GPT-6.1 Sol

Between a flagship that costs a great deal and a cheap model for bulk work, OpenAI places GPT-6.1 Sol in the middle. It released the model on 29 September 2026 as gpt-6.1-sol on its Responses and Chat Completions APIs, and its API changelog of that day lists multi-agent delegation to sub-agents as a beta feature.

Where OpenAI positions it

In OpenAI's model guide, GPT-6 Astra is the pick for the hardest reasoning and coding, GPT-6 Luna suits high-volume work, and Sol sits between them, meant to balance intelligence against cost. Its standard price is roughly a fifth of Astra's and identical to that of GPT-6 Sol and Claude Sonnet 5.5. The Decoder, covering the release on its own account, reports that paying ChatGPT Work and Codex users get Sol as well as API customers, and that OpenAI is holding back a planned flagship, GPT-6.1 Astra, because of safety concerns.

Specifications

OpenAI's model page lists a 1.05 million token context window, output of up to 128,000 tokens, a knowledge cutoff of April 2026, and text and image input with text output. Reasoning effort has five settings from low to max. Hosted tools supported include web and file search, a code interpreter, computer use and MCP. Customers can choose US or EU data residency, but the faster tier cannot be combined with EU residency. Long-context requests, batch processing, flex processing and the faster tier are each priced differently from the standard rate.

What the benchmarks say

OpenAI's own benchmarks, which it calls preliminary, show a tie with Astra on DeepSWE v1.1, a result 2.1 points behind it on OSWorld 2.0, and a lead over Claude Opus 5.5 on GDP.pdf and AutomationBench. Independent testing by Artificial Analysis gives it 52 on the Intelligence Index (version 4.3.2) at maximum effort, compared with Claude Sonnet 5.5 on 56 and Claude Opus 5.5 on 58 in that version, and the index took 67 million output tokens to complete.