Skip to content

OpenAI · Solving complex problems and programming

GPT-5.6 Sol

A model for the hardest logic problems, mathematics and autonomous coding. It makes sense where results can be verified through tests or additional sources.

Compare with another model

Costs
High · standard pricing is $5 per million input tokens and $30 per million output tokens; Fast mode costs twice as much
Speed
Medium to slow depending on thoroughness; Fast mode can be up to 2.5× faster
Availability
Online service from OpenAI and the Codex tool
Input length
Varies by product and mode
Inputs
Text and, depending on the service used, image inputs

Data checked . Release date unverified.

Indicative capability profile

Where the model is strong

The five levels are our clear summary of the results below. They are not a ranking that applies to every task.

Analysis One of the best results in logic tasks and mathematics.
Coding Very strong in both coding and terminal work.
Autonomous work across multiple steps Suitable for demanding work in multiple steps with results checked.
Writing and content Strong language capabilities, but unnecessarily costly for everyday content.
Speed The most thorough setting increases both cost and processing time.

Ideal use

When to choose it

  • Advanced coding and terminal work
  • Mathematical and analytical problems
  • Tasks with measurable results and automated checks

Usage boundaries

When to choose another model

  • Simple communication and routine content
  • Low-cost processing of large volumes of data

Traceable supporting sources

Data and measurement sources

Distinguish between the manufacturer's documentation and the results of a specific test. Measurements also depend on the settings and the task set used.

LiveBench

LiveBench 2026-06-25

82.4 overall · 91.7 for problem solving

The highest overall score in the specified version of LiveBench.

Open original source ↗
Artificial Analysis

Coding Agent Index

80 points in Codex

The result measures the model together with the Codex tool it worked in.

Open original source ↗
OpenAI

Price and Fast mode

$5 input · $30 output · Fast at twice the price

The price is per million tokens with a short context; Fast mode can be up to 2.5× faster.

Open original source ↗

Trends over time

Related events from AI Radar

Search more →
OpenAI major

OpenAI cuts prices for GPT-5.6 Luna and Terra models, says optimization of its own infrastructure by Sol made this possible

Anyone using OpenAI API for their own projects or tools with a low-cost model (Luna) will pay significantly less for the same work than before, as illustrated by Simon Willison, who switched from Gemini to Luna because of the price.

Companies running products on cheaper GPT-5.6 models can reduce inference costs by as much as tens of percent or consider switching from competing providers (Google, Anthropic), intensifying price pressure across the AI industry.

✓ 2
DeepSeek important

DeepSeek has released V4 Flash 0731 with performance close to GPT-5.6 Luna at a significantly lower price

Anyone choosing a model for development with LLMs based on value for money gains another inexpensive option with performance close to more expensive models, but the result depends substantially on the selected reasoning level.

According to the reported figures, companies running LLMs at higher volumes can reduce costs per task by approximately 60 % compared with GPT-5.6 Luna at comparable performance, with additional savings on repeated context thanks to a cache discount of 98 % compared with the usual 90 %.

✓ 2
Anthropic worth noting update

OpenAI claims that GPT-5.6 Sol outperformed Opus 5 on ARC-AGI-3 – but only with its own testing environment

New: OpenAI achieves 38.3 % on ARC-AGI-3 with its own API setup (Retained Reasoning + Compaction); In the official harness without these features, GPT-5.6 Sol scores only 7.8 %; ARC-AGI-3 was deliberately designed without external aids to measure raw model performance; Retained Reasoning and Compaction are not part of the official testing environment; Benchmark results reflect both the model and the technical setup around it

When reading benchmark rankings, you need to verify that the figures being compared come from the same testing environment – a difference of 7.8 % versus 38.3 % for the same model shows that a score alone, without methodological context, can be misleading.

Companies selecting an AI model for production deployment should not rely on marketing benchmark figures, because the same model can show significantly different performance depending on which additional API features (e.g. preserving context between steps) the provider used during testing.

1 source

Next step

Compare the model using your actual task.

A benchmark narrows the selection. A short trial on your data determines the choice.

Open comparison →