MiniMax M3.1 Flash Is Space Bunny: What Changed from M2.7 and M3, and What It Costs

In late September, a model with no maker's name on it topped OpenRouter's daily chart, ahead of DeepSeek V4.1 Flash, and stayed OpenCode's most-used model for five days straight. It was listed as Space Bunny Alpha, and it was free. That model was MiniMax M3.1 Flash. As of 19 October 2026 you can call it per token on AIHubMix as minimax-m3.1-flash, at $0.165 per million input tokens and $0.66 per million output tokens.

The short version:

• It replaces M2.7. Starting 19 October, MiniMax sends M Plan and Token Plan requests for M2.5 and M2.7 to M3.1 Flash. Both older families retire on 18 November.
• It costs about 43% less than M3 per token on AIHubMix. MiniMax hasn't published a per-token price of its own yet.
• Thinking is always on, and the default is the deepest level. On a small bug fix, the default used about 7× the output tokens of the low setting and took about 6× as long. Both settings passed every test.

From Space Bunny to a model ID

The dates and chart positions come from CellCog's Space Bunny timeline, which compiles OpenRouter and OpenCode posts. During the stealth week, the community identified the model from its tokenizer before MiniMax said anything. cheaty's tokenizer comparison on X found that Space Bunny's text tokenizer "perfectly matches that of the MiniMax M3". Other testers matched 36 of 36 and 50 of 50 probe strings. Space Bunny's published specs were also identical to what MiniMax later documented: 1M context, text, image, and video input, and thinking that can't be turned off. One early user summed it up as "insanely fast, but it needs clear and specific instructions, otherwise it's too lazy." Keep that in mind when you write prompts for it.

MiniMax's docs have since dropped the "Preview" label. The model list now calls it MiniMax-M3.1-Flash, a "frontier multimodal coding model with 1M context window and tunable thinking depth." Every response from AIHubMix also reports that upstream model name.

What changed from M2.7

For anyone on a MiniMax plan, M2.7 is the model M3.1 Flash replaced overnight. Here are the differences that matter in practice:

Video input is in MiniMax's docs. AIHubMix currently lists text and image input for this model. Base64 images and plain image URLs both worked in our tests.

A few things stayed the same. Both models think on every request. Both speak MiniMax's Anthropic-compatible and OpenAI-compatible APIs. Both only accept tool_choice: auto according to MiniMax's API reference. If your M2.7 code never touched thinking settings or caching, it will run on M3.1 Flash. It just runs differently, and the thinking level explains most of that.

The new control is effort. MiniMax's text generation guide lists low, medium, high, xhigh, and max, and says the model uses max when the field is omitted. M2.7 had nothing to set. So when your request includes no effort setting, M3.1 Flash thinks as hard as it can.

How close it gets to M3, and where M3 still wins

M3 is the tier above: MiniMax's open-weight flagship, released 1 June with the same 1M context. On AIHubMix, M3's output price is 1.75× Flash's ($1.152 vs $0.66). That's a real gap, but not a different weight class.

The honest answer to "how close is Flash to M3" is that nobody has published numbers for Flash yet. MiniMax released no benchmark table for M3.1 Flash, and Artificial Analysis has no entry for it. M3 has both. MiniMax reported 59.0% on SWE-Bench Pro (self-reported), and Artificial Analysis scores M3 at 29.2 on its Intelligence Index. Any M3.1 Flash benchmark table you see is either M3's or unsourced.

M3 still clearly wins in four places:

• Evidence. M3 has independent scores; Flash has anecdotes and our own tests.
• Open weights. M3's weights are on Hugging Face under the MiniMax Community License. Flash has none yet; MiniMax has said it plans to release them, with no date.
• Thinking off. M3 can skip thinking entirely. Flash can't, so latency-critical single-shot calls are harder to make fast.
• Priority tier. MiniMax sells a faster Priority tier for M3 at 1.5× price. Flash rejects the service_tier field.

Flash wins on price per token and on the effort dial, which lets one model cover both quick fixes and long reasoning.

Price and real cost

Prices are per million tokens, checked 19 October 2026. MiniMax's figures come from its pay-as-you-go pricing page. M3's official price includes what MiniMax labels a "Permanent 50% off".

M3's official price doubles above 512K input tokens. AIHubMix lists one flat price for M3.1 Flash.

On AIHubMix, Flash's per-token price is about 44% below M2.7 ($0.165 ÷ $0.2958 ≈ 0.56) and about 43% below M3 ($0.165 ÷ $0.288 ≈ 0.57). Output prices follow the same ratio.

Per-token price isn't the whole bill, though. With always-on thinking, the effort level decides how many tokens a task takes. To put a number on it, AIHubMix ran one realistic task on 19 October. All figures below were measured by AIHubMix; nobody else has reproduced them yet.

The task: fix a billing function with two bugs. A pause that runs past the end of the period double-counts one day, and overlapping pauses get subtracted twice. The model returns the corrected function, and a local harness runs 215 hidden test cases. Each setting ran three times through AIHubMix's Chat Completions endpoint.

