The most effective way to launch a foundation model in 2026 is no longer a splashy press release or a coordinated social media campaign. Instead, it is the quiet, high-volume stress test of an anonymous identity. When MiniMax officially unveiled its M3.1 Flash on October 19, the company was not introducing a new product to the market; it was formalizing the existence of a model that had already processed over 200 trillion tokens under the pseudonym “Space Bunny.”
This stealth-to-scale deployment playbook represents a shift in how AI labs validate their infrastructure. By operating as an anonymous provider on OpenRouter, MiniMax bypassed the initial skepticism that often greets new entrants. Space Bunny held the number one spot on OpenRouter rankings for ten consecutive days, processing 56.7 trillion tokens on that platform alone — the highest volume ever recorded for a stealth model. This strategy allowed the lab to gather real-world performance data and prove demand before the brand name was even attached to the inference engine. The agent sandbox layer has become a utility, forcing labs to compete on raw throughput and reliability rather than just marketing.
The Economics of Stealth #
M3.1 Flash enters a market hyper-focused on per-call pricing and the token-based utility models that define modern AI. MiniMax has opted for a flat-rate structure, charging $0.15 per million tokens for input and $0.60 for output, with a cached input rate of $0.015. This pricing strategy is a direct challenge to competitors like DeepSeek, whose V4.1 Flash model utilizes a tiered system. While DeepSeek offers parity with MiniMax’s $0.15 input rate during off-peak hours, its peak pricing — $0.30 for input and $1.20 for output — makes the MiniMax flat rate effectively twice as cheap for developers operating during standard business hours in the US and Europe. This aggressive monetization strategy mirrors the broader trend where labs are increasingly monetizing the retrieval layer to capture value from high-volume users.
This pricing, combined with the mixture-of-experts architecture inherited from the M3 predecessor, positions M3.1 Flash as a high-throughput commodity. According to vendor-reported data, the model achieves output speeds exceeding 200 tokens per second (TPS), with a P50 latency of over 170 TPS, and represents a 70% cost reduction compared to the M3. For builders navigating the current ecosystem, these metrics suggest a focus on high-volume, low-latency integration rather than specialized, high-cost reasoning tasks.
Benchmarks and Transparency #
Despite the scale of its pre-deployment, the official launch of M3.1 Flash remains opaque. MiniMax has not published a technical paper or a formal model card, and there has been no independent replication of the performance metrics provided by the vendor. The company reports the following vendor-reported benchmarks:
- CyberGym: 82%
- OSWorld V2: 61.7%
- DeepSWE v1.1: 73.7%
- BankerToolBench: 76.2%
- ALE: 29.6%
These figures, while impressive on paper, must be viewed through the lens of a lab that has prioritized market penetration over academic transparency. The model, which features a 1-million-token context window and multimodal capabilities, is currently restricted to subscription-only access via the M Plan and MiniMax Code. While the company has promised to release open weights around October 23 and has teased a future Pro tier, the current ecosystem is left to evaluate the model based on its anonymous performance history rather than verifiable technical documentation.
The Path Forward #
MiniMax’s trajectory — from a 2022 startup to a platform claiming over 200 million individual users — highlights the efficacy of the stealth-to-scale model. By treating the open-market leaderboard as a sandbox, the lab successfully de-risked its infrastructure before the official product launch. For investors and builders, the question is no longer whether a model can perform, but how it sustains its cost-to-performance ratio once it moves from the shadows of an anonymous provider to the scrutiny of a public API. As the industry moves toward the promised open-weights release, the focus will shift from the cleverness of the deployment strategy to the long-term reliability of the underlying architecture.