GPT-6 Sol: What Changed From GPT-5.6 Sol, and What Did Not

The headline on GPT-6 Sol was the price. OpenAI cut it to $2 per million input tokens and $10 per million output, half of what GPT-5.6 Sol was listed at, and attributed the cut to caching and inference efficiency. That is a real change, but it is the least interesting part of the release if you are deciding whether to move a production workload.

The more useful questions are what the capability claims are based on, how far Sol sits below the flagship Astra, and what it means that OpenAI has already published a newer Sol. This guide takes them in that order.

What changed from GPT-5.6 Sol

OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, following the flagship GPT-6 Astra earlier that month. The company says both were trained with methods similar to Astra's and improve on the equivalent GPT-5.6 models in professional work, factuality, coding, computer use and alignment, while using fewer tokens. It also reported that on its internal factuality evaluation Sol makes roughly half as many mistakes as its predecessor.

The concrete, checkable changes are these:

• Price: $2 input and $10 output per million tokens, down from a reported $4 and $20 for GPT-5.6 Sol. Cached input is $0.20.
• Context: 1,050,000 tokens, with up to 922,000 tokens of input and 128,000 of output.
• Knowledge cutoff: 20 April 2026.
• Reasoning effort: six levels, none, low, medium (default), high, xhigh and max.
• Modalities: text and image in, text out.

OpenAI's model page describes Sol as built for complex coding and agentic workflows. That matches where the vendor chose to put its benchmark claims.

Reading OpenAI's benchmark claims

The launch figures that reached the press were comparisons rather than raw scores. On OSWorld 2.0, OpenAI said Sol at xhigh effort performs close to Claude Opus 5 at medium effort for about 80% less cost. On AutomationBench, a test of workflows spanning several applications, it said Sol beat Claude Opus 5 at max effort while costing 9% as much per task.

Both claims are about cost-adjusted performance, which is the right framing for a mid-tier model. They are also vendor-run, on vendor-chosen settings, against a competitor configuration the vendor picked. Notice that the OSWorld comparison puts Sol at xhigh against a rival at medium. That is a fair way to show what Sol can reach, but it means Sol's cost advantage on your workload will depend on whether you need xhigh to get there.

Vendor numbers measured against an independent harness tend to come down rather than up, so treat these as an upper bound on what you will see.

Sol against Astra

Astra lists at $10 input and $50 output, five times Sol on both. OpenAI does not claim Sol matches Astra on the hardest work; it positions Sol as the model that carries Astra's methods into everyday professional use at a lower cost. The spec sheets are otherwise close: the same window, the same output limit, the same image input.

In practice the split looks like this. Sol is a reasonable default for coding agents, document work and multi-step tool use. Astra earns its price on long unattended runs, hard computer-use tasks and problems where Sol has already failed. If you are unsure, start on Sol and escalate. A five-times price gap means Astra only pays for itself when it saves a retry or a human correction most of the time. The GPT-6 Astra guide covers the flagship in detail.

There is a middle step as well. GPT-6 Sol Pro is Sol with pro reasoning mode enabled, billed at Sol's rates for more tokens. For a hard task, it is often worth trying before paying Astra prices.

Sol has already been superseded

OpenAI's model page for GPT-6 Sol now notes that GPT-6.1 Sol is the newer Sol model. That matters for anyone starting fresh: the 6.1 release is the one OpenAI will optimise documentation and defaults around. One documented difference is that GPT-6.1 Sol does not accept none reasoning effort, while GPT-6 Sol does.

That makes GPT-6 Sol the choice when you need reasoning switched off entirely, for latency or for Chat Completions tool calls, or when you have already validated your prompts against it and want stability. The GPT-6.1 Sol guide covers what changed in the point release. When a dated snapshot is offered, pin it in production; an undated alias can move under you.

Practical API notes

A few details from OpenAI's documentation that affect how you integrate Sol:

• Use the Responses API for tools. On Chat Completions, Sol supports function calling only when reasoning effort is none. The Responses API supports function calling at any effort and adds hosted tools including web search, file search, code interpreter, computer use and MCP.
• Watch the 272K line. Prompts above 272K input tokens are billed at twice the input and cache rates and one and a half times the output rate, for the entire request.
• Batch is half price. OpenAI bills Batch and Flex at 50% of standard. The catalogue lists openai/gpt-6-sol:batch at $1 input and $5 output per million tokens at the time of writing, the same model on an asynchronous endpoint.

Deciding whether to move

If you are on GPT-5.6 Sol today, the price cut alone justifies a test. Take twenty tasks from your own history, run them on both models at the same effort, and compare three things: pass rate, output tokens per task and cost per completed task. A model that claims to use fewer tokens should show it in that second column.

If you are coming from another vendor, run the same set against your current model and Sol at medium and high. How to benchmark LLMs on your own work has the harness setup. The result will tell you more than any of the comparisons above.