M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens, and MiniMax sells it through its own coding product as well as its API. That points to a bid to be the default model for high-volume agent work. The "~25x cheaper than Opus and GPT" headline never says which model or which price it is measured against. And the anonymous Space Bunny run shows the model was used heavily while free, which is not the same as proof that it is good.
Summary
MiniMax has released M3.1 Flash, a model built for agent tasks. It is available through the MiniMax API, the M Plan subscription and MiniMax Code. The company reports 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2. It lists prices of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, and says agent tasks cost 70% less than on its previous model. According to our archive, the model is the one that ran anonymously on OpenRouter as "Space Bunny", and MiniMax plans to release the weights later. All benchmark and cost claims come from the company and have not been independently verified.
What happened
MiniMax's announcement presents M3.1 Flash as a model for agent workflows, coding and computer use. The headline scores are 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2. The list prices are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax also says agent-task usage costs 70% less than on the model it replaces. That figure is the company's own, and the supplied material does not show how it was calculated.
The model is available now through three MiniMax channels. The first is the raw API. The second is M Plan, a subscription. The third is MiniMax Code, the company's own coding product. MiniMax says weights will follow later, but the supplied material gives no date and no license.
Our archive records that the model was previously listed on OpenRouter as "Space Bunny". MiniMax says that anonymous listing served about 200 trillion tokens. We could not find the passage with that figure in the supplied excerpt of the post, so it stays a MiniMax claim.
The supplied material includes no evaluation methodology and no third-party replication. The "toe-to-toe with Opus and GPT at ~25x lower cost" framing is the vendor's comparison. It does not name the competing model versions or the prices behind the ratio.
What the archive adds
Our earlier Russian-language note (t.me/qwgai/3988) gave the same scores and prices. It flagged them as MiniMax's own figures rather than an independent comparison. Our archive's canonical entry ties the announcement to three things: the benchmarks, the pricing and the planned release of the weights. It also confirms that M3.1 Flash was previously known as "Space Bunny" on OpenRouter.
The dossier adds two pieces of context. First, according to OpenRouter's listing, Space Bunny Alpha launched on September 23, 2026 as a free, anonymous third-party model with a one-million-token context window and coding and multimodal capabilities. If MiniMax's roughly 200-trillion-token figure is right, that volume built up in about 26 days, from September 23 to the October 19 announcement, on a listing that cost nothing and had no brand on it.
Second, MiniMax's HKEX IPO prospectus says gross margin improved as the company rolled out AI-native products and optimized inference costs, while R&D spending remained very large relative to revenue. Cheaper serving per token therefore matters to MiniMax's finances. The prospectus does not show that this particular launch was driven by revenue.
The dossier also frames the competition. Open-weight labs such as DeepSeek, Qwen, Kimi and GLM compete at the low-price end. The closed labs sell cheap tiers: Claude Haiku, GPT "mini" and Gemini "Flash". The supplied material does not give their current list prices.
Trend signal
The single case documented here runs in three steps. A free, anonymous model appears on an aggregator. It collects a large volume of traffic. About four weeks later, the vendor names it, prices it and promises open weights. We have one example in the evidence, so this is a sequence MiniMax followed, not an established industry method.
The wider pattern we can defend is about where the cheap end of the market competes. The fight is shifting toward agent loops rather than chat. Agent loops resend the same system prompt, tool schemas and repository context on every turn, so token volume is large and repetitive. In that setting, the price per completed task matters more than peak capability.
M3.1 Flash's cache ratio is not unusual. Cached input costs one tenth of normal input, a ratio that is common among major API providers. What is aggressive is the absolute rate. MiniMax states the 70% saving in terms of agent tasks rather than per-token price, which tells you which workload it is targeting.
Why now
1) Usage while the model was free. According to OpenRouter's listing, Space Bunny Alpha launched on September 23, 2026 as a free, anonymous model (d2), and our archive confirms it became M3.1 Flash (d1). MiniMax says it served about 200 trillion tokens, which shows demand at a price of zero but says little about quality. 2) Serving economics. MiniMax's HKEX prospectus links gross-margin gains to inference-cost optimization against heavy R&D spending (d3), so a model that is cheaper to serve matters financially. 3) Agent workloads. The company frames its 70% saving around agent tasks and sells through its own coding product, MiniMax Code (primary). That suggests it is targeting high-volume, cache-heavy agent traffic.
Connections to other events
Our read is that the launch is built to make M3.1 Flash the default cheap model inside coding-agent tools, and MiniMax's own product, not just the API, is part of that plan. Here is the case, split into what is documented and what we infer.
