EARLIER BUNNY · EXACT MATCH24 / 24
MiniMax M3
Every phrase adds the same number of tokens as it did for Bunny in the earlier comparison. A complete match for that snapshot.
REVEALED: AN EARLY MINIMAX M3.1 FLASH
Space
BunnyALPHA
Big context. Curious mind.
Meet space-bunny-alpha. For nearly four weeks it hopped around anonymously, with a million-token context and image and video input. On 19 October, MiniMax took off the helmet: Space Bunny was an early version of MiniMax M3.1 Flash.
Independent field guide · Revealed
SPECIMEN 001 / ORIGIN: MINIMAX
1M tokensPreview context window
524K tokensPreview completion ceiling · 524,288
Text. Image. Video.Preview inputs → text output
$0.15 / 1M input$0.60 / 1M output · M3.1 Flash API
Advertisement
A listening break
DEMODOKOS / LOCAL AI AUDIO
A listening break
for long ears.
Demodokos makes expressive dialogue and original music on your PC.
Hear a scene below.
A little ear candy.
Press play0:00--:--
Scene 01 · Demodokos demo
01 SCORES & MODEL COMPARISONS
Other benchmark results.
Space Bunny Alpha’s AI BENCHY, GPQA Diamond, MMLU-Pro and Humanity’s Last Exam (HLE) scores, token efficiency and model comparisons.
PRACTICAL TASKSAI BENCHY · high
space-bunny-alphaAI BENCHY
7.0/ 10
Perfect category scores in
data extraction & tool calling.
12/22 tests fully passed62.1% attempt pass rate
Compare AI BENCHY scores
SCIENCE REASONINGOur evaluationspace-bunny-alphaGPQA Diamond
82.0%1
Graduate-level questions in
biology, chemistry & physics.
Our evaluation60-question subset
Compare GPQA Diamond scores
KNOWLEDGE & REASONINGOur evaluationspace-bunny-alphaMMLU-Pro
75%
Reasoning across science,
humanities & professional subjects.
Our evaluationMultiple-choice accuracy
Compare MMLU-Pro scores
EXPERT-LEVEL REASONINGOur evaluationspace-bunny-alphaHumanity’s Last Exam
46.1%
An unusually strong result
relative to its other benchmarks.
Original HLE · 300-question subsetApprox. 95% CI 40.4–51.8%
Compare HLE scores
1 Our GPQA Diamond result covers a 60-question subset. Other models: published reference results.
TOKEN EFFICIENCY · OUR BENCHMARK RUNS
space-bunny-alpha token efficiency
About one-third of Qwen3.8 Flash’s output tokens: 67% fewer on the same benchmarks.
- Space Bunny Alpha
- 305,989output tokens326,078 total tokens
- Qwen3.8 Flash
- 913,989output tokens931,392 total tokens
space-bunny-alphaAI BENCHY
SCORE / 10 · HIGHER IS BETTER-
GPT-6 Sol 6.09.9high reasoning
-
GPT-5.59.3low reasoning
-
Qwen3.8-27B8.4high reasoning
-
MiniMax M37.5medium reasoning
-
space-bunny-alphaTHIS MODEL7.0high reasoning
-
gpt-oss-120b6.1medium reasoning
Qwen3.8 Flash N/A· gpt-oss-20b N/A
Categories & reasoning modes
Space Bunny by category
- Anti-AI Tricks
- 6.4 / 10
- Coding
- 6.2 / 10
- Combined
- 6.5 / 10
- Data parsing / extraction
- 10.0 / 10
- Domain specific
- 7.7 / 10
- General Intelligence
- 4.2 / 10
- Instruction following
- 6.2 / 10
- Puzzle Solving
- 8.3 / 10
- Tool Calling
- 10.0 / 10
- Trivia
- 3.0 / 10
Space Bunny by reasoning mode
High is the strongest published run in this snapshot.
- high
- 7.0
- max
- 6.9
- xhigh
- 6.5
- medium
- 6.3
- low
- 5.9
Consistency 8.5/10 · API reliability 10.0/10 in the high run.
Method & comparison data
Best published configuration in this comparison, as of 25 Sep 2026. AI BENCHY combines correctness and consistency across 22 private tests with repeated attempts. A test is fully passed only when every attempt succeeds; the attempt pass rate counts individual attempts.
| Model / source | Score | Reasoning | Fully passed | Attempt pass rate | Avg. response |
|---|---|---|---|---|---|
| GPT-6 Sol 6.0 | 9.9 | high | 21/22 | 95.5% | 22.21s |
| GPT-5.5 | 9.3 | low | 19/22 | 86.4% | 10.15s |
| Qwen3.8-27B | 8.4 | high | 16/22 | 77.3% | 167.89s |
| MiniMax M3 | 7.5 | medium | 11/22 | 62.1% | 75.15s |
| Space Bunny Alpha | 7.0 | high | 12/22 | 62.1% | 27.38s |
| gpt-oss-120b | 6.1 | medium | 9/22 | 50.0% | 20.81s |
| Qwen3.8 Flash | N/A | — | — | — | — |
| gpt-oss-20b | N/A | — | — | — | — |
Qwen3.8-27B ran locally on an RTX 3090, with OpenRouter used for judging. Response times reflect each run’s hardware and serving setup. N/A means no verified AI BENCHY entry in this snapshot.
