MiniMax M3.1 Flash costs $0.15/M input tokens and turns out to be the stealth "Space Bunny" model. Here's the pricing, benchmarks, and how it compares to Claude Haiku 5.5.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
MiniMax M3.1 Flash is a new 428-billion-parameter mixture-of-experts model out of China, priced at $0.15 per million input tokens and $0.60 per million output tokens, roughly a third of what its predecessor M3 cost. The twist: MiniMax confirmed on October 19, 2026 that the model had already been running in the wild for weeks under the codename "Space Bunny," quietly topping OpenRouter and OpenCode leaderboards before anyone knew who built it.
If you've been anywhere near an LLM leaderboard this month, you've probably seen the Space Bunny memes. Turns out the joke was on all of us. It wasn't a scrappy open-source lab or a mystery startup. It was MiniMax testing its own flagship Flash model in stealth mode, and it worked well enough that people were recommending it without knowing who made it.
Space Bunny showed up anonymously on OpenRouter in early October and immediately started eating tokens. According to OpenRouter's own numbers, it was the #1 model on the platform for 10 straight days, processing 56.7 trillion tokens, the most of any stealth model the platform has ever run. It also led OpenCode for 12 days and reportedly crossed 200 trillion tokens total across both venues before "hopping off" on October 5.
Then on October 19, MiniMax's official account posted the reveal: "Welcome back. Space Bunny. MiniMax-M3.1-Flash is coming." A few hours later came the full announcement, framed with the tagline "Space Performance + Bunny Speed," confirming the stealth model and the shipped product were the same thing.
It's a smart marketing move, honestly. Instead of publishing benchmark slides nobody trusts anymore, MiniMax let the model earn trust anonymously first. A model that wins a blind leaderboard fight for ten days is a stronger pitch than any chart with your own logo on it.
Here's the number everyone's fixating on, and for good reason:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Context |
|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.60 | 1M |
| DeepSeek V4.1 Flash-1017 | ~$0.15 (off-peak) | ~$0.60 (off-peak) | 1M+ |
| Claude Haiku 5.5 (≤100K) | $0.10 | $0.50 | 200K |
| Claude Opus 5 | $12.00 | $12.00 (blended ~$24 combined) | 200K |
Artificial Analysis puts the real-world cost at about $0.22 per task on its Intelligence Index, and a workload of 1M input plus 1M output tokens runs you roughly $0.75 total. That undercuts Claude Opus 5 by close to 30x on paper. It's not quite as cheap as Claude Haiku 5.5's headline input rate, but Haiku's pricing jumps 5x past 100K tokens, and MiniMax holds its rate across the full 1M context window.
This is the same price war we covered with DeepSeek V4.1 Flash a few weeks back, and it isn't slowing down. Every Chinese lab releasing a "Flash" model this year has been racing the others to the bottom on price, and MiniMax just set a new floor, again.
MiniMax says M3.1 Flash runs nearly 3x faster than M3 and cuts agentic task costs by 70%. The benchmark numbers back up the speed claim at least:
So it's not dominating any single benchmark. It's clustering near the top of the cheap-model tier, which is exactly where MiniMax wanted to land.
There's one number worth flagging before you get excited, though: a 93% hallucination rate on Artificial Analysis's AA-Omniscience test. That sounds disqualifying until you read what it measures. The model answers 56% of factual questions correctly, which beats DeepSeek V4.1 Flash's 43% and GLM-5.3 Flash's 28% on the same test. The catch is that M3.1 Flash almost never says "I don't know" — it guesses on nearly everything, and when it's wrong, it's confidently wrong. For a coding agent that's mostly checking terminal output and test results, that's survivable. For anything where you're trusting the model's factual claims unsupervised, it's a real problem.
Here's where I'd slow down before switching your agent pipeline over. The Hacker News thread on the launch is a lot less impressed than MiniMax's own slides. Several commenters who tried Space Bunny before the reveal called it underwhelming for real coding work, needing a lot of hand-holding to produce usable output. One person flatly said benchmark scores are "effectively meaningless" given how labs choose which comparisons to publish, and pointed out MiniMax's marketing leans on older models like Opus 5 and GPT-5.6 Sol rather than current frontier releases.
That said, it's not all negative. One developer reported it matched DeepSeek V4 Pro on a specialized programming benchmark, and another said it tied Opus 5.5 on data analytics work at a third of the price, though it burned through noticeably more tokens getting there. Token-hungry but cheap enough that it doesn't matter is a trade I'd actually take for a lot of background agent work.
