MiniMax M3.1 Flash is out — benchmarks, speed and API price

MiniMax released M3.1 Flash on October 19, 2026 — a fast, inexpensive model built for agents and coding. The API price is $0.15 per million input tokens and $0.60 per million output tokens, and cached input costs $0.015. MiniMax also confirmed that Space Bunny, the anonymous model that processed more than 200 trillion tokens, was an early version of M3.1 Flash.

What MiniMax announced#

• Agents and coding first. MiniMax calls M3.1 Flash a generational step up in agentic tasks, coding, computer use and cybersecurity.
• Cheaper and faster than its predecessor. By MiniMax's numbers, agentic tasks cost about 70% less than on the previous model, and inference is nearly three times faster. Output speed reaches 200+ tokens per second, with a median above 170.
• Where it runs. The model is available in the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Weights come later — MiniMax's post on X names October 23.
• More to come. A Pro version is announced, without a date.

Benchmarks as reported by MiniMax#

These are the developer's own results, not independent tests:

MiniMax says M3.1 Flash ranks first among Chinese models on OSWorld V2 and ALE. Until independent leaderboards add it, read these as claims and test the model on your own tasks.

Price compared with earlier MiniMax models#

Prices per 1M tokens: MiniMax's checked on October 20, 2026, CheapTokens' are current. CheapTokens has no separate cached-input price.

Per token, M3.1 Flash costs half as much as M3. The 70% figure is about whole agent tasks: MiniMax claims the new model also needs fewer tokens to finish them. That second part is worth checking on your own workload.

What it means if you build agents#

Long agent sessions resend the same history on every step, so the cached-input price matters as much as the headline one. At $0.015 per million, a well-cached M3.1 Flash session can cost much less than its $0.15 list price suggests.

Speed matters too. At 170–200 tokens per second, a 30-step coding session spends less time waiting on the model.

Two caveats before switching. The benchmarks come from MiniMax, and independent checks will follow once the weights are out. And since only MiniMax serves the model for now, you depend on a single provider.

Where CheapTokens fits#

On CheapTokens you can run MiniMax-M2.7 — an earlier, separate MiniMax model with a 200K context and up to 8,192 output tokens per answer. The M3.1 Flash results above are MiniMax's figures for the new model and don't carry over to M2.7, so judge it on your own tasks.

Where M2.7 already does the job — tool-calling agent loops, multi-step automation, summarizing a long history that fits in 200K — it costs $0.05 per 1M input tokens and $0.15 per 1M output tokens, with no peak hours. The API is OpenAI-compatible, and one prepaid balance also covers DeepSeek and GLM when a step needs a different model. Here's a 30-step agent task (1.05M input and 15K output tokens) on each model at current prices, next to OpenRouter's reference prices:

Compare what your own workload would cost:

Current prices: