GPT-6 Astra
Lineage
- Company
- OpenAI
- Family
- GPT-6 Astra
- Assistants that offer it
- ChatGPTpaid plan, Plus, Pro (official page ↗)
Perplexitypaid plan, Max (official page ↗) - API price
- 10 $ per million tokens read, 50 $ written (developers.openai.com/api/docs/pricing, read on 5 November 2026)
Links read on 10/11/2026 on company pages, the press as a fallback; a link given for a family applies only to its latest version.
Rank in each benchmark
- General intelligenceEpoch AI, ECI index169.2166.2 to 172.52nd out of 71
- User preferenceLMArena, Elo score147428th out of 68
- Agent tasksEpoch AI, APEX-Agents64.7 %53.7 to 75.713th out of 52
- Code repairEpoch AI, DeepSWE74.1 %1st out of 26
- Coding tasksEpoch AI, CursorBenchnot measured
Each segment of the bar is worth a tenth of the ranking; the 1st groups the best ranked. Rank among models released in 2026 and measured by the benchmark. Sources: Epoch AI, "Capabilities & benchmarking" (CC BY 4.0); LMArena, text with style control (CC BY 4.0); data as of 10 November 2026.
Effort settings
| Setting | Electricity use of a task | Tasks completed | Tokens per task |
|---|---|---|---|
| DeepSWE, code repair | |||
| low | 29 to 50 Wh | 67.0 % | 10,580 |
| medium | 56 to 96 Wh | 72.8 % | 20,362 |
| high | 73 to 130 Wh | 73.2 % | 26,506 |
| xhigh | 82 to 140 Wh | 74.1 % | 29,557 |
| max | 170 to 290 Wh | 73.2 % | 61,149 |
Estimation methodology
The effort setting says how much the model reasons before answering: the higher it is, the more tokens it writes, and the more electricity the task needs. Tokens per task are measured by Epoch AI.
Electricity: Projet Celsius estimate. Tokens of the task × electricity ofwriting 1,000 tokens for the size of the model (AI Impact Database, IA-0011), × the data centre overhead (PUE). The scope covers computing on the servers and cooling; it does not cover chip manufacturing, training, the network or the device. CursorBench counts tokens read and written together: the range runs from "all read" to "all written". DeepSWE counts tokens written: the range starts from writing alone.
Assumed model size: large class, deduced from the output price (developers.openai.com/api/docs/pricing), assumption. If the model were medium-sized, the estimate would be 4 times lower. Benchmark: a quick question to an assistant needs about 0.3 Wh.
GPT-6 Astra among OpenAI's models, general intelligence
Epoch AI ECI index and its 95% interval, by release date; 14 OpenAI models released in 2026 and measured by the ECI index
Dot: the index measured, at the release dateLine: the 95% intervalGPT-6 Astra5 names in the "Table" tab
Source: Epoch AI, "Capabilities & benchmarking", ECI index. CC BY 4.0 licence.
On a coding task, GPT-6 Astra needs 170 to 290 Wh at its most demanding setting, as much as about 570 to 970 quick questions to an assistant (about 0.3 Wh each).
Continue in the ranking
The full ranking →CompanyOpenAI14models measured, of which 10 with an entryIts assistant, its models, its governance and the electricity of its data centres.The entry →AssistantChatGPT62out of 100, 3rd out of 10One of the 2 assistants in the ranking that offer GPT-6 Astra.The entry →RankingAI modelsAll models measured1stof 26, code repairThe ranking where GPT-6 Astra is best placed.The ranking →The hub newsletter
This month's ranking, news on the impact of AI and one figure from the AI Impact Database. One newsletter a month at most, free of charge. Past issues