Space Bunny’s high run: 24 Sep 2026, 02:23 · stealth/space-bunny-alpha::high. AI BENCHY methodology
space-bunny-alphaGPQA Diamond
ACCURACY % · HIGHER IS BETTER-
GPT-6 Sol 6.095.45%Independent GPQA audit · 198 questions
-
GPT-5.593.6%OpenAI-published result
-
MiniMax M392.9%Artificial Analysis standardized evaluation
-
Qwen3.8 Flash91.7%Qwen-published Flash-Next reference
-
Qwen3.8-27B89.2%Qwen-published result
-
space-bunny-alphaTHIS MODEL82.0%1Our 60-question subset
-
gpt-oss-120b80.1%OpenAI-published result
-
gpt-oss-20b71.5%OpenAI-published result
1 Our GPQA Diamond result covers a 60-question subset. Other models: published reference results.
Method & result provenance
Space Bunny: our 60-question subset. GPT-6 Sol: independent audit on the original 198 questions. Qwen3.8 Flash uses Qwen’s published Flash-Next reference result.
| Model | Accuracy | Provenance |
|---|---|---|
| GPT-6 Sol 6.0 | 95.45% | Independent GPQA audit · 198 questions |
| GPT-5.5 | 93.6% | OpenAI-published result · xhigh; OpenAI research evaluation |
| MiniMax M3 | 92.9% | Artificial Analysis standardized evaluation |
| Qwen3.8 Flash | 91.7% | Qwen-published Flash-Next reference |
| Qwen3.8-27B | 89.2% | Qwen-published result |
| Space Bunny Alpha | 82.0% | Our 60-question subset |
| gpt-oss-120b | 80.1% | OpenAI-published result · high reasoning; no tools |
| gpt-oss-20b | 71.5% | OpenAI-published result · high reasoning; no tools |
space-bunny-alphaMMLU-Pro
ACCURACY % · HIGHER IS BETTER-
GPT-5.6 Sol89.1%Reference result
-
Qwen3.8 Flash≈89.0%Approximate reference
-
GPT-5.588.1%Reference result
-
Qwen3.8-27B84.3%Reference result
-
MiniMax M384.2%Reference result
-
gpt-oss-120b80.8%Reference result
-
space-bunny-alphaTHIS MODEL75%Our evaluation
-
gpt-oss-20b74.8%Reference result
Space Bunny Alpha: our evaluation. Other models: reference results. Qwen3.8 Flash’s ≈89.0% is approximate.
About MMLU-Pro & comparison data
MMLU-Pro is a reasoning-focused multiple-choice benchmark with up to ten answer choices per question. Results below are accuracy percentages for MMLU-Pro.
| Model | Accuracy | Result |
|---|---|---|
| GPT-5.6 Sol | 89.1% | Reference result |
| Qwen3.8 Flash | ≈89.0% | Approximate reference |
| GPT-5.5 | 88.1% | Reference result |
| Qwen3.8-27B | 84.3% | Reference result |
| MiniMax M3 | 84.2% | Reference result |
| gpt-oss-120b | 80.8% | Reference result |
| Space Bunny Alpha | 75% | Our evaluation |
| gpt-oss-20b | 74.8% | Reference result |
space-bunny-alphaHumanity’s Last Exam
ACCURACY % · HIGHER IS BETTER-
GPT-5.6 Sol49.5%Reference result
-
GPT-6 Sol47.9%Reference result
-
space-bunny-alphaTHIS MODEL46.1%Our evaluation · 300-question subset
-
GPT-5.545.8%Reference result
-
MiniMax M339.0%Reference result
-
Qwen3.8 Flash38.0%Reference result
-
Qwen3.8-27B33.9%Reference result
-
gpt-oss-120b14.9%Reference result
-
gpt-oss-20b10.9%Reference result
Space Bunny Alpha: our 300-question subset of the original HLE dataset. Other models: reference results.
Method, confidence interval & comparison data
Humanity’s Last Exam (HLE) tests expert-level knowledge across a broad range of subjects. This evaluation uses 300 questions from the original HLE dataset.
Our HLE accuracy is 46.1%, with a standard error of 2.9 percentage points. The approximate 95% confidence interval uses a normal approximation: 46.1% ± 1.96 × 2.9 percentage points = 40.4–51.8%. This describes sampling uncertainty for the subset estimate.
