Space Bunny Alpha › Benchmarks
Space Bunny Alpha Benchmarks: What's Been Measured
There are no official benchmark scores for Space Bunny Alpha. Neither TokenRA nor the anonymous provider has published an evaluation table, so there is no MMLU, SWE-bench or AIME number to cite. What does exist is a set of operational measurements and one independent tokenizer analysis — below is each of them, with the source of every figure labelled.
Last updated 25 September 2026 · Space Bunny Alpha was released 23 September 2026
How we label sources on this page
Anonymous models attract a lot of confident numbers. We separate what the provider states from what a third party has measured from what is only a hypothesis.
Why there is no benchmark table for this model
Space Bunny Alpha is a stealth release: a lab puts a model into public testing under a codename with the provider hidden. In that arrangement there is no model card, no paper, no evaluation appendix and no vendor to publish one. The TokenRA listing contains the model specifications, while the OpenRouter listing displays live operational metrics — neither is a quality evaluation table.
This is worth stating plainly because it is the reason a search for space bunny alpha benchmarks is unhelpful right now. At the time of writing the results for that query are dominated by the model's own listing pages and by evaluations of other stealth models — Union Alpha, Ox Alpha, Horizon Alpha. Nobody has published a benchmark page for this model. That absence is the finding.
Operational metrics Listing-reported
These operational figures are reproduced from the OpenRouter listing as observed on 25 September 2026. The displayed three-day window runs from Sep 22, 10 AM to Sep 25, 10 AM. They describe gateway service operation, not model quality, and are not independent measurements by this site.
| Metric | Value (25 Sep 2026) | What it means |
|---|---|---|
| Throughput | 80 tok/s (P50, best across providers) | How fast the model writes |
| Latency | 1.05 s (P50) | Total round-trip time |
| Uptime (3d) | 99.76% | At least one provider was responding |
| Availability (3d) | 98.24% | Inference was successfully served |
| OpenRouter availability (24h) | 98.57% | Inference successfully served over the last 24 hours |
| Price | Free | No charge for prompt or completion tokens |
Source: OpenRouter listing, observed 25 September 2026. It defines availability as the percentage of time inference was successfully served and says it uses the next-best provider when one returns an error. The 98.24% figure is therefore a listing-reported OpenRouter gateway metric, not an independent test by this site, not a TokenRA metric, and not proof that the two routes share one upstream provider.
Specifications Provider-listed
| Spec | Value |
|---|---|
| Context window | 1,000,000 tokens |
| Max output | 524,288 tokens |
| Input modalities | Text, image, video |
| Output | Text |
| Reasoning | Adjustable reasoning effort |
| Tool calling | Supported (tools, tool_choice) |
| Structured output | response_format JSON, without JSON-schema enforcement |
| Provider / gateway | Anonymous third-party provider; TokenRA and OpenRouter are separate gateway paths, and shared upstream is not confirmed |
| Released | 23 September 2026 |
Independent measurements Independent
The only third-party measurement we can point to comes from stealthprint, an open-source fingerprinting toolkit (MIT licence) whose space-bunny case was published on 24 September 2026. Its method is a vocabulary differential: send the same probe text to the mystery model and to candidate tokenizers, and compare token counts. Because the probe differential cancels the chat-template constant,leaving only the model's own tokenizer count.
| Measurement | Result | What it supports |
|---|---|---|
| Tokenizer vocabulary (24 probes) | 24 / 24 matched MiniMax | The strongest public identity signal. Same run scored Kimi 13/24 and GLM 7/24, so this is not a generic "everything looks similar" result. |
| Chat-wrapper overhead | +143 tokens, zero drift | Stable serving path — the gateway template does not change between calls. |
| Context ceiling | ≥ 1M tokens (binary search) | Confirms the listed 1M window is real, and rules out the older m2.5 generation. |
| Same-gateway delta | Identity with minimax-m3 / m2.5 | Points to the same serving family as the named MiniMax models. |
| Vision stack | Adapter-class vision | Image input is real and size-scaled, not a placeholder. |
| Backend pool | Homogeneous | Repeated identical probes returned a consistent backend. |
| Self-identification | Improvised ChatGPT claim | Deliberately weighted at zero. For stealth models the self-report is frequently bait — see the caveat below. |
Source: stealthprint case study, 24 September 2026 (MIT). We did not run these tests; we are reporting them with attribution. Methodology builds on public prior work — tokenizer-fingerprint and ox-alpha-identification-public.
