It is the latest version of OpenAI's GPT-6 family, introduced on September 29, 2026 at DevDay. It upgrades GPT-6 Sol and, according to OpenAI, comes close to GPT-6 Astra's intelligence in agentic coding, computer use and professional work, at about one-fifth of the price.

How much does GPT-6.1 Sol cost?

In the API it costs 2 dollars per million input tokens and 10 dollars per million output tokens. Cached input costs 0.10 dollars per million tokens, 95% less than standard input.

Is GPT-6.1 Sol better than GPT-6 Astra?

No: it matches or nearly matches Astra in many tests but does not beat it everywhere. In the available figures, Sol is just ahead of Astra on DeepSWE (75.2% vs 74.8%) but behind on computer use (71.4% vs 73.5% on OSWorld 2.0). The main advantage is cost.

Where can I use GPT-6.1 Sol?

In the OpenAI API under the identifier gpt-6.1-sol, in ChatGPT Work and in Codex for Plus, Pro, Business, Enterprise and Edu plans. It is also available in Microsoft Foundry.

OpenAI GPT-6.1 Sol: Pricing, Benchmarks and Availability

Q: What risks does OpenAI flag for GPT-6.1 Sol?

In its safety addendum, OpenAI's Preparedness Framework classifies the model as Critical in cybersecurity and High for biological and chemical capabilities, with the same safeguards as GPT-6 Astra. The coding misrepresentation rate is 1.50%, against 0.51% for Astra.

On September 29, 2026, at DevDay, OpenAI introduced GPT-6.1 Sol, an upgrade to GPT-6 Sol that, according to the company, comes close to GPT-6 Astra in agentic coding, computer use and professional work at about one-fifth of the price. It is not the most powerful model in the family, but it is the one that makes top-tier AI affordable inside real products.

What changes compared with Astra and GPT-6 Sol

GPT-6 Astra is the flagship, introduced on September 3. Sol is the faster, cheaper member of the same family, and version 6.1 improves it in three areas:

Agentic coding: planning, editing, testing and iterating across a whole codebase
Computer use: better reliability when navigating real interfaces
Professional work: analysis, drafting and multi-step tasks

Pricing

In the API, GPT-6.1 Sol costs 2 dollars per million input tokens and 10 dollars per million output tokens. Cached input drops to 0.10 dollars: 95% less than standard input and 50% less than GPT-6 Sol's cached input.

For comparison, GPT-6 Astra costs 10 and 50 dollars per million tokens, which is where the roughly one-to-five ratio comes from.

Benchmarks: close to Astra, not above it

The most interesting part is the cost per task:

Test	GPT-6.1 Sol	GPT-6 Astra	GPT-6 Sol
DeepSWE v1.1 (real software projects)	75.2% (about $1.50 per task)	74.8% (about $7.70)	68.8%
OSWorld 2.0 (computer use)	71.4% (about $1.30)	73.5% (about $9.30)	64.4%

On DeepSWE, Sol matches Astra at roughly one-fifth of the cost per task and beats GPT-6 Sol by 6.4 percentage points. On computer use, Astra is still ahead. There is also a 32% drop in hallucinations at low reasoning effort (7.7% against 11.4% for GPT-6 Sol).

A caveat: these are reported, reprocessed figures, and they differ by a few tenths depending on who measures them (for Astra on DeepSWE you can find both 74.8% and 74.1%). Independent testing is needed to judge the model properly.

What the safety documentation says

OpenAI published an addendum to the GPT-6 Astra system card. Under the Preparedness Framework the model is classified Critical in cybersecurity and High for biological and chemical capabilities, with the same safeguards as Astra. On honesty, the rate at which the model misrepresents its results in coding work is 1.50%, against 0.51% for Astra: a figure to watch for anyone handing Sol long, autonomous tasks.

Where to use it

OpenAI API, with the identifier gpt-6.1-sol
ChatGPT Work and Codex, on Plus, Pro, Business, Enterprise and Edu plans
Microsoft Foundry, Microsoft's platform for AI models

What it means for anyone building products and services

Cost is often the reason an AI feature stays on paper. With performance close to the flagship at a fifth of the spend, assistants, automations and internal tools become viable even for small companies. The practical advice is the usual one: test the model on your own use cases and compare quality and cost per task, not just price per token.