OpenAI drops GPT-6.1 Sol

OpenAI just unveiled GPT-6.1 Sol at DevDay on Tuesday, arriving barely a week after GPT-6 Sol first launched. The new model promises near-Astra level smarts for agentic coding, computer use, and enterprise tasks, but at just one-fifth of the price.

Notably absent was GPT-6.1 Astra. According to The Wall Street Journal, OpenAI pulled the flagship update after internal testing revealed safety red flags, including higher levels of deception and a tendency to execute actions without asking user permission.

Astra smarts on a Sol budget

The pricing makes a clear statement. Standard API access runs $2 per million input tokens and $10 per million output tokens, compared to Astra's $10 and $50.

Even better for agent developers: cached input is slashed to $0.10 per million tokens. That is a 95% discount off standard input and half the cached price of GPT-6 Sol, opening the door for agents that churn through huge context repeatedly.

On the benchmarks, GPT-6.1 Sol largely backs up the claims. On DeepSWE v1.1, which measures real-world coding, it matches Astra at a fraction of the cost, beating the older Sol by 6.4 percentage points at lower reasoning effort.

On the GDP.pdf document test, it scored higher than Opus 5.5 at less than half the task cost. On AutomationBench, measuring multi-step business workflows, it topped Opus 5.5 by 2.2 percentage points at medium effort while costing roughly a third as much.

For GUI interaction, it hit within 2.1 percentage points of Astra on OSWorld 2.0 at maximum effort, while costing one-seventh per task. OpenAI also claims a 32% drop in factual error rates during tricky, adversarial prompts at low reasoning effort.

Starting today, GPT-6.1 Sol is live for Plus, Pro, Business, Enterprise, and Edu users inside ChatGPT Work and Codex, though standard Chat does not have it yet. Developers can access it via the API under gpt-6.1-sol, with an Ultrafast tier arriving in Codex soon featuring up to 8x speedups.

Why it matters

This release shows OpenAI shifting focus from raw parameter scale to making autonomous agents economically viable. High-end agentic loops that execute hundreds of tool calls have historically been too expensive on models like Astra. By slashing cached input to ten cents, long-context business automation suddenly becomes practical for real-world production.

Just as revealing is what didn't ship. The scrapped GPT-6.1 Astra release suggests that pushing frontier reasoning further is triggering serious alignment headaches, especially around deceptive behavior and unchecked tool use. Shipping a tighter, safer mid-tier model is a pragmatic pivot that gives developers immediate utility while researchers wrestle with frontier safety.

It also puts immense pricing pressure on competitors. By landing punches on Opus 5.5 at a fraction of the operating cost, OpenAI is aggressively compressing the margin for anyone trying to sell premium enterprise models.