Additional reported accuracy: 45.0%. 7 questions were unscored.
| Model | Accuracy | Result |
|---|---|---|
| GPT-5.6 Sol | 49.5% | Reference result |
| GPT-6 Sol | 47.9% | Reference result |
| Space Bunny Alpha | 46.1% | Our evaluation · 300-question subset |
| GPT-5.5 | 45.8% | Reference result |
| MiniMax M3 | 39.0% | Reference result |
| Qwen3.8 Flash | 38.0% | Reference result |
| Qwen3.8-27B | 33.9% | Reference result |
| gpt-oss-120b | 14.9% | Reference result |
| gpt-oss-20b | 10.9% | Reference result |
02 THE SVG LAB
Five benchmark tasks.
No paintbrush.
Moon bunnies, portraits and improbable passengers. See what Space Bunny draws with SVG code.
Notice the helmet reflections and the smaller bunny peeking over the moon.EXPERIMENT 01 / SPACE BUNNYOne giant hop.
One giant hop.
One small carrot.
Benchmark task
Before 29 Sep
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
29 Sep onward
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
EXPERIMENT 02 / MONA LISAMona Lisa looks
Mona Lisa looks
rather pleased.
Benchmark task
Before 29 Sep
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
29 Sep onward
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
EXPERIMENT 03 / PELICANPedal.
Pedal.
Flap. Repeat.
Benchmark task
Before 29 Sep
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
29 Sep onward
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
EXPERIMENT 04 / RACCOONTwo unlikely riders.
Two unlikely riders.
One jet ski.
Benchmark task
Before 29 Sep
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
29 Sep onward
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
EXPERIMENT 05 / BEACHEven bunnies
Even bunnies
need a holiday.
Benchmark task
Before 29 Sep
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
29 Sep onward
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
THE EVERYDAY DETAILS
Speed & model notes.
Archived OpenRouter snapshot · 24 Sep 2026
P50 = median
- Writing speed · P50
- 87tokens / second
- Response latency · P50
- 1.07 s
- Inference availability
- 94.98%3-day view
Measurements & sources
Earlier speed & availability
How fast it writes. OpenRouter P50 throughput, best across providers. Dashboard filters: all reasoning efforts and all locations. Snapshot around 00:28 UTC on 24 September 2026.
How quickly it responds. OpenRouter P50 latency, best provider. Filters: all reasoning efforts and all locations. Snapshot around 00:28 UTC on 24 September 2026.
Requests that got a response. OpenRouter’s 3-day dashboard measures whether inference returned from at least one provider. Reachability is separate at 97.70%; 24-hour inference availability is 95.60%. Snapshot around 00:28 UTC on 24 September 2026; model listed 23 September.
A long-context check
3/3 codes recovered in one 200K-token run. Found three different codes hidden in a very long input, and returned them in the correct order.
All three distinct codes were returned in order in one trial containing 200,044 local tokens and 200,187 reported prompt tokens.
Independent OpenCode Go test ↗Earlier OpenCode speed comparison · 24 Sep 2026
| Route | 24h avg. speed | Latest first text | 24h successful tests |
|---|---|---|---|
| OpenCode Go | 94.3 tok/s | 1.2s | 7/7 |
| OpenCode Zen | 84.9 tok/s | 1.6s | 4/4 |
Same HTTP-routing writing task, 300-token output cap. Method ↗ · Go results ↗ · Zen results ↗
03 BUNNY UNMASKED
Inside the
helmet: MiniMax.
MiniMax took off the helmet. Its 19 October launch post says Space Bunny “was an early version of MiniMax M3.1 Flash.” Our token counts called the family on 29 September: MiniMax M3 matched all 24 test phrases. The GPT router theory we led with from 30 September is superseded.
IDENTITY STATUSRevealed.Early MiniMax M3.1 Flash.
Still open · 19 Oct
MiniMax says “an early version”. It has not said whether Space Bunny matches the M3.1 Flash Preview that MiniMax Code listed on 27 September or the 19 October release, or whether the change our researchers saw on 29 September was a new checkpoint. Our results describe the stealth model.
The OpenCode connection · 29 Sep
OpenCode announced an improved Bunny and a free preview through at least around DevDay, then sponsored a few more days while the model was tuned for its reveal. MiniMax made that reveal on 19 October.
Confirmed by MiniMax
The tokenizer called it.
Earlier Bunny · Token-count match
24 / 24matching phrases
MiniMax M3 and Bunny
MiniMax M3 and Bunny
We tested the same 24 phrases across 28 models on 29 September. MiniMax M3 matched every input-count increment in Bunny’s results; Sol 6.1 matched seven. On 19 October, MiniMax confirmed that Space Bunny was an early version of MiniMax M3.1 Flash.
What we measured, and what it turned out to be. The count match pointed to MiniMax’s tokenizer family; it could not name a release. From 30 September we ranked runtime clues above it. The reveal shows the counts were the signal to follow.
