The Backdrop
In late September, an anonymous model quietly appeared on OpenRouter and OpenCode: stealth/space-bunny-alpha. No vendor name, no model card, zero pricing — yet it shipped with a one-million-token context window, text, image and video input, tool calling, and adjustable reasoning effort. Within days it climbed to the top of daily usage on both platforms. The community traced its fingerprint through tokenizer behavior and error formats, and nearly everything points to MiniMax's unreleased M3.1 — which MiniMax has never confirmed. On September 27, MiniMax switched on M3.1-Flash-Preview inside its own coding agent, MiniMax Code: again no model card, no public API, no price. This complete anonymous pre-launch trail deserves a close look from anyone orchestrating multiple models.
Fingerprinting: How the Community Pinned Down an Anonymous Model
Within hours of the stealth launch, developers ran three kinds of fingerprint tests. First, tokenizer comparison: feeding the same text corpus to Space Bunny and MiniMax's known model family and checking whether token splits match. One study on OpenCode found a 24/24 token match against the MiniMax family; a broader measurement set reported 50/50. Second, adversarial probes: deliberately malformed inputs to compare error formats and truncation behavior. Third, breadcrumbs in official code: MiniMax's open-source MiniMax Code repository already contained test code referencing MiniMax-M3.1, a one-million-token context, and low/high/max effort tiers — closely matching Space Bunny's published traits. The evidence chain is strong, but the technically correct description remains: a MiniMax-family fingerprint, specific version unconfirmed.
Why Vendors Love Going Anonymous
The pattern is no accident. Launching anonymously and for free outsources stress testing to real traffic across the internet: no model card means no accountability for benchmark numbers; no price means no backlash when the free window ends; and if the model underperforms, pulling the anonymous route leaves no public record. For Chinese labs, OpenRouter has become a de facto pre-launch hotline. MiniMax's M3, open-sourced in June, is a natively multimodal MoE with 428 billion total parameters and 23 billion active, scoring 80.5% on SWE-bench Verified and 59.0% on the harder SWE-Bench Pro — VentureBeat reported it beat GPT-5.5 and Gemini 3.1 Pro on that benchmark at 5–10% of the cost. With that track record, using free anonymous routes to collect real-world load data for M3.1 makes perfect sense. Meanwhile the host platform itself just leveled up: OpenRouter was announced as acquired by Stripe, with media reports putting the figure around $7.5 billion, cementing the aggregation layer's position as the traffic gateway.
Three Ledges Callers Must Clear
For anyone wiring an anonymous endpoint into production, the risks concentrate in three places. First, mutable identity: today it points to M3.1; tomorrow the underlying checkpoint may silently change, invalidating every performance conclusion you measured. Second, data terms: OpenRouter's page explicitly warns that the anonymous provider may retain prompts and completions, while OpenCode's free route is separately labeled zero-retention and no-training — same model, two routes, completely different terms. Third, uncertain lifetime: the endpoint may vanish when the preview window ends, and a free price list can reset to nothing at any time. The community consensus is pragmatic: treat Space Bunny as an evaluation endpoint, benchmark it against your workhorse models on your own repositories, but never bind your critical path to it.
Where a Unified Gateway Fits: Evaluation and Fallback, Each in Its Place
Endpoints like this — unidentified, disposable, free but with unclear terms — are exactly why multi-model calls need a unified interface. Take router.accels.tech as an example. Accels, a Singapore-based company, focuses on three things: stability (workhorse models run on official and reliable hosted endpoints with automatic failover), model coverage (mainstream and hot new models live in a single catalog, wired in as they launch), and unified billing (one key, one invoice — no registering and reconciling across every platform). A fitting use for this very topic: evaluate the anonymous model and your primary model side by side through the unified interface, changing almost nothing in your code — just the model string.
from openai import OpenAI
client = OpenAI(
base_url="https://router.accels.tech/v1",
api_key="your-accels-key",
)
resp = client.chat.completions.create(
model="stealth/space-bunny-alpha",
messages=[{"role": "user", "content": "Review the boundary-condition handling in this code."}],
)
print(resp.choices[0].message.content)
Once evaluation passes, put the anonymous endpoint into the fallback chain: your primary model first, the free anonymous model after. If the stealth route goes offline, requests automatically land on the primary model and the business never notices. Evaluation, comparison, and fallback all happen under the same base_url.
Closing
Anonymous launches will only multiply: vendors want free real-world load data, aggregation platforms want exclusive first-look stories, and both sides get what they need. For developers, the point is not guessing whether it really is M3.1 — it is making sure your invocation layer can swap, compare, and fall back at any time. Consolidate evaluation and routing behind one unified gateway, and switching models is just editing a string. Check whether your usual models are already on router.accels.tech, and toss the next anonymous hit into your evaluation queue while you are at it. If this analysis helped, tell us in the comments whether you have ever wired in an anonymous endpoint.
Sources
- Web Pulse — MiniMax slips a new coding model into its agent tool without a price tag: https://wpnews.pro/news/minimax-slips-a-new-coding-model-into-its-agent-tool-without-a-price-tag
- BlockBeats — OpenRouter 上线新匿名模型 Space Bunny，多项指纹指向 MiniMax M3.1: https://www.theblockbeats.info/flash/368770
- BuildFastWithAI — Space Bunny Review: 1M Context, Coding, Speed & Is It Worth Using?: https://blog.buildfastwithai.com/space-bunny-review
- AGI Hunt — AI News Daily 2026-09-28: https://agihunt.info/en/daily/2026-09-28
Top comments (0)