Summary:

Google opens Carbon to paying developers and Ultra subscribers. Independent benchmark snapshots show workload-specific tradeoffs, while introductory API prices end November 30.

Google opened Gemini 4 Carbon to paying developers and enterprises on November 10, expanding beyond the cybersecurity cohort announced two weeks earlier. Google AI Ultra subscribers are also getting access across the Gemini app, AI Mode in Search and Antigravity.

The practical change is access: teams can now evaluate Carbon through the paid Gemini API rather than wait for the initial Fairwind program. The public model and Carbon for cyber still have different access routes, so the latter’s security results should not be treated as ordinary API capabilities.

Where Carbon is available

  • Developers: the paid Gemini API in AI Studio and Android Studio. The API model identifier is gemini-4-carbon.
  • Enterprise teams: Google lists Gemini for business, Antigravity and CodeMender among the access surfaces.
  • Consumers: rollout to Google AI Ultra subscribers in the Gemini app and AI Mode. This is not a free-tier announcement.
  • Security defenders: Carbon for cyber remains available through the Fairwind application process.

This follows our earlier coverage of Carbon’s restricted introduction. The November 10 release is the broader-access milestone, not a second first announcement.

Budget for December, not just the launch price

The standard API rate is $2 per million input tokens and $10 per million output tokens through November 30, 2026. Google’s pricing documentation lists $4 and $20 respectively from December 1. Output billing includes thinking tokens; Carbon has no free API tier.

That makes a two-period budget useful. For an illustrative workload of one million input tokens and one million output tokens, the listed standard token charge is $12 during the introductory period and $24 afterward, before other applicable charges. Actual spending depends on usage: a model that reasons longer can consume more billed tokens even when its per-token price is competitive.

Three benchmark snapshots, with different answers

These comparisons come from benchmark owners and an outside evaluator, rather than a single vendor’s headline score. They cover coding, app automation and visual knowledge work. Compare models within each row; the percentages measure different outcomes.

Four-row scorecard: Carbon leads selected CursorBench success and AutomationBench results, Argon has the lowest Cursor cost, and Fable leads MKA weighted success.

  • CursorBench 4.0: Carbon High scores 60.4%, Fable 5.2 Max 60.0%, Opus 5.5 Max 57.8% and Argon High 52.9%. Cursor cautions that small score differences may not be statistically meaningful. Reported costs are $9.22, $32.79, $13.43 and $4.57 per task respectively, applying token prices to measured usage. Carbon’s near-tie with Fable is more interesting here than a categorical claim that it is better at coding.
  • AutomationBench 1.0.6: Zapier’s October 26–27 prerelease report gives Carbon High 53.9%, Argon High 51.3% and Opus 5.5 Max 42.5%. Carbon passed 354 of 657 tasks, with success requiring every final-state assertion to pass. These are dated prerelease configurations, not a fresh test of every November 10 public endpoint. The report does not provide a Fable 5.2 result.
  • MKA Climb: Artificial Analysis reports 54% for Carbon High, 57% for Fable 5.2 and 53% for Opus 5.5. This is weighted rubric satisfaction, not the strict task pass rate used by Zapier. The benchmark was developed by Google DeepMind; these figures use Artificial Analysis’s implementation. Claude configurations use maximum effort where supported, with the evaluator’s default fallback otherwise. No Argon value is included in this snapshot.

The useful conclusion is workload-specific: Carbon is a strong candidate for difficult coding and automation, while this visual knowledge-work sample favors Fable. None of the rows is a universal model ranking. Benchmark cost figures are dated measurements, not a promise of future production spending.

A practical evaluation before switching

Choose a small set of tasks your current model already handles, then compare completion quality, human corrections, latency and total billed tokens. Include one difficult multi-step task and one routine task: the more expensive reasoning path may earn its place on the former without making sense for the latter.

For security work, confirm the access tier first. Google’s cybersecurity evaluation methodology identifies Fairwind access for those results. Keep that separate from the public Carbon route when deciding what to test or procure.

Sources and image credits

Original PromptsZone charts use the sources above. The unaltered official Gemini family logo is used as a labeled fallback rather than a distinct Carbon logo, from Google’s press image library. Source: Google. Google and Gemini are trademarks of Google LLC. PromptsZone is not affiliated with or endorsed by Google.