GPT-6.1 Sol vs GPT-6 Sol

These are the two current Sol models, released a week apart, and on most of the table they're the same model at the same price. The decision comes down to a pair of rows — cached input and the bottom of the effort range — and which one your workload leans on.

Where they match

Input and output cost the same at every tier both publish. The context window and maximum output are identical, both publish the long-context repricing rule, and both offer Standard, Batch, Flex and Fast. Their request and token limits per minute match tier for tier. GPT-6.1 Sol's knowledge cutoff is a little newer.

The cached-input row

GPT-6.1 Sol bills cached input at 0.05x its input rate; GPT-6 Sol at the usual 0.1x. On a short, uncached request that difference is invisible. On an agentic Codex session that re-sends a large, mostly cached context every turn, cached input can be most of the tokens billed, and this one row decides which model is cheaper for the whole session — always in GPT-6.1 Sol's favor.

The effort floor

GPT-6 Sol accepts none; GPT-6.1 Sol doesn't, and starts at low. If you run latency-sensitive calls with reasoning switched off, GPT-6 Sol keeps that option and GPT-6.1 Sol forces some reasoning — and reasoning tokens bill as output. OpenAI also tells existing GPT-6 Sol users to review its migration guidance before switching.

A gap in the record

GPT-6.1 Sol's rate-limit table has no batch-queue column, so this site shows no batch queue limit for it rather than borrowing GPT-6 Sol's.

How to decide

Cache-heavy agentic work favors GPT-6.1 Sol, and it's the Codex CLI's default. Work that depends on none effort favors GPT-6 Sol. The migration page covers the switch itself.