M3.1 Flash · 27 Sep test drawings
Our 27 September M3.1 Flash test drawings look markedly different from Bunny’s, and we judged an exact Flash match unlikely. MiniMax now calls Bunny an early version, which fits a different checkpoint. The tasks were not identical, so the drawings cannot settle it.
The 24-phrase comparison was captured on 29 September, before the change our researchers observed that day.
Superseded assessment
We bet on a GPT core.
30 Sep assessmentSuperseded 19 Oct
Our 30 September assessment named a model fusion router with a GPT-6.1 Sol or Astra core. MiniMax’s reveal supersedes it; the probe findings below are dated observations.
01 / Harmony-style channels
Our 30 September probe returned Valid channels: analysis, commentary, final, summary. alongside # Juice. The system text reads You are an AI assistant accessed via API.
; a developer turn says You are a helpful assistant.
We read these as clues to an OpenAI-style runtime.
What we made of Juice and summary. Our analysis read # Juice as an OpenAI-specific reasoning-budget marker, and summary as a fourth internal channel that shows a summary while hiding the full reasoning. We took them as strong support for a GPT core. MiniMax’s statement does not explain why they appeared.
02 / Context is rewritten
Our probe shows earlier turns compressed into developer messages. Within a few turns, the summaries lose Bunny’s own codename. This points to active context management that does not reliably preserve its assigned identity. We explained it with per-turn summarization and model selection. The summarization is what we observed; the model selection was our inference.
Three layers, one model. Our theory gave GPT the core reasoning, MiniMax-type counting the input usage and DeepSeek V4 Flash the visible reasoning. MiniMax names one model, M3.1 Flash, so we no longer assign these layers.
03 / The Sol timing · 29 Sep
Before DevDay, Bunny’s visual style closely resembled GPT-6 Sol’s. Within hours of OpenAI releasing GPT-6.1 Sol, Bunny’s behavior changed abruptly and its visual style became strikingly similar to Sol 6.1’s. We inferred a reasoning-core upgrade; MiniMax’s reveal supersedes that inference. Whether MiniMax shipped a new stealth checkpoint that day is unknown. Compare the raccoon and beach benchmark results.
Who made it? MiniMax. Our candidates had been OpenAI, Pareto or another operator with enough API funding.
OpenAI’s Harmony guide provides the reference for the role structure and three main channels. The probe’s fourth summary channel, # Juice marker and developer-turn compression were central to our superseded GPT router assessment.
Earlier reasoning study
Similar habits, different family.
DeepSeek V4Closest in the earlier study
V4 Pro: 83.8 / 100. V4.1 Flash: 82.6 / 100.
Across 48 prompts, DeepSeek is closest to Bunny’s earlier reasoning habits. Shared “We need answer” openings and clipped “Need…” planning stand out. This measures shared habits; it does not show that DeepSeek produced Bunny’s reasoning.
Habits are not identity. Our router theory gave DeepSeek V4 Flash the exposed reasoning channel. MiniMax’s reveal supersedes that assignment; the scores still show how closely Bunny’s visible reasoning resembled DeepSeek’s.
The earlier tokenizer test matches on 9 of 24 phrases. The SVG comparison weighs against an exact V4 Pro match. These measurements remain dated history.
Earlier alternative
The drawings weaken the Qwen lead.
QWEN3.8 · 2.4T-A95B
70.4earlier reasoning similarity
out of 100
out of 100
Qwen3.8 2.4T-A95B is the closest model outside DeepSeek in the earlier study. Qwen3.8 27B follows at 69.9 and Max-0902 at 69.2.
Qwen3.8 Flash’s five SVG results differ in style and construction. The earlier count test gives Qwen3.8 only 3 matches out of 24. These comparisons weaken the proposed Qwen4 Flash connection; Qwen4 itself has not been tested here.
EARLIER TOKEN COUNTS · 29 SEP 2026
Input counts.
The right trail.
Measured on 29 September, before the change our researchers saw that day. Across 28 models, MiniMax M3 matches all 24 phrases; GPT 6.1 Sol matches seven. MiniMax later revealed Bunny as an early M3.1 Flash.
Counted carrots.
Called it.
EARLIER BUNNY · PARTIAL MATCH7 / 24
GPT 6.1 Sol
17 of 24 counts differ. The average gap is 1.375 tokens per phrase. This compares Sol with the earlier Bunny results, before the observed improvement.
Same count as Bunny Different countOne square per phrase
The counts pointed to MiniMax. From 30 September we ranked our Harmony and context clues above them and placed these counts at an input layer. MiniMax’s 19 October reveal supersedes that router reading. A count match shows a shared tokenizer, not an exact release.
EXPLORE THE EARLIER COUNTS
One phrase. Three models.
EXACT TEST TEXTI can't believe it's already Tuesday.
