OpenAI didn't ship the GPT-6.1 Astra model it had lined up. It shelved that one on Monday after it failed internal safety tests. Instead, at DevDay on Tuesday, it announced GPT-6.1 Sol, which OpenAI says nearly matches GPT-6 Astra on coding, computer use and professional work "at one-fifth of Astra's standard input and output token prices."

GPT-6.1 Sol: near-Astra intelligence for a fifth of the price.

It’s the most cost-efficient model for its performance available today. pic.twitter.com/hH8PnE2Oxk

— OpenAI (@OpenAI) September 29, 2026

Start with the price. In the API, gpt-6.1-sol costs $2 per million input tokens and $10 per million output tokens, against Astra's $10 and $50, according to OpenAI's pricing page. That's the same list price as GPT-6 Sol, which launched only a week earlier. The real change is cached input, now $0.10 per million tokens. That's 95% below standard input and half of GPT-6 Sol's cached rate, which helps agents that keep re-sending the same long context.

OpenAI's own benchmark numbers make the case:

• DeepSWE v1.1 (real-codebase engineering): matches GPT-6 Astra at roughly a fifth of the cost, and beats GPT-6 Sol's best score by 6.4 points.
• AutomationBench (business workflows): 2.2 points above Anthropic's Claude Opus 5.5 at medium effort, at about a third of the cost.
• OSWorld 2.0 (computer use): within 2.1 points of Astra at maximum effort, at about one-seventh of the cost per task.
• Terminal-Bench Science 0.1: $5.47 per task on average, versus $23.21 for Opus 5.5 and $23.80 for Astra. Astra still posts the top score (68.1%), and OpenAI recommends it for the hardest research.
• Factuality: on deliberately tricky prompts, answers with a factual error fall from 11.4% to 7.7% at low effort, about 32% fewer than GPT-6 Sol.

Given this week, the safety section may get the closest reading. OpenAI says 6.1 Sol shows "substantial improvements over GPT-6 Sol in our alignment evaluations, bringing it closer to GPT-6 Astra." It had lower failure rates on respecting explicit restrictions and avoiding unauthorized outcomes in agentic tasks, and OpenAI "observed no attempts to bypass an automated safety reviewer." Staying within scope was reportedly where 6.1 Astra fell short.

Sam Altman played down the Astra decision in an interview with CNBC: "Often we build a model, we test it, it doesn't meet our standards, we change it, we launch it later."

Where can you use it? Not in regular ChatGPT chat, yet. OpenAI's announcement says it's "available starting today to all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex." The help center's release notes are more cautious: OpenAI is "rolling it out in ChatGPT Work and Codex, starting with Pro users and expanding to Plus, Business, Enterprise, and Edu." So if it's missing from your Plus account, that may be why. Enterprise and Edu admins also have to enable it.

Developers can use it in the API now, and it's the new default model in Codex CLI 0.159.1, per the Codex changelog. An Ultrafast version, which OpenAI says will generate tokens up to 8x faster in Codex, is due "in the coming days."

If OpenAI's numbers hold up outside its own tests, Sol is the model most people will actually use. Astra stays the pick when only the best will do.