
  • Each function response has to match the id, name and order of the call before it, and Google prefers tools to run one at a time.

Input and output share a single window of 1,048,576 tokens. The output count includes thinking and tool calls. If you leave max_output_tokens unset, you get 65,536.

Long outputs arrive in segments

This is the most unusual change. A single foreground segment tops out at 65,536 output tokens when streaming, or 32,768 without streaming. To go further, Carbon uses "long decodes", which join segments together until the job finishes or the shared token window runs out.

Teams have three options:

  • Automatic SDK continuation: Google recommends this. It needs minimum SDK versions, such as Python google-genai 2.34.0 for the Interactions API.
  • A server-managed background loop.
  • Your own loop: you pass a continuation_token yourself.

The Interactions API reports requires_continuation when a segment hits its cap, incomplete when the overall budget is used up, and completed when the job is done. Stored interactions stay on Google's servers for seven days.

Some rules apply:

  • Long decoding only works on the v1beta1 endpoint.
  • Changing the model, system instruction or tool list between segments returns a 400.
  • Google asks you to keep the thinking level and built-in tools the same across segments too, even though changing them isn't rejected.

Google suggests long decodes for hard maths, formal proofs and the toughest coding tasks. It advises against them for chat, quick answers and anything where latency matters.

What it costs

Carbon has an introductory price that matches Argon: US$2 per million input tokens and US$10 per million output tokens. Thinking tokens are billed as output. AI Studio's pricing page shows both models doubling from 1 December 2026, to US$4 for input and US$20 for output. Google's announcement gives 30 November as the change date. Batch and Flex runs cost half the Standard rate, and Priority costs 1.8 times Standard. Cached input costs US$0.10 per million tokens now and will rise to US$0.20.

Google bills every token it processes. That includes thinking and, when grounding is switched on, internal tool instructions and function declarations. Grounding with Google Search is charged per query on top of that. Token counts from all segments are added together, so one long decode can produce a large bill from a single request.

How it scores outside Google

Google's own results are strong. It reports first place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, and first of 23 model configurations on Balyasny's financial benchmark. Independent testers support part of that picture.

Artificial Analysis

  • Carbon tops the Intelligence Index v5 with a score of 59 and leads four of its nine evaluations.
  • It trails Claude Opus 5.5 on the firm's coding agent index, 59 to 62.
  • It sits behind three rivals on the AA-Briefcase knowledge work test.
  • Its hallucination rate rose from Argon's 4% to 11%. That's still the second lowest of any model the firm has measured, after Argon.
  • Running the index cost US$7.76 per task. That's close to Argon's US$7.68 and about 23% below Opus 5.5.

Vals AI

  • Carbon ranks third of 36 on the Vals Index, at 70.45%.
  • It's first on Vals' finance index, Legal Research Bench and Tax Agent Bench.
  • It's ninth on Vibe Code Bench.
  • At high reasoning effort, a Vals Index run averaged close to 48 minutes and US$10.51 per test.

Limits and catches

  • Carbon runs in global, US and EU regions only. There's no Australian or New Zealand data region, which matters for teams with data residency commitments.
  • RAG Engine, tuning, the Live API and fixed quota aren't supported.
  • Many headline results come from Google itself, and some come from internal or pre-release runs. Google also says resistance to prompt injection still needs work.
  • Carbon uses a lot of tokens, so a modest per-token price can still mean expensive tasks.

What we'd do from Auckland

We'd run our hardest real tasks on both Argon and Carbon before 1 December, while prices are still at the introductory rate. Then we'd judge each model by what a finished job costs, because per-token rates hide how much Carbon spends on thinking.

We'd also:

  • budget in New Zealand dollars with room for exchange rate swings;
  • check whether US or EU processing is acceptable for our data;
  • reorder prompts so static content comes first and benefits from implicit caching.

For routine chat, summarising and day-to-day coding, Argon costs the same and is the sensible default. We'd save Carbon for long, hard jobs where the extra tokens pay off.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member