What these numbers are not
Every figure above is an identity measurement — it tells you which toolchain the model resembles. None of it is a capability measurement. A 24/24 tokenizer match says nothing about whether the model writes better or worse code than its named sibling. Do not read this table as a quality ranking.
Two specific traps worth naming:
- The self-report is bait. When asked who it is, this model reportedly produces an improvised claim about being a ChatGPT model. Self-descriptions from stealth models are frequently wrong by design. Any page that quotes it as an answer is quoting noise.
- A "blazing-fast" adjective is not a benchmark. The listing describes blazing-fast inference while the latency field carries a live number. Marketing adjectives and measured throughput are different kinds of claim; only the second one is checkable.
What people are actually running on it Provider-listed
The listing publishes the public apps sending the most traffic to this model. This is the closest thing to a real-world workload benchmark that exists right now, and the pattern is unambiguous:
| App | Tokens | Type |
|---|---|---|
| Cline | 54.1B | Coding agent (IDE) |
| Claude Code | 50.4B | Agentic coding tool |
| Hermes Agent | 32.1B | Open-source autonomous agent |
| DeepSeek Harness | 30B | Agent harness |
| Kilo Code | 28.9B | Agentic coding tool |
Every top consumer is a coding agent. Five for five, with no chat app and no consumer product in the list. That is the strongest available signal about what this model is good at: long-horizon agentic coding with large contexts, exactly the workload the 1M window and tool calling support. It is not a benchmark score, but it is measured usage rather than a claim — and it is the number we would weigh most heavily when deciding whether to build on this model.
Token volumes are cumulative and grow continuously; the ranking above is a 24 September 2026 snapshot.
How to measure it yourself
Because no official table exists, your own numbers are the only ones that describe your workload. Four steps, in order of cost:
- Run your own task set. Take ten real prompts from your actual use case, run them against
space-bunny-alphaand against the model you would otherwise pay for, and score the outputs blind. This is the only comparison that answers your question. - Measure your own latency and cost. The P50 figures above are global; your numbers depend on your region, prompt length and reasoning effort setting.
- Test the availability gap. Rerun your set enough times to see whether the current OpenRouter snapshot generalises to your route. Decide what your fallback is before you depend on a free preview.
- Reproduce the fingerprint if you want to verify identity. The stealthprint toolkit is open source:
stealthprint tokenizer --tokenizers tok/reruns the vocabulary differential with your own tokenizer files.
Our own test results are in progress and will be published on the home page as they complete. We would rather show an empty slot than a borrowed number.
Space Bunny Alpha benchmarks: FAQ
Are there official benchmark scores for Space Bunny Alpha?
No. As of 24 September 2026 neither TokenRA nor the anonymous provider has published an evaluation table. There is no MMLU, SWE-bench or AIME score to cite, so any figure quoted for this model is either inherited from another model or invented.
What has actually been measured?
Two categories. OpenRouter listing-reported operational metrics (80 tok/s P50, 1.05 s P50 latency, 99.76% three-day uptime, 98.24% three-day availability, free pricing) and one independent analysis: 24/24 tokenizer vocabulary match, +143-token wrapper overhead, and a verified context ceiling of at least 1M tokens.
How fast is Space Bunny Alpha?
The OpenRouter listing reports roughly 80 tokens per second at P50 and 1.05 seconds P50 latency in the 25 September 2026 snapshot — live operational figures that move over time, describing serving speed rather than model quality.
Does it have coding benchmark results?
None have been published. The listing claims strong coding capability, and the five heaviest consumers are all coding agents, but no coding benchmark score has been released by the provider or measured independently.