Gemini 4 Carbon vs GPT-6 Astra
| Gemini 4 Carbon | GPT-6 Astra | |
|---|---|---|
| Released | ||
| Input / 1M tokens | $2 | $10 |
| Cached input / 1M tokens | $0.10 | $1.00 |
| Output / 1M tokens | $10 | $50 |
| Context window | No data | 1,050,000 |
| Output limit | No data | 128,000 |
Both labs report
Terminal-Bench 4.0
Gemini 4 Carbon65.2%Run by Artificial Analysis · mini-swe-agent harness; high thinking; average of 3 runs
Terminal-Bench Science 0.1
Gemini 4 Carbon64.8%Run by Artificial Analysis · mini-swe-agent harness; high thinking; average of 3 runs
AutomationBench 1.0.6Private set
GPT-6 Astra41.4%Run by Zapier · highest score at any effort; set named in the GPT-6.1 Sol post, which prints the same figure
OSWorld 2.1Offline set, partial score
Gemini 4 Carbon81.5%82-task offline subset, no internet access; highest thinking; max of 3 runs, one attempt per run
Reported for one model
OSWorld 2.0 (v2026.08.08)Offline set, partial score
ChartographyNo tools
GPT-6 AstraNo data
SEC-bench Pro 260505
Gemini 4 Carbon88.2%183-task JavaScript engine split (May 2026); pass@1, mean of 3 attempts; Antigravity harness; LLM judge for crashes in both builds
Agents' Last Exam v1Measure not stated
Agents' Last Exam v1.1Score
Gemini 4 Carbon77.4%122 tasks; ALE-Claw harness; 5-hour window; October 17, 2026 task files; safety filters on