MiniMax slips a new coding model into its agent tool without a price tag

MiniMax launched M3.1-Flash-Preview inside MiniMax Code on September 27, 2026, skipping the usual model card, benchmark report and public API listing entirely. The release lands days after developers fingerprinted a free, anonymous OpenRouter model called Space Bunny Alpha and concluded it's an early build of the same system.

You won't find M3.1-Flash-Preview on OpenRouter. You won't find a price for it anywhere. MiniMax just turned it on inside MiniMax Code, its own coding agent product, and let users start running it. According to MiniMax's own announcement on X from its MiniMax_Agent account, the model "debuts today on MiniMax Code" and is "built for everyday development, fast, reliable, and ready for real work, from quick bug fixes to full features." That's the entire announcement. No model card, no benchmark numbers, no API endpoint developers can call directly.

What is public: M3.1-Flash-Preview supports a context window up to 1 million tokens, and MiniMax Code now exposes five reasoning-effort levels for it: low, medium, high, xhigh, and a new tier called max. Higher effort levels mean the model thinks longer and produces more output tokens at greater latency, the same trade-off every reasoning model makes, but the tier list itself, five distinct steps instead of the usual three, suggests MiniMax is tuning this specifically for developers who want to dial cost and speed against task difficulty on the fly.

Four days before M3.1-Flash-Preview showed up in MiniMax Code, an anonymous model called Space Bunny Alpha appeared on OpenRouter under the identifier stealth/space-bunny-alpha, priced at zero dollars for both input and output during its preview window. Developers didn't wait for MiniMax to say anything. They ran tokenizer comparisons and error-behavior tests, the same fingerprinting techniques the community uses whenever a stealth model shows up trying to get free benchmark data out of the wild. An independent fingerprinting test sent 24 token-count probes through OpenCode's Space Bunny route and found all 24 matched MiniMax models tested on the same gateway; a separate measurement set put the match at 50 out of 50, according to reporting from The Neuron and CellCog. MiniMax hasn't confirmed that Space Bunny Alpha is M3.1-Flash-Preview. But the pattern, free anonymous benchmark run followed days later by a quiet in-product launch of a similarly specced model, isn't new, and it's not accidental. It's how you collect real-world performance data without putting your name on a model that might embarrass you.

This is the second coding-focused release from MiniMax in recent months. StartupFortune covered the company's earlier coding-only model, M3, which scored 80.5% on SWE-bench Verified and 59.0% on the harder SWE-Bench Pro. VentureBeat reported at the time that M3 beat GPT-5.5 and Gemini 3.1 Pro on that harder benchmark, at what the outlet described as 5 to 10 percent of the cost. M3's official API pricing sits at $0.30 per million input tokens and $1.20 per million output tokens, with cached input priced at $0.06. For comparison, Claude Sonnet 5 runs $2 per million input tokens and $10 per million output, and Claude Opus 5 runs $5 and $25. MiniMax's M2.7, the generation before M3, was already roughly 10 times cheaper than Claude Sonnet on a per-token basis.

Frankly, the pricing gap is the whole story here, and MiniMax knows it. A company that can undercut Anthropic and OpenAI by an order of magnitude on input tokens doesn't need flashy benchmark charts to get developers to try the next release. It just needs to ship fast enough that switching costs never get the chance to build up. M3.1-Flash-Preview is MiniMax's third meaningfully distinct coding model release in under a year, and each one has landed while the last one was still being benchmarked by outside labs.

The timing isn't isolated to MiniMax, either. Huawei used its annual conference in Shanghai to unveil the Atlas 960 SuperPoD computing cluster, a chip and cluster upgrade that AP reported came days before the Trump-Xi meeting in Washington on September 24. Huawei's Ascend chips are increasingly part of the enterprise procurement story underneath China's model race, as Chinese AI companies push harder toward domestic silicon rather than Nvidia hardware.

The strategy isn't secrecy for its own sake. It's iteration speed treated as the product. Western labs publish a model card and a blog post before letting anyone touch a new release. MiniMax is choosing to let the model answer for itself first, and worry about the paperwork later, if at all.

Also read: Alibaba's Qwen3.8-27B nearly matches Claude Opus 5.5 motion graphics on a 4090 • An AI math benchmark problem on Apéry-style proofs just got marked solved • Seaplanes Are Changing How Travelers Reach Palawan's Remote Islands

This article is posted in Entrepreneurship News, check it out for more related stories.