What is documented, with the main caveat first. A free, unbranded model with a one-million-token context window was listed on OpenRouter from September 23 (d2). MiniMax says it served about 200 trillion tokens before the October 19 announcement. Free models attract a lot of traffic simply because they cost nothing, so that volume shows people were willing to route to it at a price of zero. It does not show the model is good. Our inference, which the announcement does not confirm, is that MiniMax gained data on load and failure patterns before setting a price. We have no information on how, or whether, that traffic was used.
The price headline. The "~25x lower cost" claim gives no comparison point. Our own arithmetic against MiniMax's list prices: a uniform 25x gap implies a competitor list price of about $3.75 per million input tokens and $15 per million output tokens. That is a flagship price band, not a cheap-tier one. The comparison that matters to anyone routing models is against Haiku, GPT mini and Gemini Flash. Against those tiers the gap will be some multiple far smaller than 25x, and the supplied material does not let us calculate it. In short, the ratio compares MiniMax's budget model with rivals' premium ones, and only holds if those flagship prices are the comparison point.
Distribution. Shipping through MiniMax Code and M Plan means MiniMax wants to own users directly, not only serve as a cheap backend for other people's tools. Holding back the weights fits that plan, though this is also our inference. The API price sets the reference point first. Open weights can then win self-hosting and data-residency buyers who will not send code to a Chinese API, without undercutting paid traffic on launch day.
Why it matters
Builders of coding agents can test M3.1 Flash for the high-volume tier of a router right away: file search, test triage and first-draft patches, with a pricier model kept for escalation. The cached-input rate only pays off if the stable part of the prompt actually stays at the front and gets cached. Prompt layout is therefore a cost lever, but no more so than with any provider that uses a similar cache ratio.
For buyers, the API puts code on MiniMax infrastructure. Many enterprises will not clear that on data-residency grounds, and for them the real decision waits on weights and a license that do not exist yet.
For Western labs' cheap tiers, the pressure works through the reference price, not a proven quality match. If independent cost-per-solved-task data confirms MiniMax's claims, a premium "mini" price becomes harder to justify. If it doesn't, the 25x framing fades.
What happens next
- MiniMax publishes M3.1 Flash weights with a stated license within 6 months of the October 19, 2026 announcement. (6 months, 55%). Watch for: An official MiniMax M3.1 Flash repository with a license file appears on Hugging Face or a similar hub.
- An independent agentic-coding evaluation reports M3.1 Flash more than 5 points below MiniMax's self-reported 73.7% on DeepSWE v1.1 or its nearest equivalent. (3 months, 45%). Watch for: A third-party leaderboard such as Artificial Analysis, SWE-bench or Terminal-Bench publishes an M3.1 Flash run along with its harness configuration.
- At least one major coding-agent tool features M3.1 Flash as a default or recommended low-cost model. (3 months, 45%). Watch for: M3.1 Flash appears as a default or recommended model in release notes or model settings for Cline, Kilo or Cursor.
Limits and uncertainty
The motive we describe is inferred from the order of events and the channels MiniMax chose. MiniMax has not stated it. The stealth run may simply have been a capacity test.
The 200-trillion-token figure is MiniMax's, and we could not trace it to a passage in the supplied excerpt. The 70% agent-task saving has no published method. The benchmark scores have not been replicated, and agent benchmarks are sensitive to how the harness is configured.
We do not have current list prices for Haiku, GPT mini or Gemini Flash in the evidence. The cheap-tier comparison is therefore stated as a question rather than calculated. A model that is cheap per token but needs more turns or retries can lose its advantage per completed task.
The weight release has no date and no license, so that part of the story may not happen.
Skill to build now
Build a harness that measures cost per resolved task. This week, take 20 closed issues from your own repo that each have a known-good fix and a test that checks it. Run each issue through your agent twice: once with your current cheap model and once with M3.1 Flash. For every run, log input, cached-input and output tokens, turn count, and whether the test passes. Divide total spend by the number of tasks solved. Then move the stable system prompt and tool schemas to the front of the prompt, run the set again, and see how much the cache discount actually changes the bill.
Reader takeaway
Don't route traffic based on the 25x headline. Take 20 to 30 real tickets and run them through your existing agent harness twice: once with your current cheap-tier model and once with M3.1 Flash. Compare cost per resolved ticket, counting cache hits, retries and turns. Move production traffic only if M3.1 Flash wins on that measure, and keep a stronger model as the fallback for the tickets it fails.
Open question
Against Haiku, GPT mini and Gemini Flash rather than flagship models, is M3.1 Flash still cheaper per completed task, or does the advantage disappear once retries and turn counts are included? And will the traffic from the free Space Bunny period stay now that the model has a price and a Chinese vendor's name on it?