Google opened Gemini 4 Carbon to paid API developers today, and the timing matters: introductory pricing at $2 per million input tokens — half the cost of Claude Opus 5.5 — expires November 30. If you’ve been waiting to evaluate Carbon, you have twenty days to do it cheaply.
The Pricing Case #
At launch, Carbon is priced at $2 per million input tokens and $10 per million output tokens. Claude Opus 5.5 runs $4/$20. Cache input tokens cost $0.10/M on Carbon versus $0.20/M for Opus 5.5. After November 30, Carbon’s standard rates jump to $4/$20 — identical to Opus 5.5.
The token price comparison understates Carbon’s cost advantage. Because Carbon tends to complete tasks with fewer tokens, Artificial Analysis puts real-world cost per task at $7.76 for Carbon versus $10.04 for Opus 5.5 — a 23% gap that persists even after the introductory period ends. For teams running high-volume agent pipelines, that delta adds up fast.
| Model | Input (intro) | Output (intro) | Input (standard) | Output (standard) | Cache Input |
|---|---|---|---|---|---|
| Gemini 4 Carbon | $2/M | $10/M | $4/M | $20/M | $0.10/M | | Claude Opus 5.5 | $4/M | $20/M | $4/M | $20/M | $0.20/M |
Benchmark Reality: Where Carbon Wins and Where It Doesn’t #
Carbon sits at 59 on the Artificial Analysis Intelligence Index v5; Claude Opus 5.5 is at 58. That one-point gap is narrow enough to be noise. The interesting divergences are elsewhere.
Carbon’s hallucination rate on SimpleQA-based testing is 11%, compared to 50% for both Opus 5.5 and GPT-6 Astra. That is not a rounding error — it is a structural difference in factual reliability that matters enormously for agents that need to cite, retrieve, and reason from real documents without fabricating details.
On business automation (AutomationBench), Carbon leads at 53.9%. On CursorBench 4.0, Carbon scores 60.4% to Opus 5.5’s 57.8%. On cybersecurity benchmarks — SEC-bench Pro, ExploitGym, CWE-bench v1 — Carbon leads all three.
Where Opus 5.5 wins: computer use. OSWorld 2.1 scores are 87.2% for Opus 5.5 versus 81.5% for Carbon. And on the Artificial Analysis coding agent index specifically, Opus 5.5 leads by three points. The two most prominent coding benchmarks give different verdicts — CursorBench favors Carbon, Artificial Analysis coding agents favor Opus 5.5. Neither result is wrong; they measure different things.
What Carbon Is Actually Built For #
Google is not trying to sell Carbon as an interactive code-completion model. The pitch is sustained agentic work: multi-step pipelines that run for hours or days, long-horizon reasoning across a 1-million-token context, and security engineering that requires precise exploit analysis and patch generation. VentureBeat reports Google says internal teams have run Carbon loops that “persist through dead ends” for days.
That framing clarifies when Carbon makes sense. Multi-package migrations, cross-subsystem root-cause investigations, documentation-heavy agent workflows that benefit from aggressive context caching — these are Carbon’s native use cases. If you’re building a user-facing coding assistant for quick completions, the Opus 5.5 lead on interactive coding benchmarks is more relevant.
Who Should Evaluate Now #
The developer community’s reaction has been cautious. Hacker News consensus: Gemini’s historical track record on coding means trust has to be earned, not benchmarked. The hallucination gap is compelling; the one-point overall index lead is not compelling enough to trigger a migration.
The practical question is simpler: what does your workload actually look like? Long-running agents on static codebases with large context requirements — Carbon is worth testing now, while the cost of experimentation is half price. Interactive coding sessions and computer-use workflows — Opus 5.5 isn’t going anywhere.
Access is live through the paid Gemini API in Google AI Studio and Android Studio. The model ID is gemini-4-carbon. Cybersecurity teams can apply separately through the Fairwind Program. The November 30 pricing deadline isn’t a soft suggestion — standard rates are twice the current price.
Source: Google AI Blog, Artificial Analysis, DEV Community benchmark analysis.