Gemini 4 Carbon Tops AI Index at 59 Points, Then Vaporizes: No Release Date, 11% Hallucinations
A nine-point index gain at unchanged per-token prices sets a new cost-performance ceiling that competitors must match. But with no public access, the practical impact for developers is zero until the rollout opens.
Gemini 4 Carbon scored 59 on Artificial Analysis' intelligence index v5, taking first place and four of nine sub-tests. The model's coding performance with its Antigravity CLI reached 59 points, just three behind Claude Opus 5.5's 62, the closest Google has come on that leaderboard. AA-Omniscience results show accuracy climbing to 58% while hallucination rate jumps from 4% to 11%, but rivals like Opus 5.5 and GPT-6 Astra both sit at 50%.
Per-task cost lands at $7.76, 23% below Opus and 24% above GPT-6 Astra. Carbon outputs 123k tokens per task versus Opus's 173k, a 29% reduction. Pricing stays at $4/$20 per million input/output tokens, identical to Argon, but the index score jumps nine points. Cached input costs $0.2 per million, which can halve the bill for cache-heavy workloads.
The announcement drew sharp criticism on Hacker News for its closed availability and timing. Users reported low quotas and hallucination issues in the web interface, though benchmark scores come from the API model. Until Carbon opens, the practical move is to benchmark your own workloads and measure how cache hits affect costs.
The hallucination rate increase from 4% to 11% is a deliberate tradeoff for a more assertive model, and the benchmark shows rivals hallucinate far more, so the absolute risk remains low.
The close coding score depends heavily on the companion Antigravity CLI, meaning model and tooling are now evaluated together; a weak CLI could erase the gap.
The synchronized release timing with internal reviews and earnings suggests the announcement serves a non-developer audience, which is a trust issue that could hurt Google's adoption.
The price anchor staying unchanged while the index jumps nine points means the cost per unit of intelligence drops sharply, making it harder for competitors to justify higher prices.