MiniMax M3.1 Flash: $0.15 Agent Model Live Now, Weights Reportedly Due Oct 23

MiniMax has released M3.1 Flash, a low-cost agent and coding model priced at $0.15 per million input tokens and $0.60 per million output tokens, available now on its API, M Plan and MiniMax Code. The weights are not out yet: MiniMax's official blog says only "later," while social posts relayed by press coverage give October 23. The model is the public version of "Space Bunny," the anonymous model that topped OpenRouter's usage charts, by MiniMax's account.

For builders the question is simple: does a $0.15 model reach near-frontier agent performance? On MiniMax's own numbers it gets close on several tasks and trails on others, and the license, size and context length are not yet disclosed. This post separates the confirmed from the pending. For the closest rival, see our write-up of DeepSeek V4.1 Flash-1017.

Update (October 19, 2026): Artificial Analysis scores and a 1M context

The Decoder, citing Artificial Analysis, reports that M3.1 Flash scores 42 on the Artificial Analysis Intelligence Index, up 13 points from MiniMax M3's 29. That ties GLM-5.3-Flash, edges Gemini 3.8 Flash (41) and trails DeepSeek V4.1 Flash-1017 (45). Reported cost per task is about $0.22 against $0.26 for DeepSeek, because the model used roughly 189 million output tokens across the benchmark versus 224 million.

The same report gives 48 percent on Humanity's Last Exam (DeepSeek 43), 22 percent on CritPt (DeepSeek 24), and 56 percent on AA-Omniscience accuracy. The weak spot: a 93 percent hallucination rate, because the model tries to answer instead of declining when unsure. It also lists a context window of up to one million tokens and text plus image input, which settles one item left open below. These are third-party figures reported by the outlet; the open-weights date of October 23 is still MiniMax's plan, not yet a release.

Quick facts

What was Space Bunny?

In recent weeks an anonymous model named Space Bunny ran on OpenRouter. MiniMax now says it processed more than 200 trillion tokens and was an earlier version of M3.1 Flash. This is the same stealth-testing pattern we covered with the Ox Alpha stealth model: a lab lets developers try a model blind, collects usage, then reveals the name. Press reports cited a lower 56.7 trillion token figure and a ten-day run at number one on the platform; MiniMax's own figure is the one above, and we have not reconciled the difference. Usage is not quality, but it does show that developers kept choosing it. If you want background on routing platforms, read our OpenRouter enterprise guide.

The official benchmark table

The table below is based on the one in MiniMax's announcement. We read the numbers directly from it; footnotes list harnesses such as Claude Code, Codex and mini-SWE-agent versions, and say that some scores come from Artificial Analysis or public leaderboards. A selection of rows:

Two readings stand out. First, it leads or ties on workflow-style agent benchmarks (Automation-Bench, BankerToolBench) and sits within a point or two of the frontier models on DeepSWE. Second, it falls behind on harder terminal and repo-level tasks (Terminal Bench 4.0, NL2Repo) and on knowledge-heavy HLE. MiniMax's page itself describes the OSWorld and ALE scores as first among Chinese models, not first overall.

Cost per task, not just per token

MiniMax also published a chart of DeepSWE v1.1 score against average cost per task on a log scale. M3.1 Flash lands around 68.1% at the high setting, 72.1% at xhigh and 73.7% at max, with a cost per task roughly a quarter of a dollar, based on the chart's axis. DeepSeek V4.1 Flash-1017 sits higher and a little cheaper on the same chart, reaching 77.0% at max. Per-task cost matters more than per-token price for agents, because a cheap model that takes more steps can erase its advantage. The chart uses public leaderboard data as of September 22, 2026 with the mini-SWE-agent harness, so it is a snapshot, not a live ranking.

How it compares with the cheap-model field

The Flash tier is getting crowded. Our small models comparison puts Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash, Gemini 3.7 Flash and GLM 5.3 Flash side by side. M3.1 Flash adds a model that is cheap, fast (200+ tokens per second) and, if the weights arrive, self-hostable. MiniMax has been moving in this direction for months, from its open video model to the MiniMax Code CLI, and Asian models were already taking a large share of routing traffic, as we noted in our OpenRouter token share report.

Which model for which job

This table is explainx.ai's reading of MiniMax's own numbers above, not a measured test.

How to read a launch table like this one

Vendor tables are useful but not neutral, and this one shows why. The footnotes disclose that different benchmarks used different harnesses, from Claude Code to Codex to mini-SWE-agent, with context windows between 200K and 1M tokens. A model can gain or lose several points just from the harness. The Agents' Last Exam row notes that scores from different subsets are not directly comparable, so blanks mean no official full-set result, not zero. And some competitor scores are "the highest reported at release," while one competitor's OSWorld score is an in-house reproduction.

Use three habits when you read launch tables. First, compare rows where the footnote names the same harness for all models. Second, favor benchmarks that resemble your work: a coding agent team should care about DeepSWE and Terminal Bench more than HLE. Third, check the cost axis: a model that matches a rival's score at one-fifth the cost per task is a different proposition from one that only matches on price per token. MiniMax's own cost chart supports that view, and it is also the first thing an independent tester can reproduce.

Timeline

• Earlier in October 2026: Space Bunny appears anonymously on OpenRouter and builds usage.
• October 17: DeepSeek's V4.1 Flash-1017 arrives with open weights, the benchmark MiniMax compares against most closely.
• October 19: MiniMax publishes the M3.1 Flash blog, lists API pricing, and says the Pro model and weights are coming.
• October 23: The date given for weights in the reported post. Not yet reached.

What is not known yet

• License. Open weights can mean MIT, Apache 2.0 or a restrictive community license. No source we found states it.
• Size and architecture. One outlet describes a 428B mixture-of-experts model (possibly the parent MiniMax M3, which MiniMax lists with up to 1M tokens of context), but MiniMax's page gives no parameter count, so treat that as unconfirmed.
• Context length. Not stated on the official page.
• The date. Oct 23 comes from social posts relayed by outlets, not from MiniMax's blog; the blog says no date. One third-party site also notes it found no formal release entry in MiniMax's release notes. Watch MiniMax's Hugging Face page.
• Independent tests. Every number above is MiniMax's. Run your own evals.

How to try it today

1. Create an API key with MiniMax and select the M3.1 Flash model name from its docs, or use the model through the M Plan or MiniMax Code.
2. Run your hardest existing agent task twice, once on your current model and once on M3.1 Flash, and compare pass rate, steps and total cost.
3. Use the cached-input price ($0.015 per million) by keeping system prompts and tool definitions stable.
4. If you plan to self-host, wait for the weights and license before committing infrastructure.
5. Keep a fallback model for terminal and repo-scale tasks, where the table shows the biggest gaps.

What this means for what you build or pay

At $0.15 and $0.60, a Flash model can handle high-volume steps such as tool routing, form filling, log triage and first-pass code edits at a small fraction of frontier prices, with a stronger model reserved for the hard final step. The risk is the usual one with vendor tables: harness choices, judge models and cherry-picked comparisons can move numbers by several points. A day-one release also means no community reports yet on tool-call reliability or long-context behavior.

Related reading

• DeepSeek V4.1 Flash-1017: open weights, MIT, benchmarks and pricing
• Small models compared: Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash and more
• OpenRouter Ox Alpha stealth model
• What is OpenRouter: enterprise guide
• MiniMax Code CLI
• MiniMax H3 open video model
• Asian models and OpenRouter token share

Details are accurate as of October 19, 2026 and based on MiniMax's announcement and press coverage; the weights release and license may change the picture.