GPT-6.1 Sol Is Almost a Flagship, Minus the Flagship Price

When the Middle Model Catches Up With the Flagship

AI models usually follow a simple rule. If you want the best result, you pay for the most expensive model. GPT-6.1 Sol bends that rule quite a bit.

It is an upgrade to OpenAI's GPT-6 Sol. In agentic coding, computer use and professional work it nearly matches GPT-6 Astra, the company's most intelligent model. Yet its standard input and output token prices are one-fifth of Astra's.

Tokens are small pieces of text that an AI model uses to count its work and your bill. Everything you send to the model counts as input tokens, and everything it replies with counts as output tokens. The lower the price per token, the cheaper the same job gets, especially when there is a lot of it. People who run AI agents will feel it most, since these programs go through a long chain of steps on their own and call the model many times in a row.

The Three GPT-6 Models in Plain Words

The GPT-6 lineup now has three models.

- GPT-6 Astra is OpenAI's smartest model, built for the best possible results.

- GPT-6.1 Sol offers intelligence almost at the Astra level for five times less.

- GPT-6 Luna is a fast and efficient solution for everyday tasks at scale.

Astra remains the company's most powerful model overall. Sol offers a new balance of capability and cost, and Luna suits simple, high-volume tasks where speed and price matter.

Cached input deserves a separate mention. It is reused context, for example a long document that the model rereads across several requests in a row. For GPT-6.1 Sol it costs $0.10 per million tokens, which is 95% below the standard input price and half of what GPT-6 Sol charged for cached input. For agent tasks that need to reuse context again and again, that is a noticeable saving.

In practice. For complex tasks you do not have to reach for the most expensive model right away. You can start with GPT-6.1 Sol and switch to Astra only where you need maximum quality.

Coding and Computer Use

To show how much stronger the model has become, OpenAI uses tests called benchmarks. A benchmark is a set of identical tasks that lets you compare different models with each other.

In DeepSWE v1.1, AI agents solve hard software development tasks in real codebases. GPT-6.1 Sol reaches the level of GPT-6 Astra here at roughly one-fifth of the cost. It also beats the best GPT-6 Sol result by 6.4 percentage points while using less reasoning effort and less money. A percentage point is the difference between two percentages, for example between 30% and 36%. And "reasoning effort" is the mode in which the model thinks for longer or shorter before it answers.

OSWorld 2.0 checks how agents handle long sequences of actions on a computer in everyday and professional tasks. On the offline set, GPT-6.1 Sol at maximum reasoning effort beats GPT-6 Sol by seven percentage points at less than half the cost. It trails Astra by only 2.1 percentage points, and a task costs about seven times less.

In practice. Computer use means the AI opens programs, clicks buttons and fills in fields on its own. It can help with routine work such as moving data between services. You should still check the result.

Documents and Business Workflows

GDP.pdf is a test of professional questions about complex PDF documents with tables, charts, diagrams and small print. The questions come from finance, healthcare, law and seven other professional fields. GPT-6.1 Sol beats Opus 5.5 with fallbacks at a task cost more than twice as low at every reasoning level tested. It approaches Astra's results at about one-fifth of the task cost.

AutomationBench 1.0.6 checks whether agents carry out business workflows correctly from start to finish. They get 47 tools for sales, marketing, operations, support, finance and HR. At medium reasoning effort, GPT-6.1 Sol beats Opus 5.5 by 2.2 percentage points at about one-third of the cost, and it beats GPT-6 Sol by 4.8 percentage points at the same setting. OpenAI also warns that the cost shown for Claude Fable 5.1 on its chart is understated, because it leaves out the fallback models that had to be called in about 40% of tasks.

In practice. If you work with reports, contracts or tables in PDF, you can ask the model to find the data you need and retell it in plain words. And if your job involves sales, marketing or support, tests like these are closest to your everyday tasks. It is still best to check the figures against the original.

When You Still Need Astra

Terminal-Bench Science 0.1 evaluates scientific workflows, including data analysis, simulation and theorem proving. At maximum reasoning effort, GPT-6.1 Sol more than doubles the GPT-6 Sol result at less than half the task cost. The average cost of one task is $5.47 against $23.21 for Opus 5.5 and $23.80 for Astra, which means costs are more than 75% lower.

Still, the highest result among the tested models, 68.1%, belongs to GPT-6 Astra. OpenAI recommends it for the hardest scientific research tasks.

You can choose like this.

- GPT-6 Astra for the hardest tasks, where you need maximum quality.

- GPT-6.1 Sol for serious work with code, documents and computers at a lower price.

- GPT-6 Luna for quick everyday tasks.

That way you do not overpay for power your task does not need, and you do not pick a model that is too simple where the cost of a mistake is high.

How Accurate and Safe the New Model Is

GPT-6.1 Sol is more accurate with facts. The biggest improvement shows up at low reasoning effort, where the share of answers with a factual error drops from 11.4% to 7.7%, a reduction of about 32%. Across all tested reasoning settings, its error rate stays within 1.9 percentage points of Astra's at less than one-fifth of the cost per task. The test counts the share of answers with at least one factual error on anonymized conversations where users pointed out a mistake by an earlier model. These requests are deliberately hard and do not reflect typical use.

In tests of alignment with human intentions and values, the model also improves on GPT-6 Sol and gets closer to Astra.

- GPT-6.1 Sol is more open about its limitations.

- It follows user intent and safety constraints more reliably.

- It is less likely to hide a broken search tool, break explicit constraints or take unauthorized actions in agent tasks.

- OpenAI found no attempts to bypass the automated safety review, just as with Astra and GPT-6 Sol.

In the broken search tool test, the check is whether an agent tells the user about the failure instead of giving the most likely answer. GPT-6.1 Sol fails to report the problem in 2.1% of cases, GPT-6 Sol in 4.9%, Astra in 1.5% and Luna in 28.7%. The tasks are chosen to expose mistakes, so the numbers do not show how often failures happen in normal work. Still, facts, numbers and links from an AI answer are better checked by you.

In practice. If an AI searches the internet for you, ask it to tell you plainly when the search did not work. This is exactly the behavior OpenAI tests in a separate benchmark.

What GPT-6.1 Sol Costs at OpenAI

Starting today, GPT-6.1 Sol is available to all Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. It is not yet available in Chat. Developers can connect it through the OpenAI API under the name gpt-6.1-sol. An API is a way to connect an AI model to your own programs and services.

Standard API prices are as follows. It is $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens.

In the coming days OpenAI will also offer GPT-6.1 Sol Ultrafast, with up to eight times faster token generation than its standard speed in Codex.

How to Try GPT-6.1 Sol in Unitool

You do not need a ChatGPT subscription for this. GPT-6.1 Sol is already in Unitool, and you can start in a couple of minutes.

1. Open a chat in Unitool and pick GPT-6.1 Sol from the model list at the top.

2. Click "Create new chat" so the new task does not mix with your earlier conversations.

3. If you need to, attach files with the plus button. The model accepts images, PDFs and text files.

4. Type your task in the "Ask me anything" field or dictate it with the microphone.

5. Check the line under the input field. It shows what the model costs, currently 0.313 for input and 1.565 for output per 1,000 tokens.

A good first request could be:

Here is a PDF report. Find the five most important figures, explain in plain words what they mean and show where in the document they are.

The model did especially well on complex PDFs with tables and fine print, so this makes a fair first test. If you need code, ask it to describe a plan first and only then write.

Pick the Model by Task

AI models update so fast that it is hard to keep track of every version. That is why it is better to choose by trying. Take your own work task, run it on GPT-6.1 Sol and then on another model, and compare which result is better and cheaper.

For that, Unitool brings AI models for text, images, video and voice into one subscription, for example ChatGPT, Claude, Gemini, Midjourney, Nano Banana, Seedance, ElevenLabs and Suno. You switch between them in the same window, without a dozen accounts. And if the task is big, turn on Unitool MCP, and the chat will make the images, voice-over and video from your description.

Open Unitool, pick GPT-6.1 Sol and hand it the task you have been wanting to get off your plate.