Reviews · NOVEMBER 10, 2026
Gemini 4 Carbon opens at $2/$10 API rates, with a December increase ahead
Google’s model targets complex coding and debugging. Small teams should test customer-facing features against completed-task costs, not introductory token prices.
Google opened Gemini 4 Carbon to paying developers and businesses November 10, following its October 28 unveiling, at introductory API rates of $2/million input tokens and $10/million output tokens.
Google’s announcement, “Gemini 4 Carbon is now available,” positions it for complex, multi-step coding and debugging. Google still recommends Gemini 4 Argon for most software engineering, as Engadget reports.
Access and verification
Access includes Google AI Studio and Gemini for business; Google AI Ultra rollout covers Antigravity, the Gemini app and AI Mode in Search. VentureBeat reports no downloadable weights, self-hosting license or country-by-country schedule. Gemini 4 Carbon for cyber requires Fairwind program membership.
Carbon checks code and rendered interfaces through build-and-verify loops. Antigravity’s /swarm preview gives agents parallel, isolated workspaces and verifier testing; Google describes Carbon coordinating, architecting and auditing alongside lighter models.
SecurityWeek reports five researchers used AlphaProof Code over four weeks to verify libsignal security guarantees, with Carbon agents generating roughly 6.5 million lines of Lean proofs. Verification covered Google’s compiled build, not Signal’s applications; parts excluded side channels.
Benchmarks and limits
Google DeepMind’s Gemini 4 Carbon, Model Card reports October 2026 evaluations. Its Carbon/Argon comparison lists Terminal-Bench 4.0 scores of 65.2% and 57.1%; FrontierSWE 2 results include 61.2%, 65.5% and 62.3%, with GPT-6 Astra among comparators. These are vendor-reported evaluations.
VentureBeat reports Anthropic’s Claude Fable 5.2 at 69.9% on FrontierSWE v2, plus Artificial Analysis’s FrontierCode 2.0 results of 49.1% and 40.6%.
Carbon accepts text, image, audio and video, outputs text, and supports up to 1 million combined input-output tokens. Its cutoff is March 2026. The card warns of hallucinations, occasional slowness or timeouts, and higher token consumption at stronger thinking settings. Extended reasoning and verification consume more thinking tokens than Argon.
Token prices versus accepted changes
Introductory rates run through November 30, 2026. December 1 standard uncached rates will become $4/million input and $20/million output, matching Claude Opus 5.5; Fable 5.2 lists $8/$40. Google’s 95% cached-input discount implies $0.10/million during promotion, $0.20 afterward.
VentureBeat’s Zapier AutomationBench figures:
| Setting | Score | Cost/attempt | Estimated cost/passing task, including failures | |---|---:|---:|---:| | Carbon medium | 52.97% | $1.65 | $3.11 | | Carbon high | 53.88% | $2.12 | $3.94 | | Fable 5.2 max | 51.45% | $4.64 | Not stated |
Artificial Analysis’s index suite instead estimates Carbon medium at $6.59/task with a score of 58, versus Fable medium at $3.80 and 56. Those estimates use standard rates; vendor effort settings aren’t equivalent. Lower token prices don’t establish lower task costs.
"The practical takeaway: compare the cost of an accepted change, including retries and review, rather than the cheapest token."
For an owner-led team, a hypothetical booking, quote or checkout feature provides a bounded pilot. Track time, corrections, extra tokens and retries alongside quality, using the same acceptance criteria across models. The Fable 5.2 testing analysis offers a comparison point. Budget the pilot against December’s scheduled rates, not just November’s promotion.