Gemini 4 Carbon opens at $2/$10—but task cost is the real test

Google opened Gemini 4 Carbon to paying developers and businesses November 10, announcing introductory API rates of $2 per million input tokens and $10 per million output tokens.

VentureBeat reports uncached rates last through November 30, then rise to $4/$20, matching Claude Opus 5.5; Claude Fable 5.2 costs $8/$40. Cheaper tokens don’t establish cheaper completed work.

Access and operating limits

Google lists AI Studio and Android Studio paid API access, plus Gemini for business API/model-picker access. Carbon is rolling out to Google AI Ultra subscribers through Antigravity, the Gemini app, and AI Mode in Search. VentureBeat says country schedules, Vertex AI availability, regions, and support commitments remain unspecified.

Google positions Carbon for long-horizon coding, research, and verification, recommending Gemini 4 Argon for most software engineering and knowledge work. Engadget emphasizes that distinction; the desk’s October 19, 2026 Argon paid-API coverage supplies predecessor context.

Google DeepMind’s model card describes Argon-based Carbon: text/image/audio/video input, text output, and up to 1 million combined input-output tokens. Configurable thinking balances quality, latency, and cost, but higher settings consume more tokens. Hallucinations, slowness, and timeouts remain risks. Its cutoff is March 2026; some domains are less current. Prospect research isn’t exempt.

What the benchmarks establish

Google self-reports 92% on its 375-task agentic honesty benchmark versus Argon’s 70%: Carbon declined missing-data fabrication 98% of the time, but flagged methodological flaws only 68%. Those aren’t guarantees of accurate buying signals.

The Google-authored paper “MKA: Evaluating Multimodal Knowledge Agents on End-to-End Professional Tasks” covers 200 tasks, 10 sectors, and 83 occupations; 88% require standalone files. Authors report 96.2% evaluator/human-expert criterion agreement across a 50-task study.

| MKA measure | Carbon | Claude Fable 5.2 | |---|---:|---:| | Climb-WSR | 54.3% | 56.6% | | Overall WSR | 71.9% | 73.3% | | Average API cost/task | $5.63 | $35.28 | | Average tokens/task | 7.05 million | 40.48 million |

The authors found the top three systems statistically indistinguishable under paired task-level analysis.

VentureBeat reports Artificial Analysis estimates using standard rates: Carbon medium costs $6.59/task at index 58, versus Fable medium’s $3.80 at 56. Rankings don’t transfer automatically between tests.

Price accepted work before December

VentureBeat reports these Zapier AutomationBench results:

| Setting | Score | Cost/attempt | Cost/passing task | |---|---:|---:|---:| | 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 specified |

"Pilot takeaway: compare the cost of an accepted result, including retries and human review—not merely the cost of generating an answer."

That’s the operator test, not a promise from token pricing. Small-team owners should compare one low-risk prospect-research, proposal-preparation, or customer-response workflow against their current process. Track accepted completions, token use, retries, and staff review. Reprice usage at December’s rates before deciding, ahead of December 1, whether Carbon merits further testing.

Sources

• https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
• https://deepmind.google/models/model-cards/gemini-4-carbon/
• https://storage.googleapis.com/deepmind-media/papers/mka_tech_paper.pdf
• https://venturebeat.com/technology/google-gemini-4-carbon-is-now-publicly-available-and-it-undercuts-claude-fable-5-2-and-opus-5-5-on-api-price
• https://www.engadget.com/2311855/gemini-4-carbon-now-available-worldwide/