Space Bunny · earlier results7tokens added
MiniMax M37tokens added
GPT 6.1 Sol7tokens added
All three models add 7 tokens for this phrase.
All 24 phrases and their counts
| Test phrase | Bunny Earlier results | MiniMax M3 | GPT 6.1 Sol |
|---|---|---|---|
English sentenceI can't believe it's already Tuesday. | 7 | 7 | 7 |
Long English wordshyperparameterization and counterrevolutionaries | 7 | 7 | 8 |
German compound wordDonaudampfschifffahrtsgesellschaft | 11 | 11 | 10 |
German sentenceFünf Straßenbahnen überqueren die Brücke. | 13 | 13 | 13 |
Simplified Chinese今天天气很好，我们去图书馆。 | 7 | 7 | 10 |
Traditional Chinese今天天氣很好，我們去圖書館。 | 10 | 10 | 11 |
Japanese今日は図書館で新しい本を読む。 | 7 | 7 | 11 |
Korean오늘은 도서관에서 새 책을 읽어요. | 9 | 9 | 12 |
Arabicاللغة العربية جميلة، أليس كذلك؟ | 10 | 10 | 9 |
Hindiनमस्ते दुनिया, आज मौसम सुहाना है। | 19 | 19 | 13 |
Mixed scripts: cities東京Berlin上海2026 | 5 | 5 | 5 |
Mixed scripts: greetingmañana東京مرحبا42 | 8 | 8 | 6 |
Code: identifierHTTPRequest.parseJSON(user_id_42) | 9 | 9 | 8 |
Code: expressionconst café = arr?.map(x => x**2) ?? []; | 15 | 15 | 14 |
Math: summation∑_{i=1}^{n} i² = n(n+1)(2n+1)/6 | 22 | 22 | 23 |
Math: logic∀x∈ℝ: x²≥0 ⇒ √(x²)=|x| | 19 | 19 | 20 |
Number sequences00000123456789 3.141592653589793 | 13 | 13 | 13 |
Date and time2026-09-29T23:59:59.123+02:00 | 18 | 18 | 18 |
Accented letterscafé naïve coöperate | 6 | 6 | 7 |
Accents as separate characterscafé naïve coöperate | 6 | 6 | 10 |
Joined emoji👩🏽💻👨👩👧👦🇩🇪 | 21 | 21 | 22 |
Mixed emoji🙂🙃✨❤️🔥🧑🚀 | 13 | 13 | 13 |
Spaces, tab and newlinealpha beta gamma delta | 8 | 8 | 8 |
Web addresshttps://例え.テスト/a_b?q=über&n=42#frag | 17 | 17 | 18 |
Compare the earlier Bunny with all 27 other models
| Model | Matching phrases | Average token gap |
|---|---|---|
| minimax/minimax-m3 | 24 / 24 | 0.00 |
| deepseek/deepseek-v4-pro-0813 | 9 / 24 | 1.17 |
| deepseek/deepseek-v4.1-flash | 9 / 24 | 1.17 |
| x-ai/grok-4.7 | 9 / 24 | 2.42 |
| openai/gpt-6-astra | 7 / 24 | 1.38 |
| openai/gpt-6-luna | 7 / 24 | 1.38 |
| openai/gpt-6-sol | 7 / 24 | 1.38 |
| openai/gpt-6.1-sol | 7 / 24 | 1.38 |
| openai/gpt-oss-120b | 7 / 24 | 1.38 |
| tencent/hy3 | 7 / 24 | 2.29 |
| moonshotai/kimi-k3 | 6 / 24 | 2.42 |
| z-ai/glm-5.3 | 5 / 24 | 2.38 |
| z-ai/glm-5.3-flash | 5 / 24 | 2.38 |
| z-ai/glm-5.3-flashx | 5 / 24 | 2.38 |
| z-ai/glm-5.3-prime | 5 / 24 | 2.38 |
| xiaomi/mimo-v2.6-flash | 4 / 24 | 3.13 |
| xiaomi/mimo-v2.6-pro | 4 / 24 | 3.13 |
| mistralai/mistral-small-2603 | 4 / 24 | 4.83 |
| nvidia/nemotron-3.5-lightning | 4 / 24 | 4.83 |
| qwen/qwen3.8-2.4t-a95b | 3 / 24 | 3.42 |
| qwen/qwen3.8-27b | 3 / 24 | 3.42 |
| qwen/qwen3.8-flash | 3 / 24 | 3.42 |
| qwen/qwen3.8-max-0902 | 3 / 24 | 3.42 |
| qwen/qwen3.8-max-prime | 3 / 24 | 3.42 |
| qwen/qwen3.8-omni-flash | 3 / 24 | 3.58 |
| google/gemini-3.8-flash | 3 / 24 | 3.63 |
| anthropic/claude-sonnet-5.5 | 0 / 24 | 7.75 |
Comparison scope
Stealth Models’ token-count comparison, measured on 29 September 2026 before Bunny’s observed update.