My take: this isn't a model I'd hand my main coding agent to today. It's a model I'd absolutely wire into a cheap first-pass agent, a bulk code-review pass, or anything high-volume and latency-sensitive where an occasional wrong answer gets caught downstream anyway. That's basically the exact use case MiniMax built it for, and it's a sensible one.
These two models are aimed at the same spot in the market: cheap, fast, good enough for agentic grunt work. Here's how they actually stack up.
| MiniMax M3.1 Flash | Claude Haiku 5.5 | |
|---|---|---|
| Input price | $0.15 / 1M | $0.10 / 1M (≤100K context) |
| Output price | $0.60 / 1M | $0.50 / 1M (≤100K context) |
| Context window | 1,000,000 tokens | 200,000 tokens |
| Terminal-Bench 4.0 | not published | 39.2% |
| Open weights | Due October 23, 2026 | Closed |
| Known weakness | High hallucination rate on factual recall | Weaker than Sonnet/Opus on complex agentic coding |
If your job needs a huge context window and you're running high-volume, low-stakes tasks, M3.1 Flash's price-per-token-at-scale is hard to beat, and it's getting open weights in a few days, which Haiku never will. If you're inside the Anthropic ecosystem already and need something that plays well with Claude Haiku 5.5's subagent patterns and tool use, Haiku is still the safer default for anything customer-facing. I wouldn't put M3.1 Flash in front of an end user without a review step. I would put it behind one.
Three ways as of this week:
It was also free on OpenRouter for 72 hours after the reveal, so if you're reading this within a few days of October 19, check whether that window is still open. Open weights are scheduled for October 23, 2026, four days after the hosted launch, so if you want to self-host it, that's the date to watch.
This whole saga is part of a bigger pattern we've been tracking across Chinese labs all year, the same dynamic we saw play out with Kimi K3 going after Claude on frontend coding. Nobody's winning the price war outright. DeepSeek, GLM, Qwen, and now MiniMax are all undercutting each other by a few cents every few weeks, and the models keep getting good enough that it's starting to matter for real workloads, not just demos.
It's a 428-billion-parameter mixture-of-experts model from MiniMax, with about 23 billion active parameters per token, a 1-million-token context window, and multimodal (text and image) input. It's built for agentic coding and high-volume, latency-sensitive tasks.
$0.15 per million input tokens and $0.60 per million output tokens through the MiniMax API. Artificial Analysis estimates roughly $0.22 per benchmark task and about $0.75 for a 1M-input/1M-output workload.
Yes. MiniMax confirmed on October 19, 2026 that the anonymous "Space Bunny" model, which led OpenRouter for 10 days and OpenCode for 12, processing over 200 trillion tokens combined, was an earlier internal build of M3.1 Flash.
It's decent, not dominant. It scores 73.7 on DeepSWE v1.1 and 62.2 on Automation-Bench, both competitive with DeepSeek V4.1 Flash. But Hacker News commenters who tested it as Space Bunny reported it needed more hand-holding than DeepSeek or GLM for real coding tasks, despite similar benchmark scores.
M3.1 Flash has a bigger context window (1M vs 200K) and holds its price flat at scale, where Haiku 5.5's rate jumps 5x past 100K tokens. Haiku 5.5 is still the more reliable pick for complex or customer-facing agentic coding, since M3.1 Flash carries a notably higher hallucination rate.
MiniMax has set October 23, 2026 as the open-weights release date, four days after the hosted API went live.
Artificial Analysis's AA-Omniscience benchmark measures how often a model gives a wrong answer instead of admitting uncertainty. M3.1 Flash answers almost every question rather than declining, which pushes both its correct-answer rate (56%) and its hallucination rate on wrong answers unusually high at the same time.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Mistral's Timestral forecasts time series on a CPU in 10-140ms, no GPU. How to install and run it in Python (5M-209M models), plus the license's $20M catch.
Mercury 3 claims 1,068 tokens/sec via diffusion, but real-world OpenRouter tests show ~160 tok/s and it loses to Claude Haiku 5.5 on DeepSWE coding (46 vs 52).
HybridAgent lets LLM agents think in latent vectors and act in text, cutting tokens up to 54.7% and cost up to 3.7x. Here's what it means for agent builders.
Comments