GPT-6 Sol Vision Reasoning (OpenAI)

Overview

GPT-6 Sol is an OpenAI GPT-6 family model built for coding and agent workflows. It reasons over very long text and can analyze images you attach. It produces text answers, and it can also return structured tool calls when you provide tools. This helps you turn screenshots, specs, and large codebases into clear next steps and concrete outputs.

What you can build

• Repo-level code review, refactors, and migration plans across many files
• Multi-step debugging from logs, stack traces, and error screenshots
• Agent workflows that search the web, look up internal files, and then write a final plan
• Visual QA for UIs: read labels, spot layout issues, and explain what changed
• Document and screenshot triage: extract key facts, tables, or visible text into summaries
• Structured extraction pipelines that return JSON for downstream processing

Inputs

• A single instruction written as plain text.
• Or a JSON-formatted chat history with roles and message content. Use this instead of a single instruction, not both.
• Up to 10 image files per run on Wiro, for visual analysis. GPT-6 Sol supports common vision formats like PNG, JPEG, WebP, and non-animated GIF.
• Optional system-level instructions that set global behavior for the run.
• An optional reasoning effort level you choose, from no reasoning to maximum effort.
• An optional setting that lets the model browse the web during the run.
• An optional verbosity level that controls how short or detailed the writing is.
• An optional output length cap, up to 128,000 generated tokens.
• Optional tool definitions, provided as JSON, so the model can call your functions.
• Optional tool rules that let you allow, forbid, require, or force a specific tool.
• An optional switch that lets the model return multiple tool calls in a single turn.
• Optional user and session identifiers to keep chat history grouped across runs.
• Optional continuation inputs that resume a prior run and supply tool results back to the model.

Outputs

The model returns a text response.

If you enable tools, the output may include one or more structured tool calls, followed by text that uses the tool results. If you request structured outputs, it can produce JSON-formatted content designed to parse cleanly.

Recommended settings

• Use the default reasoning level for most work, then increase it when tasks require deeper multi-step planning.
• Set reasoning to the lowest level when you need function calling through Chat Completions tool flows.
• Keep verbosity low for short, spec-like answers. Increase it for explanations and writeups.

Limitations

• Output is text. Images are supported as input only.
• Audio and video inputs are not supported.
• The context window is 1,050,000 tokens, with up to 128,000 output tokens.
• The published knowledge cutoff for GPT-6 Sol is April 20, 2026.
• Tool calling works best through the Responses API tool system. In Chat Completions, function calling is supported only when reasoning is set to none.
• When reasoning is enabled (not none), some sampling-style controls used on older model families may be unsupported.
• Vision quality depends on the image. Blurry photos, heavy compression, tiny text, watermarks, and messy screenshots can cause misses or wrong reads.
• Very large images can be rejected by vision preprocessing limits. Downscale images when this happens.

Safety & compliance

Use GPT-6 Sol in line with OpenAI Usage Policies. Don’t use it to facilitate wrongdoing, exploit vulnerabilities without authorization, create sexual violence content, or generate instructions for weapons.

If you build agent flows with browsing or tools, keep clear boundaries. Log tool actions and require human review for high-stakes decisions like medical, legal, financial, or employment outcomes.

Some OpenAI access programs can change safeguard behavior for cyber workflows. Only enable those modes if your organization is approved and your use is authorized.