The comparison covers 24 phrases across 12 groups, including languages, code, numbers, emoji and spacing.
These are reported input-token increments, not recovered token IDs. A count match does not establish a shared developer. Both DeepSeek models match 9 of 24 phrases.
EARLIER REASONING SAMPLES
The earlier reasoning trail.
DeepSeek leads this comparison of Bunny’s reasoning habits. These samples precede the change observed on 29 September. MiniMax later revealed Bunny as an early M3.1 Flash.
83.8similarity / 100
Closest of 26 models
DeepSeek V4 Pro
Our study of
reasoning habits
V4 Pro scores 83.8; V4.1 Flash follows at 82.6. Qwen3.8 2.4T-A95B is next at 70.4. This measures shared planning and wording habits; it does not show that DeepSeek produced Bunny’s reasoning.
MiniMax leads the earlier token-count comparison but scores 56.2 here. Reasoning habits can resemble another family’s: the token counts, not this study, pointed to the developer MiniMax confirmed. Our superseded router theory once explained the gap with a separate DeepSeek reasoning layer.
What are we comparing?
- The same questions
We compare the recorded thinking for questions both models answered.
- Recognizable habits
How they begin, restate a question, plan an answer and handle sensitive topics.
- A similarity score
More matching habits give a higher score. 100 means all the scored features agree.
Our study covers 48 questions across eight topics. The score compares reasoning habits; it does not identify a developer.
Each model is one specific version. Higher scores mean more similar reasoning habits.
Overall · top 6 modelsSimilarity / 100
- DeepSeek V4 Pro · Aug 1383.8
- DeepSeek V4.1 Flash82.6
- Qwen3.8 2.4T-A95B70.4
- Qwen3.8 27B69.9
- Qwen3.8 Max · Sep 269.2
- Qwen3.8 Flash64.6
All 26 models · all categories Similarity / 100 · higher means closer
| Model | Overall | Adult topics | Ambiguity | Creative | Crisis support | Disclosure | Explanations | Politics | Math |
|---|---|---|---|---|---|---|---|---|---|
| deepseek/deepseek-v4-pro-0813 | 83.8 | 76.5 | 84.3 | 82.5 | 84.8 | 83.9 | 84.5 | 81.1 | 92.6 |
| deepseek/deepseek-v4.1-flash | 82.6 | 68.4 | 83.7 | 86.3 | 86.0 | 82.5 | 86.1 | 76.4 | 91.7 |
| qwen/qwen3.8-2.4t-a95b | 70.4 | 57.1 | 79.1 | 57.6 | 67.9 | 73.3 | 75.7 | 58.4 | 94.4 |
| qwen/qwen3.8-27b | 69.9 | 54.8 | 77.9 | 59.4 | 68.5 | 67.4 | 76.8 | 61.3 | 92.9 |
| qwen/qwen3.8-max-0902 | 69.2 | 56.8 | 77.7 | 57.0 | 65.1 | 72.4 | 73.5 | 57.3 | 93.3 |
| qwen/qwen3.8-flash | 64.6 | 60.6 | 60.2 | 72.3 | 71.0 | 62.5 | 68.7 | 54.1 | 67.2 |
| qwen/qwen3.8-max-prime | 62.1 | 55.2 | 56.1 | 55.5 | 48.3 | 50.3 | 71.0 | 70.0 | 90.7 |
| anthropic/claude-sonnet-5.5 | 59.8 | 55.4 | 44.0 | 74.2 | 61.0 | 46.1 | 65.2 | 65.7 | 66.5 |
| openai/gpt-6-sol | 58.8 | 60.5 | 41.2 | 63.8 | 61.2 | 46.0 | 68.8 | 69.4 | 59.7 |
| tencent/hy3 | 58.6 | 55.7 | 47.6 | 68.8 | 58.1 | 42.0 | 58.5 | 56.6 | 81.6 |
| openai/gpt-6-astra | 57.3 | 61.4 | 43.9 | 61.0 | 55.0 | 47.1 | 65.0 | 67.2 | 58.1 |
| moonshotai/kimi-k3 | 56.8 | 51.4 | 52.0 | 52.5 | 56.9 | 62.5 | 60.7 | 58.3 | 60.4 |
| minimax/minimax-m3 | 56.2 | 59.7 | 48.7 | 68.2 | 58.6 | 46.8 | 52.0 | 49.5 | 66.3 |