Here is the arithmetic for the default row. A request used 464 prompt tokens, about 192 of them served from cache, plus 6,673 output tokens. That comes to (272 × $0.165 + 192 × $0.0165 + 6,673 × $0.66) ÷ 1,000,000 ≈ $0.0045. Across 1,000 fixes like this, that's about $4.45 at the default and $0.64 at low, with the same test results.

Three runs per setting is a small sample, and run-to-run variance was large: the default ranged from 4,589 to 9,680 output tokens. Still, the direction is clear. If you migrate from M2.7 without setting an effort level, the lower per-token price can disappear into thinking tokens on harder tasks. The effort post in this series (linked at the end) has the full five-level ladder.

Where it isn't the best choice

• Interactive work at the default setting. A small single-function fix took 39 to 82 seconds at the default level. If a person is waiting, set low or medium, or use a faster model. Claude Haiku 5.5 finished the same fix in about 7 seconds in our tests.
• Anything that needs a published score. If a procurement review or an SLA needs an independent benchmark, M3 or DeepSeek V4.1 Flash have them and Flash doesn't yet.
• Calls that must run without thinking. Flash can't skip thinking. On short classification prompts, adaptive thinking spent almost nothing (0 to 6 tokens), but you can't guarantee zero.
• Self-hosting. Without published weights, there's nothing to deploy.

Who should switch now, who should wait, who should stay

• M Plan and Token Plan users on M2.5 or M2.7: you've already switched, because MiniMax reroutes those model names today. Re-run your evals and add an effort setting to your requests.
• Pay-as-you-go users on M2.5 or M2.7: you have until 18 November. Moving to M3.1 Flash on AIHubMix means changing the model ID and adding reasoning_effort. The migration post in this series covers the errors you'll hit.
• Teams already on M3 for coding agents: try Flash at low or medium on a slice of traffic. Its per-token price is 43% lower, and on our task every setting passed.
• Stay on M3 if you need open weights, thinking off, independent benchmarks, or the Priority tier.

The fastest check is to send your own prompt to MiniMax M3.1 Flash on AIHubMix and to M3 with the same key. The side-by-side page shows both price cards. The AIHubMix quick start covers keys and base URLs:

import os
from openai import OpenAI

client = OpenAI(base_url="https://aihubmix.com/v1", api_key=os.environ["AIHUBMIX_API_KEY"])

resp = client.chat.completions.create(
    model="minimax-m3.1-flash",
    reasoning_effort="low",  # omit it and the model runs at max
    messages=[{"role": "user", "content": "Find the bug: def avg(xs): return sum(xs) / len(xs) - 1"}],
)
print(resp.choices[0].message.content)
print(resp.usage)

FAQ

Is MiniMax M3.1 Flash the same model as Space Bunny Alpha?
Yes. Space Bunny Alpha was the anonymous listing of M3.1 Flash on OpenRouter and OpenCode from 23 September 2026. Community tokenizer tests matched it to MiniMax before MiniMax named the model on 27 September.

What is the model ID on AIHubMix?
minimax-m3.1-flash. MiniMax's own docs use MiniMax-M3.1-Flash, and earlier material used MiniMax-M3.1-Flash-Preview.

How much does MiniMax M3.1 Flash cost?
On AIHubMix it costs $0.165 per million input tokens, $0.66 per million output tokens, and $0.0165 per million cached input tokens. MiniMax had not published its own per-token price as of 19 October 2026.

Can I turn thinking off?
No. The model always thinks. The lowest setting is effort low, which on short prompts often spends very few thinking tokens.

What happens to my M2.7 requests?
On MiniMax's M Plan and Token Plan, requests that name M2.5 or M2.7 go to M3.1 Flash from 19 October 2026. On MiniMax pay-as-you-go, existing users can call M2.5 and M2.7 until 18 November 2026, after which those requests stop working.

Is M3.1 Flash better than M3?
There's no published benchmark to say so. Flash is cheaper per token and has effort levels. M3 has open weights, independent scores, a Priority tier, and can run without thinking.

Does it support images and video?
MiniMax documents text, image, and video input. AIHubMix currently lists text and image input for this model. Both base64 images and image URLs worked in our tests.

Keep reading: the MiniMax M3.1 Flash series

• If you're moving code off M2.5 or M2.7 before the 18 November retirement, this post lists every error and silent change we hit: Moving Off MiniMax M2.5/M2.7 Before Nov 18: Fixing M3.1 Flash's "requires adaptive thinking" 400
• The default effort level drove most of the cost above. This post measures all five levels on the same bug fix: MiniMax M3.1 Flash Effort Levels: Why the Default Cost 7× Low on the Same Bug Fix
• To see how Flash compares with the two other cheap coding models released the same month, on the same task: MiniMax M3.1 Flash vs DeepSeek V4.1 Flash vs Claude Haiku 5.5: One Bug Fix, Three Bills

Sources

• MiniMax: Models
• MiniMax: Model Invocation (text generation guide)
• AIHubMix: MiniMax M3.1 Flash
• CellCog: Space Bunny Alpha, the Stealth Model and the MiniMax Clues
• Developers Digest: MiniMax M2.5 and M2.7 Retirement