GPT-5.6 Sol Guide: The July Flagship, Three Generations Later

GPT-5.6 Sol had a short reign as OpenAI's flagship. It shipped on 9 July 2026; by the end of September OpenAI had released GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol. Yet Sol is still worth understanding properly, for two reasons. Many of the API features people now take for granted arrived with it, and its price has moved enough that the version of Sol you read about in July is not the one you would pay for today.

OpenAI's model page calls it the "GPT-5.6 flagship model for complex professional work". OpenRouter's catalogue adds that it is "particularly strong at command-line and multi-step coding tasks". OpenAI has not published a parameter count or architecture.

What GPT-5.6 added, beyond the model

The changelog entry for the release is mostly about platform features, and they apply across Sol, Terra and Luna:

• The max reasoning effort level, above xhigh. Sol accepts none, low, medium (default), high, xhigh and max.
• Pro mode as a request setting, reasoning.mode: "pro", billed at the model's standard token rate instead of a separate Pro price.
• Persisted reasoning, reusable across Sol, Terra and Luna, so a conversation can change tier without discarding its reasoning state.
• Programmatic Tool Calling and a beta of multi-agent orchestration in the Responses API.
• Explicit prompt caching controls, with cache writes billed at 1.25 times the uncached input rate.
• Images at original dimensions as input.

The specification is otherwise familiar: a 1,050,000-token window with up to 922,000 input and 128,000 output tokens, a knowledge cutoff of 16 February 2026, text and image in, text out, on Chat Completions, Responses and Batch. The gpt-5.6 alias routes to Sol. If you want the version with pro mode switched on, the GPT-5.6 Sol Pro guide covers it.

The price has moved, and may move again

Sol currently lists at $4 per million input tokens, $0.40 cached and $20 output, on both OpenAI's page and OpenRouter. That is not the launch price. On 21 August OpenAI cut it, describing the change as 20% lower input and 33% lower output pricing. Its page now lists GPT-5.5 at $5 and $30, so Sol is cheaper per token than the model it replaced.

Two qualifications. OpenAI calls the current rate promotional pricing "available at least through November 21, 2026", so budget for the possibility that it rises. And the long-context rule still applies: prompts over 272K input tokens bill at twice the input rate and one and a half times the output rate for the whole request.

OpenRouter also lists a :batch variant at $2 and $10 at the time of writing — the same model on an asynchronous discounted queue. For evaluation runs and offline processing, that halves the bill.

Latency at the top of the dial

Six effort levels make Sol flexible, and max makes it slow. Artificial Analysis benchmarked Sol at max and recorded a time to first token of roughly 97 seconds on OpenAI's API at the time of writing, against a median of under four seconds across the models it tracks. That is fine for an unattended agent working through a migration and unusable for anything a person waits on.

OpenAI's answer is Fast mode, introduced on 30 July as a replacement for Priority Processing. For Sol it "delivers up to 2.5x faster speeds than standard processing at twice the price". Requests already tagged as priority use it automatically. It is a reasonable trade for interactive agent sessions where developer time is the expensive resource. Best model for low latency covers when time to first token matters more than raw capability.

The cheaper fix is to stop defaulting to high effort. Most tool-calling steps in an agent loop do not need xhigh, let alone max. Reserve the top levels for planning and hard debugging turns.

The GDPval result, read carefully

GPT-5.6 Sol at max effort scored 1,747.8 on GDPval-AA v2, second behind Claude Fable 5 Max at 1,815 as of late August 2026. GDPval measures performance on realistic professional deliverables rather than puzzles, which makes it more relevant than most benchmarks to the "complex professional work" OpenAI positions Sol for.

Two caveats. The score is for the max setting, which is also the slow and verbose one; at medium you are using a different trade-off, and the score does not tell you how much of it survives. And the ranking predates the GPT-6 releases, so do not read "second" as current. GDPval explained covers what the benchmark does and does not capture.

GPT-6 Sol is now the comparison that matters

On 22 September OpenAI released GPT-6 Sol at $2 input, $0.20 cached and $10 output per million tokens — half of GPT-5.6 Sol's current rate. GPT-6 Astra, released on 3 September, is what OpenAI now calls its most capable model, and OpenAI's own guidance describes it as generally better than GPT-5.6 Sol at staying coherent during long tasks. Artificial Analysis has marked its GPT-5.6 Sol page as deprecated and points readers to GPT-6 Sol.

So the honest position in October 2026: GPT-5.6 Sol is not the model to start a new project on without testing the GPT-6 options. The GPT-6 Sol guide is the direct comparison.

Who should still choose Sol

• Teams with a validated Sol integration. Prompts, graders and agent scaffolds tuned against it are an asset; re-validation is a real cost.
• Mixed-tier routers on GPT-5.6. If Luna or Terra handle most traffic, Sol is the escalation target that shares their persisted reasoning. The GPT-5.6 Luna guide covers the cheap end of that arrangement.
• Workloads where your own evaluation still favours it. Price is half the comparison; tasks completed is the other half.

Whatever you choose, set effort explicitly, measure cost per completed task rather than per token, and re-run the comparison when the promotional price expires.