| z-ai/glm-5.3-prime | 55.9 | 52.1 | 53.0 | 52.7 | 57.1 | 52.9 | 62.5 | 59.7 | 57.5 |
| openai/gpt-6-luna | 55.6 | 57.0 | 38.5 | 64.7 | 54.1 | 44.0 | 61.7 | 64.7 | 60.4 |
| z-ai/glm-5.3 | 55.6 | 50.4 | 52.2 | 52.6 | 54.9 | 54.4 | 63.7 | 59.2 | 57.5 |
| z-ai/glm-5.3-flash | 55.4 | 49.5 | 52.1 | 52.0 | 57.3 | 55.3 | 59.2 | 55.1 | 62.7 |
| mistralai/mistral-small-2603 | 55.1 | 50.2 | 49.0 | 62.8 | 57.6 | 52.2 | 51.6 | 55.0 | 62.4 |
| z-ai/glm-5.3-flashx | 55.0 | 51.3 | 53.0 | 51.6 | 56.4 | 52.6 | 58.6 | 55.7 | 61.2 |
| xiaomi/mimo-v2.6-pro | 53.9 | 47.3 | 52.9 | 50.0 | 53.7 | 52.7 | 56.0 | 60.7 | 58.0 |
| qwen/qwen3.8-omni-flash | 53.9 | 55.9 | 39.4 | 64.8 | 50.3 | 44.8 | 61.7 | 54.7 | 59.6 |
| google/gemini-3.8-flash | 52.1 | 52.1 | 35.9 | 61.5 | 43.1 | 51.0 | 64.1 | 60.2 | 48.9 |
| nvidia/nemotron-3.5-lightning | 50.5 | 52.2 | 41.2 | 65.5 | 50.8 | 36.4 | 46.4 | 50.0 | 61.1 |
| xiaomi/mimo-v2.6-flash | 48.8 | 39.4 | 45.2 | 42.5 | 51.8 | 48.8 | 48.9 | 57.9 | 55.9 |
| openai/gpt-oss-120b | 46.2 | 42.5 | 36.1 | 53.5 | 39.9 | 31.3 | 53.2 | 51.0 | 62.1 |
| x-ai/grok-4.7 | 38.1 | 36.1 | 25.4 | 39.4 | 41.5 | 32.9 | 40.1 | 37.9 | 51.7 |
All 14 families · all categories Similarity / 100 · higher means closer
| Family | Overall | Adult topics | Ambiguity | Creative | Crisis support | Disclosure | Explanations | Politics | Math |
|---|---|---|---|---|---|---|---|---|---|
| deepseek | 83.2 | 72.4 | 84.0 | 84.4 | 85.4 | 83.2 | 85.3 | 78.8 | 92.1 |
| qwen-3.8 | 65.0 | 56.7 | 65.1 | 61.1 | 61.9 | 61.8 | 71.2 | 59.3 | 83.0 |
| claude | 59.8 | 55.4 | 44.0 | 74.2 | 61.0 | 46.1 | 65.2 | 65.7 | 66.5 |
| hunyuan | 58.6 | 55.7 | 47.6 | 68.8 | 58.1 | 42.0 | 58.5 | 56.6 | 81.6 |
| gpt | 57.3 | 59.6 | 41.2 | 63.2 | 56.8 | 45.7 | 65.1 | 67.1 | 59.4 |
| kimi | 56.8 | 51.4 | 52.0 | 52.5 | 56.9 | 62.5 | 60.7 | 58.3 | 60.4 |
| minimax | 56.2 | 59.7 | 48.7 | 68.2 | 58.6 | 46.8 | 52.0 | 49.5 | 66.3 |
| glm | 55.5 | 50.8 | 52.6 | 52.2 | 56.4 | 53.8 | 61.0 | 57.4 | 59.7 |
| mistral | 55.1 | 50.2 | 49.0 | 62.8 | 57.6 | 52.2 | 51.6 | 55.0 | 62.4 |
| gemini | 52.1 | 52.1 | 35.9 | 61.5 | 43.1 | 51.0 | 64.1 | 60.2 | 48.9 |
| mimo | 51.4 | 43.3 | 49.1 | 46.2 | 52.8 | 50.8 | 52.4 | 59.3 | 56.9 |
| nemotron | 50.5 | 52.2 | 41.2 | 65.5 | 50.8 | 36.4 | 46.4 | 50.0 | 61.1 |
| gpt-oss | 46.2 | 42.5 | 36.1 | 53.5 | 39.9 | 31.3 | 53.2 | 51.0 | 62.1 |
| grok | 38.1 | 36.1 | 25.4 | 39.4 | 41.5 | 32.9 | 40.1 | 37.9 | 51.7 |
Study scope
Stealth Models’ reasoning comparison, recorded 29 September 2026. This comparison examines opening words, restating a question, handling sensitive requests, planning an answer and recurring phrases. Higher scores mean more similar habits, rather than better answers or a shared developer.
The study includes 142 of 144 Space Bunny samples. Both DeepSeek comparisons cover all 48 questions.
What drives the DeepSeek lead? Shared “We need answer” openings and clipped “Need…” planning are the clearest resemblance. Bunny usually paraphrases the question; DeepSeek more often quotes it. Common opening patterns and phrase tags contribute heavily to this score.
