Gemini 4 Carbon Launches: API Price, MKA Results, and Agent Limits

Google first announced Gemini 4 Carbon on October 28, 2026. On November 10, 2026, it expanded availability to paying API customers and began rolling the model out through selected consumer products. The announcement is unusually relevant to AI-agent developers because Google positioned Carbon not as a blanket replacement for its faster models, but as a higher-reasoning option for longer, more difficult tasks.

The practical questions are less dramatic than a model leaderboard: Who can use it now, what does it cost, and how much evidence supports the agent-performance claims?

What was announced on November 10?

In its launch announcement, Google described Carbon as an extension of Gemini 4 Argon, aimed at complex debugging, multi-step software engineering, document-heavy knowledge work, scientific reasoning and cybersecurity.

The published availability and pricing are specific:

• Developer API: Google’s Gemini API through Google AI Studio, available to paying developers.
• Enterprise: Gemini for business customers.
• Consumer and agent products: A phased rollout for Google AI Ultra subscribers in the Gemini app, Google Antigravity and AI Mode in Search. A rollout is not proof that every eligible account already has access.
• Introductory API price: $2 per million input tokens and $10 per million output tokens, the same announced introductory rates as Gemini 4 Argon. Google says the introductory period ends on November 30, 2026; afterward, $4 per million input tokens and $20 per million output tokens apply to the public Carbon model. These are provider-stated rates, not independently observed billing.
• Context and modality: The official model card lists a combined context window of up to one million tokens, accepting text, images, audio and video as inputs. It describes text output; the context limit is not a promise of consistent retrieval accuracy across all million tokens.

Google is also previewing /swarm in Antigravity, which is designed to coordinate several AI agents across workspaces. Preview matters here: an announced preview should not be presented as a fully tested, generally available production feature.

A more informative benchmark than a single score

The same day, Google and Artificial Analysis introduced Multimodal Knowledge Agent (MKA), a benchmark intended to test whether an agent can use visually dense professional files to produce a useful deliverable.

This is different from asking a chatbot a fact question. Tasks involve source files such as diagrams, charts, scans, tables and documents, followed by outputs that may need to be rendered or checked.

According to the benchmark announcement and Artificial Analysis methodology and live results:

• The evaluation uses 200 private tasks across 10 industry areas; 10 separate public examples help readers inspect task formats.
• Models use a shared Stirrup agent harness, with a stated upper limit of 250 turns per task. This controls part of the workflow, not every possible provider-specific difference.
• A separate model-based evaluator checks deliverables against weighted criteria. Artificial Analysis describes its implementation, and Google reports 96% agreement with domain experts in a comparison exercise. Agreement on that check does not guarantee that every future judgment is correct.
• The published MKA Climb weighted satisfaction rate lists Claude Fable 5.2 at 57% and Gemini 4 Carbon at 54%, in the configurations shown on the leaderboard.

Those figures describe this particular rubric and model configuration, not a 54% success probability for every real-world office task. The three-percentage-point difference is not itself a statistical significance test.

Google highlights that Carbon completed MKA tasks at more than six times lower average cost per task than the highest-scoring model. That is a benchmark-specific comparison, not a general statement that Carbon is six times cheaper for every API workload. The provider also claims strong results on several other agent benchmarks; users should inspect each evaluation’s methodology and cost assumptions rather than merging unrelated rankings.

Is the MKA evaluation independent?

There are two separate facts to hold together.

Google co-developed and announced the benchmark with Artificial Analysis. Artificial Analysis runs its evaluation and publishes its own methodology and results. That provides more visibility than an uninspectable vendor-created score, but it does not make the task design wholly unrelated to the company whose model is being evaluated.

The 200 scored tasks are not public, which helps limit benchmark contamination but prevents outsiders from reproducing the entire hidden test suite. The 10 sample tasks are useful for inspecting the evaluation format, not for inferring that they alone validate the headline scores.

The evaluator is itself an AI model. Rubric agreement and separate grading reduce certain risks, but no single model-as-judge metric can establish the correctness of every spreadsheet formula, chart, or engineering deliverable in an actual deployment.

Where Carbon may make economic sense

An agent team deciding whether to call Carbon should test complete tasks, not only model answers. Carbon and Argon share the announced introductory per-token price, but Google says Carbon can use more reasoning tokens on demanding workflows. A fixed token rate does not imply identical costs per completed job.

For a controlled adoption decision, use the same task set and record:

1. The exact model identifier and revision.
2. Successful final deliverables, including factual checks and formatting constraints.
3. Total input and output tokens, retries, tool calls and wall-clock time.
4. Total cost per accepted result, not just the cost per token.
5. Failure handling: whether the agent asks for clarification, loops, calls unsafe tools or returns an unsupported answer.

For coding agents, the last item matters especially. Our earlier MCP test-evidence audit showed why an observed successful test command and proof that the final revision is validated are different claims. A model benchmark does not remove the need for a trustworthy execution and review workflow.

The takeaway

Gemini 4 Carbon is a real November 10 release, with a specified developer API price and a new multimodal agent benchmark. Those facts are more useful than the generic claim that it is simply “the smartest model.”

The substantive unresolved question is whether Carbon’s deeper reasoning produces more accepted outcomes for the total cost and time in an organization’s own tasks. MKA supplies one structured point of comparison, but its task set, grader and configuration have boundaries that should stay visible.

Primary sources

• Google: Gemini 4 Carbon availability and pricing, November 10, 2026
• Google: Original Carbon announcement, October 28, 2026
• Google: Introduction to the MKA benchmark
• Artificial Analysis: MKA methodology and scores
• Google DeepMind: Gemini 4 Carbon model card

Editorial note: This article is an independent analysis of November 10 announcements and publicly accessible benchmark information. Agentic Pulse did not run an independent Gemini 4 Carbon benchmark and has received no compensation or editorial direction from Google or Artificial Analysis.