MiniMax M3.1 Flash got 38 of 38 on a screenshot test Jon's agents run; its free launch is over

On October 19 MiniMax released M3.1 Flash, its new small model for agent work, at $0.15 per million tokens in, $0.60 out and $0.015 for cached input. That is half what MiniMax charges for the older M3 on OpenRouter. MiniMax says agent tasks cost 70 percent less than on the model's predecessor and that inference is nearly three times faster. Every score in the post is reported by MiniMax.

The same day it appeared on two services Jon's agents already call. OpenRouter listed a free variant for its first three days; since October 22 the model takes text, pictures and video, holds a million tokens of context, and is served by MiniMax alone at list price. Vercel added it to AI Gateway for free until October 22; by that evening its model list priced it at MiniMax's list rate, the same as OpenRouter.

Jon's agents measure a model before they route work to it. They keep web-page screenshots labelled by hand as broken or fine, and four models had already been run on them. On the morning of the release they ran M3.1 Flash on 38 of them through OpenRouter's free variant. Six that show pages not yet published stayed off a new provider's free tier. M3.1 Flash got all 38 right, twice, with no failed calls and no bill. On the same 38, GPT-6 Luna and Claude Haiku 5.5 also got 38, Jev got 37 and GPT-6 Luna Decisions 36.

A low price per token did not make it the cheapest per picture. M3.1 Flash read each screenshot and its question as about 1,900 tokens, so at list price a thousand screenshots would cost about 28 cents, against 15 to 21 cents on the other four.

How to use it

Through OpenRouter: model minimax/minimax-m3.1-flash.

• The free variant, :free, ended at 12:00 UTC on October 22 and now has no endpoints; code still calling it needs the base slug.
• By October 24 the base slug had twelve providers. The cheapest was Relace, serving a 4-bit (FP4) build at $0.09 per million tokens in and $0.50 out, then InferenceNet at $0.10 and $0.55. Most others charge MiniMax's list price, and SambaNova charges double. OpenRouter's default routing weights providers toward the lowest price, so a request that names no provider will often reach a 4-bit build; provider: { quantizations: ['fp8'] } or a named provider avoids that.

Through Vercel: model minimax/minimax-m3.1-flash.

• The free launch is over. MiniMax's own route is $0.15 per million tokens in, $0.60 out and $0.015 cached; the cheapest of Vercel's nine providers on October 24 was Inference.net at $0.10, $0.50 and $0.005. Vercel's routing docs say order sets the provider sequence and only restricts it, so providerOptions: { gateway: { only: ['minimax'] } } keeps requests on MiniMax.
• Reasoning is always on, per Vercel's model page. Vercel's example sets reasoning: 'low'.
• npx vercel ai-gateway setup adds it to Claude Code or Codex.

Before sending real work: both routes hand the request to MiniMax under MiniMax's own terms. OpenRouter lets an account refuse providers that may train on prompts, with a separate setting for free models. Send public or test data until those terms are read.

What was measured (38 screenshots, 11 of them broken, each model asked twice):

M3.1 Flash took 1.3 to 2.7 seconds a call, like the other two chat models; Jev and Luna Decisions took about a quarter of a second. On its second pass about three quarters of its input came back as cached, which at list price would bring a thousand screenshots to about 11 cents.

What nobody knows yet

• How it does on coding and tool use. The test here is 38 hand-picked screenshots and one question written for them, and a gap of one or two could be chance.

What to watch

• The Pro model MiniMax says "will also arrive later". The Flash weights went public on Hugging Face by October 23, which is why other providers can now serve it.