If you have been routing coding or tool-use traffic through DeepSeek Flash tiers, Claude, or another cheap “fast” model, Shanghai’s MiniMax just gave you a new line item to price. On October 19, 2026, the lab formally launched MiniMax M3.1 Flash — and confirmed that the anonymous OpenRouter-era model known as Space Bunny was an early version of the same stack. MiniMax says that stealth run processed more than 200 trillion tokens before the unmasking.
The product pitch is not a general chat demo. MiniMax positions M3.1 Flash for agentic workflows, software engineering, computer use, and even cybersecurity evals, with online inference it claims is roughly 3× faster than its predecessor and ~70% cheaper on agentic usage cost. Those speed and cost percentages are company claims; the checkable part for overseas builders is the published API table and whether the model fits your stack today.
What you can use today
API access is live on MiniMax’s international platform at platform.minimax.io, with Chinese billing and docs at platform.minimax.cn. Flat pay-as-you-go rates for MiniMax-M3.1-Flash (per 1M tokens):
- $0.15 input · $0.60 output · $0.015 cached input (USD, international docs)
- ¥1.00 input · ¥4.00 output · ¥0.10 cache read (CNY, Chinese pricing overview)
At those USD rates, one million uncached input tokens plus one million output tokens lands at $0.75 in model charges — before your own orchestration costs. MiniMax also ships a desktop MiniMax Code app (agent.minimax.io/download) and an M Plan subscription path alongside raw API keys. Platform docs describe a 1M-token context window, multimodal chat inputs (text/image/video), OpenAI- and Anthropic-compatible endpoints, and an adjustable reasoning effort knob (low through max).
MiniMax says everyday output can exceed 200 tokens/second, with median (P50) speeds consistently above 170 TPS, attributed partly to more aggressive sparsity. Treat throughput as workload-dependent until you measure it on your prompts and region.
Illustration: the overseas builder angle — a cheap, fast agent/coding API you can meter today, not a dark cyber-infographic.
Company benchmarks vs what changes the bill
MiniMax’s launch post puts M3.1 Flash near or among top reported scores on several agent-leaning suites: AutomationBench-AA / Automation-Bench for multi-app enterprise workflows; 76.2 on BankerToolBench; 73.7% on DeepSWE v1.1 for software tasks; 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam (ALE) for computer use (first among Chinese models on those two, per MiniMax); and about 82% on CyberGym. In the company’s framing, those results sit close to named DeepSeek Flash and Claude Opus 5 comparison points on selected charts — again, MiniMax’s tables, not a joint bake-off you can replay from a public leaderboard alone.
Independent evaluator Artificial Analysis, as reported by VentureBeat, scored M3.1 Flash 42 on its Intelligence Index (up from 29 for prior MiniMax M3), with average task cost falling from roughly $0.51 to $0.22. That places it near Z.ai’s GLM-5.3 Flash and behind DeepSeek V4.1 Flash-1017’s reported 45 on the same index — useful context if you care about third-party cost-normalized intelligence, not only vendor slides.
The caveat: confident wrongness
VentureBeat’s reporting also flags a reliability issue that matters for unsupervised agents. On Artificial Analysis’s AA-Omniscience factual benchmark, M3.1 Flash answered more than half of questions correctly (~56%), but when it was not correct it overwhelmingly produced wrong answers instead of abstaining — a ~93% “hallucination rate” under that benchmark’s definition of non-correct responses. That is not “93% of all answers are false”; it is a measure of how the model fails. The resulting overall Omniscience score was weak (~16) despite decent raw recall.
For coding agents with tests and retrieval, that trade-off may be manageable. For finance, CRM writes, or research agents that update records without a human in the loop, you will want grounding, refusal policies, or review — MiniMax’s own agentic scores do not erase that independent finding.
Call the API now via platform.minimax.io (or platform.minimax.cn if you bill in CNY). Point OpenAI- or Anthropic-compatible clients at MiniMax’s documented base URLs with model id MiniMax-M3.1-Flash. Try MiniMax Code if you want a desktop coding agent shell. Plan for open weights on October 23, 2026 — MiniMax has said weights will ship then, but parameter count and license terms are not public yet, so do not assume commercial self-host rights until the package drops. A higher-capability Pro tier is teased without a date.
How to read the launch
Chinese Flash models are racing on the same axes: tokens per dollar, tokens per second, and whether agents can finish multi-step work without a frontier Opus bill. M3.1 Flash’s published $0.15 / $0.60 / $0.015 table puts it in the same cheap band as DeepSeek’s off-peak Flash pricing in recent third-party comparisons, with MiniMax arguing it wins on selected automation and coding charts and on hosted speed.
The honest overseas takeaway is narrower. Use it if your workload is tool-heavy coding or computer-use where you already verify outputs. Re-check the October 23 weight drop before you design a self-host path. And budget for guardrails if factual abstention matters more than raw Automation-Bench bragging rights — that is where VentureBeat’s Artificial Analysis numbers still bite.