Artificial Analysis: GPT-6.2 Astra did not improve in Coding Agent Index
In Artificial Analysis’s measurements, GPT-6.2 Astra rose to second place behind Claude Fable 5.2. In the Coding Agent Index, however, Astra stayed level with its predecessor, and GPT-6.2 Sol’s score fell by three points.
OpenAI released two new models on 17 November. GPT-6.2 Astra is intended for demanding professional work and GPT-6.2 Sol for everyday use. Artificial Analysis, which compares models, published its own results the same day.
Artificial Analysis’s Intelligence Index v5 combines nine tests covering knowledge work, coding, science and general knowledge. Astra scored five points more than GPT-6 Astra on the index and overtook Google’s Gemini 4 Carbon (60) and Anthropic’s Claude Opus 5.5 (59). In Artificial Analysis’s chart, Astra has 62 points and the leader, Claude Fable 5.2, has 65, The Decoder reported.
Most of the Pareto frontier for cost per task consists of OpenAI models, Artificial Analysis said. The frontier contains the models for which no other model is both cheaper and better. Sol scored 57 points and matches GPT-6 Astra at about 15 per cent of GPT-6 Astra’s cost per task.
A strong overall ranking does not, however, show how a model performs in a coding agent. Artificial Analysis measures this with its Coding Agent Index. The test uses OpenAI’s Codex harness, the program that controls the agent: it gives the model its tools and runs its steps. The models run at maximum reasoning effort.
On this index, Astra stayed level with GPT-6 Astra, and Sol lost three points compared with GPT-6.1 Sol. Sol’s drop comes from the SWE-Atlas-QnA-AA test, where its score fell by 14 points. Sol omits key facts from its answers more often than its predecessor.
The index score does not cover all agentic coding, however. An image in Artificial Analysis’s post on X shows both new models improving on their predecessors’ results by a few percentage points in individual agentic coding benchmarks such as FrontierCode-AA 2.0 and Terminal-Bench 4. Astra and Sol also sit on the cost frontier at several reasoning effort levels.
On the AA-Omniscience test, Astra’s hallucination rate at maximum reasoning effort was 59 per cent and Sol’s 55 per cent. The metric shows how often a model guesses the answer to a factual question without knowing the correct answer. The questions are exceptionally hard, so the figures do not reflect error rates in everyday use, The Decoder said.
Plus, Pro, Enterprise and Edu subscribers can use the models in ChatGPT Work and Codex. Developers can access them through OpenAI’s API. Astra’s API price is $8 per 1 million input tokens and $40 per 1 million output tokens. That is 20 per cent lower than GPT-6 Astra and, according to Simon Willison, the same as Claude Fable 5.2. Sol costs $2 for input and $10 for output per 1 million tokens.
Read more: How an agent works · Models