The Extended Brief
Why is Gemini 4 Carbon taking so long?

Brief by The AI News AI newsroom · Nov 18, 2026, 6:01 AM EST edition
Original reporting by Hacker News · published Nov 17, 2026, 11:14 PM EST
Gemini 4 Carbon's near-hour runtimes on real coding tasks — versus minutes for rivals — make speed, not capability, the deciding factor for adopting it.
Key points
- On the authors' private eval, Gemini 4 Carbon took close to an hour per bug-fixing task where Opus and Astra took minutes. source ↗
- Theo (t3.gg) measured Gemini 4 Carbon low at 22m21s versus 9m33s for Opus 5.5 high on November 11, 2026. source ↗
- Each eval task gives the agent a pre-fix codebase and a symptom-only bug report, requiring diagnosis and a minimal fix. source ↗
- The authors traced Carbon's runs to separate inference speed — queueing, latency, generation — from the model simply doing more work. source ↗
- Cost estimates used subscription subsidization rates of 40× for Opus, 13× for Astra, 3× for Carbon, and 17× for Hydrogen. source ↗
The data
22m21s
Gemini 4 Carbon (low) on a real-world task
Opus 5.5 (high) finished the comparable run in 9m33s.
Rates applied to $200-plan subscription pricing when estimating cost per trial.
Numbers from the original article, machine-verified against its text
Practical applications
- Before adopting Gemini 4 Carbon for agentic coding, run it on your own multi-step tasks and measure wall-clock time per task, not per-token latency.
- When profiling a slow agent, log per-step inference time separately from step count so you can distinguish provider slowness from the model doing extra work.
- Recompute cost-per-task comparisons at unsubsidized API list prices, since subscription subsidization varies widely (40× for Opus versus 3× for Carbon).
Context
Agentic coding evals hand a model a codebase and a bug report, then let it act autonomously over many steps to diagnose and fix the bug. Total task time is the product of how long each model call takes (inference speed) and how many steps the model takes. A model can therefore be slow either because the provider serves it slowly or because it does more work per task.
What to watch
- The authors' full breakdown of how much of Carbon's slowness is inference versus extra work, which this excerpt cuts off before reporting.
- A Google response or model update addressing Carbon's speed would show whether the bottleneck is fixable.
Related briefs
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- Anthropic's IPO roadshow video
- In 27,280 coding agent runs, 64% of inference spend went to runs that failed
- SenseNova-Loop-9B vs its own Preview: AutomationBench 14.33 to 36.67, GAIA 37.86 to 69.90 (9B Looped Transformer)
Editorial score 3.8 / 5 · significance 3.5 · novelty 4.0 · edge 4.0 · perspective 4.0
Desks: Engineering · Business
Topics: Benchmarks & evals · Inference · Pricing & economics
Evidence basis: Reviewed from the article's full text
This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.