If you are adding Gemini 4 Carbon to a long-document workflow, check how your software carries earlier work into later requests. ARC Prize's results page, dated November 10, reports 63.9% on ARC-AGI-3 with the Standard harness at medium reasoning, against 99.1% with the Provider Adapter at the same reasoning level. This is not a writing test. It is a warning against treating a model name as a complete configuration.
Notes and reasoning state are different things
The benchmark organizer's explanation describes two ways of carrying work forward. The Standard harness lets the model retain notes it chooses to keep. The Provider Adapter preserves opaque reasoning state between requests and compacts longer conversations so that earlier work can be reused. A harness is the software around a model that manages its execution and task continuity.
Before switching clients or adding an intermediary service, our recommendation is to check the integration documentation and ask the supplier specific questions. Does it support preserving Carbon's reasoning state across requests, or does it pass only text history and notes? Does it support long-conversation compaction, and which layer handles that? Ask for configuration documentation or inspectable request records. Comparing outputs can reveal a quality difference, but cannot by itself diagnose the loss of opaque state.
The same page also documents an incomplete evaluation. At the highest reasoning setting, the Standard harness completed only 30 of 55 games. Long gaps between actions caused game-session expirations, so ARC Prize omitted that score. This is a reason to check timeouts and task completion during a trial, not evidence that reasoning state was lost.
Use an official route as a reference, and record the bill
In its availability announcement, Google lists the paid Gemini API in AI Studio as a developer route. Google AI Ultra subscribers can also try Carbon in Antigravity. If you have access, keeping a reference output from an official route is a useful starting point before moving an entire workflow.
The public model's introductory API rates are $2 per million input tokens and $10 per million output tokens. After the introductory period expires on November 30, 2026, the rates become $4 and $20 respectively. Google also says Carbon uses more tokens than Argon for exploration and verification. The token rate alone will not tell you what a completed assignment costs.
Google still positions Argon as the preferred choice for most software engineering and knowledge work, reserving Carbon for demanding tasks such as complex multi-step coding, debugging and open-ended scientific research. Our recommendation follows that distinction: if an ordinary document workflow already meets your requirements, switching to Carbon need not be the priority. A difficult assignment that remains unresolved after repeated revisions is a better candidate for a small comparison.
Compare identical briefs, not impressions
For technical documentation, our recommendation is to keep the source material, writing requirements and reasoning setting fixed. Check factual errors, missing steps and consistency after successive revisions, alongside total spending. Change only the client or connection route at a time so that any difference is easier to investigate.