OpenAI releases GPT-6.1 Sol at DevDay 2026

Background

DevDay is OpenAI's developer event, where the company presents new models alongside new API features. By the 2026 edition, the GPT-6 family was the reference point against which new releases were described: GPT-6 Astra and GPT-6 Sol are the two earlier models that OpenAI itself used for comparison when introducing the next version^[1][2].

What was announced

On 29 September 2026, at DevDay, OpenAI released GPT-6.1 Sol^[2]. According to TechCrunch, OpenAI said the new model nearly matched GPT-6 Astra while costing less^[1]. That combination — capability close to a larger model, at a lower price — was the headline framing of the launch.

Not everything planned for the day shipped. A GPT-6.1 Astra had been planned, but it was held back after it regressed on safety evaluations^[2]. In other words, the version that would have sat above Sol was pulled because its safety results got worse, not better, during development.

Alongside the model, OpenAI added computer use to its Agents API at DevDay 2026^[3]. Computer use is the ability for an agent to operate a computer interface — reading a screen and issuing clicks and keystrokes — rather than only calling dedicated tool connectors.

What changed in the model

OpenAI described GPT-6.1 Sol as an improvement over GPT-6 Sol in three areas: programming and debugging, understanding documents, and executing multistep workflows^[1]. These are the tasks where a model has to hold a plan across many steps and keep track of state, so the comparison to the previous Sol model is the relevant one for people already using it.

Availability was staged. TechCrunch reported that GPT-6.1 Sol was offered in ChatGPT Work and Codex but not yet in ChatGPT chat^[1]. The first users were therefore people doing work tasks and coding rather than the general chat audience.

On safety, OpenAI said it observed no attempts to circumvent the automated safety reviewer, and that this was consistent with what it had seen for GPT-6 Astra and GPT-6 Sol^[1]. The claim is narrow: it concerns attempts to get around the automated reviewer, not the overall safety profile of the model.

What followed

GPT-6.1 Terra followed on 26 October 2026, less than a month after Sol^[2]. The 6.1 line therefore expanded past a single release shortly after DevDay, while the withheld Astra version did not appear in this window.

Why it matters

The release is a compact example of how a frontier model launch is packaged. Capability is stated relative to an existing, larger model rather than in the abstract; price is part of the claim; and the rollout is split by product surface, with work and coding tools first and the main chat product later^[1].

The withholding of GPT-6.1 Astra shows the safety evaluation acting as a gate rather than a formality: a model that was planned for release did not ship because its safety results regressed^[2]. And the addition of computer use to the Agents API in the same event ties the model release to a change in what agents built on the platform can do^[3].

Read together, the sequence — Sol in late September, Terra in late October, Astra postponed — describes a release cadence organised around named variants within a version number rather than a single flagship launch^[2].