Migrating: GPT-6 Sol to GPT-6.1 Sol

OpenAI specifically flags this migration: if you already use gpt-6-sol, it says to review its migration guidance before switching. That's notable for a move between two models released a week apart at the same input and output price, and the reason is the effort floor.

The change that breaks requests

GPT-6 Sol accepts none reasoning effort; GPT-6.1 Sol doesn't, and doesn't accept minimal either. Requests that set either need low instead. Two knock-on effects follow. Reasoning tokens start appearing where there were none, billed as output, so a latency-tuned call gets a little slower and a little dearer per request. And on the API, the sampling parameters that only work with none — temperature, top_p and log probabilities — have to come out. Function calling changes too: GPT-6 Sol supports it in Chat Completions only with effort set to none, while GPT-6.1 Sol requires the Responses API for any tool calling.

The change that saves money

Cached input. GPT-6.1 Sol bills cache reads at 0.05x its input rate instead of 0.1x. Input, output and cache-write rates are unchanged, so every request is the same price or cheaper, and cache-heavy agentic sessions get noticeably cheaper.

What carries over

Context window, maximum output, long-context rule, and request and token limits per minute. GPT-6.1 Sol's knowledge cutoff is slightly newer. Its rate-limit table has no batch-queue column, so if you plan Batch capacity, note that no published queue limit exists for it.

Before switching

If your traffic never sets none, this is close to a free upgrade — and it's the CLI's default. If it does, test the same tasks at low and check whether the added latency and output tokens are acceptable before moving. GPT-6.1 Sol vs GPT-6 Sol has the full comparison.