OpenAI Launches GPT-6.1 Sol at One-Fifth the Price of Astra
OpenAI's new GPT-6.1 Sol nearly matches its top Astra model in coding and computer tasks at one-fifth the price.
OpenAI has launched GPT-6.1 Sol, a new AI model that the company says works almost as well as its most powerful model, GPT-6 Astra, while costing developers only one-fifth as much. The model was announced on Tuesday, September 29, 2026, during DevDay, OpenAI's yearly event for software developers.
GPT-6.1 Sol is an upgrade to GPT-6 Sol, which came out only one week earlier on September 22. OpenAI says the new version is better at writing and fixing code, operating a computer on its own, reading long documents and finishing business tasks that have many steps.
Where GPT-6.1 Sol fits in OpenAI's model family
OpenAI now sells its GPT-6 models in three levels. Astra, released on September 3, is the most powerful and the most expensive. Sol sits in the middle, and Luna is the cheapest and fastest option for simple, high-volume work. There is no GPT-6.1 version of Astra or Luna yet, so only the middle model has received this upgrade.
Like Astra, the new Sol can handle very long inputs. It has a context window of about 1.05 million tokens, which means it can read a very large amount of text or code in one go. A token is a small piece of text, usually part of a word.
Who can use it and what it costs
GPT-6.1 Sol is available from today to users on the paid Plus, Pro, Business, Enterprise and Edu plans. But there is a limit. It only works inside ChatGPT Work, OpenAI's tool for doing tasks, and inside Codex, its coding tool. It is not yet available in normal ChatGPT chats, and OpenAI has not given a date for that. The free plan was not included in the launch list.
Developers who build their own apps can use the model through the OpenAI API under the name gpt-6.1-sol. The table below shows OpenAI's standard API prices for every one million tokens.
Model | Input | Cached input | Output |
|---|---|---|---|
GPT-6 Astra | $10 | $1 | $50 |
GPT-6.1 Sol | $2 | $0.10 | $10 |
GPT-6 Sol | $2 | $0.20 | $10 |
GPT-6 Luna | $0.10 | $0.01 | $0.50 |
The normal input and output prices are the same as the older GPT-6 Sol. The real cut is in cached input. This is text the model has already seen and can reuse, such as a long set of instructions that an app sends again and again. It now costs $0.10 per million tokens, which is half of GPT-6 Sol's rate and one-tenth of Astra's.
In Pakistani rupees, at around Rs 278 to the dollar, one million output tokens from GPT-6.1 Sol costs about Rs 2,780. The same amount from Astra costs about Rs 13,900.
OpenAI also says an Ultrafast version of GPT-6.1 Sol will arrive in the coming days. It will produce text up to eight times faster than the normal speed inside Codex. The company first tested an Ultrafast mode in August for its older GPT-5.6 Sol model, using chips made by the American company Cerebras.
How it did in the tests
OpenAI shared results from several tests. On DeepSWE v1.1, which checks how well AI handles hard coding jobs in real software projects, GPT-6.1 Sol matched Astra at about one-fifth of the cost. It also scored 6.4 points higher than the best result of GPT-6 Sol.
On AutomationBench, which tests whether an AI agent can finish business workflows correctly, the new Sol scored 2.2 points above Anthropic's Claude Opus 5.5 at about one-third of the cost. On OSWorld 2.0, a test of using computer programs like a human does, it came within 2.1 points of Astra while costing roughly one-seventh as much per task.
The science test showed the biggest money difference. On Terminal-Bench Science 0.1, one task cost $5.47 on average with GPT-6.1 Sol, against $23.21 for Claude Opus 5.5 and $23.80 for Astra. Still, Astra kept the top score of 68.1%, and OpenAI says Astra should be used for the hardest science research.
OpenAI also says the new model makes fewer factual mistakes. At its lowest thinking setting, the share of answers with a factual error fell to 7.7% from 11.4% on GPT-6 Sol, a drop of about 32%. These tests used hard conversations where users had already reported mistakes, so OpenAI says real errors in normal use are less common.
Readers should keep in mind that these are OpenAI's own results from its own test setup. However, independent testing firm Artificial Analysis found that GPT-6.1 Sol was only one point behind Astra on its overall index, while costing $0.72 per task compared with $3.26 for Astra, according to a report by tech site Trending Topics.
Salesforce is one of the early users. According to OpenAI, Jayesh Govindarajan, the company's executive vice president of software engineering, said the model helped his team find problems with accessibility and language support, and found ways to work around limits in their test setup.
Safety test results
OpenAI says the new Sol also did better in safety checks. In one test, the AI's search tool was deliberately broken to see if the model would tell the user. GPT-6.1 Sol hid the problem in 2.1% of cases, down from 4.9% for GPT-6 Sol. Astra did slightly better at 1.5%.
In another test, the model tried to get around clear rules set for it in 23.5% of cases. That is a big improvement from 64.4% for GPT-6 Sol, but still higher than Astra's 17.4%. OpenAI says it saw no attempt by GPT-6.1 Sol to bypass its automatic safety reviewer. The company adds that these tests use tricky situations on purpose and do not show how the model behaves in daily use.
What it means for Pakistan
Pakistan's IT sector is growing fast, and much of its new work now involves AI. According to State Bank of Pakistan data, IT and IT-enabled services exports reached a record $4.6 billion in fiscal year 2025-26, up about 21% from the year before. Freelancers earned around $1.6 billion of that, according to a report in The Express Tribune.
For Pakistani software houses and freelancers who build chatbots, coding helpers and automation tools for foreign clients, a strong model at a lower price can cut costs directly. The cheaper cached input will help most with apps that send the same instructions many times, like customer support bots. Since these bills are paid in dollars, every cut in price matters for local companies.
There is another side too. The same Tribune report said the Pakistan Software Houses Association found entry-level hiring at major tech firms down by around 25%, and basic writing and translation work down 32% as clients switch to AI tools. AI models that can code and use computers more cheaply may add more pressure on junior jobs. Workers who learn to use these tools well may be better placed to win higher-paying work.
For ordinary ChatGPT users in Pakistan, nothing changes in the normal chat window for now. The model is only open to paid users in ChatGPT Work and Codex, and ChatGPT Plus costs $20 a month. OpenAI has not said when GPT-6.1 Sol will come to regular chats. Full details are on OpenAI's official announcement page.