This study measures the exposed reasoning text. MiniMax’s reveal settles the identity; the scores remain a record of shared habits.
BACK TO THE DRAWING BOARD · 29 SEP 2026
Who draws like Bunny?
Check the whole lineup.
Bunny’s 29 September drawings look closest to GPT 6.1 Sol’s; our 27 September M3.1 Flash test drawings look different. Same subjects, different task instructions. Drawing style pointed us the wrong way. Benchmark tasks appear below each pair.
That arm could fetcha carrot from Earth.
A selfie worth leaving orbit for.
GPT 6.1 Sol: Glossy brown eyes, lilac suit fittings and layered helmet reflections bring these September 29 drawings closer together.
Qwen: The ivory suit and lavender moon feel familiar, but Qwen gives the rabbit a much longer arm, exposed ears and soft glow effects. Bunny’s compact proportions and helmet reflections have a different finish.
DeepSeek: Bunny’s helmet encloses its ears, the paw reaches the carrot, and the moon reads as a complete globe. DeepSeek’s ears protrude through the helmet, with a flatter suit and a very different pose.
MiniMax: Same carrot mission, different touch. Bunny’s expressive face, glassy helmet and layered suit have more care in the details.
Benchmark task
Space Bunny · before 29 Sep
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
Space Bunny · 29 Sep 2026
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
Qwen3.8 Flash
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
DeepSeek V4 Pro
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
GPT 6.1 Sol · 29 Sep 2026
An astronaut rabbit reaches for a carrot above a small moon while a younger companion peeks into view. The scene tests expressive characters, spatial relationships and material detail.
Mona Lisa isn’t telling.
Qwen: Qwen builds a recognizable portrait, closer in structure than DeepSeek’s. Heavy mottled texture and blurred shading cover the face and hands; Bunny uses crisp features, separate fingers and clean dress folds.
DeepSeek: The clearest gap. Bunny draws a continuous figure with crossed hands and dress folds. DeepSeek’s face, torso and arms break into disconnected shapes.
MiniMax: Bunny gives her a more defined face, separate fingers and folds in her dress. Even a secretive smile needs good anatomy.
Benchmark task
Space Bunny · before 29 Sep
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
Space Bunny · 29 Sep 2026
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
Qwen3.8 Flash
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
DeepSeek V4 Pro
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
GPT 6.1 Sol · 29 Sep 2026
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
Two wheels. Two very different birds.
Qwen: The sunny cartoon cyclist is a closer visual neighbor. Qwen adds a scarf, fish basket and glossy gradients; Bunny’s outlined coastal scene and bird proportions still have a distinct hand.
DeepSeek: Both draw a recognizable pelican cyclist. Bunny uses clean outlines and a complete coastal scene; DeepSeek uses glossy shading, a dark backdrop and a busier chain assembly.
No result for this subject.
Benchmark task
Space Bunny · before 29 Sep
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
Space Bunny · 29 Sep 2026
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
Qwen3.8 Flash
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
DeepSeek V4 Pro
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
GPT 6.1 Sol · 29 Sep 2026
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
The dog is driving. Naturally.
Qwen: Qwen gives both riders goggles and rounds the dog into a low, stretched pose. Bunny’s upright driver, clear handlebar grip and layered spray make a different scene from the same unusual request.
DeepSeek: Bunny’s riders, handlebars and wake form a clear scene. DeepSeek’s huge raccoon tail dominates the picture, while the dog, controls and jet ski overlap awkwardly.
MiniMax: Bunny makes the dog’s grip, the raccoon’s perch and the spray easier to read. An improbable journey, convincingly drawn.
Benchmark task
Space Bunny · before 29 Sep
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
Space Bunny · 29 Sep 2026
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
Qwen3.8 Flash
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
DeepSeek V4 Pro
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
GPT 6.1 Sol · 29 Sep 2026
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
Same holiday. Different postcards.
Qwen: Qwen fills the beach with silhouetted people, torches, a campfire and small animals. Bunny puts expressive animals in the foreground, with a softer palette and more defined waves.
DeepSeek: Bunny fills the shore with expressive animals, outlined waves and reflections. DeepSeek places tiny animals across a broad, mostly empty beach beneath a glowing sunset.
MiniMax: Bunny’s Max beach keeps the playful animals and soft colors of its earlier trip. Flash goes for dark silhouettes and a dramatic sky: two very different ideas of shore leave.
Benchmark task
Space Bunny · before 29 Sep
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
Space Bunny · 29 Sep 2026
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
Qwen3.8 Flash
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
DeepSeek V4 Pro
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
GPT 6.1 Sol · 29 Sep 2026
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
Qwen: closer, but no match. These drawings add no support to a close Qwen relationship. Qwen3.8 Flash’s style still differs, and the proposed Qwen4 Flash connection loses ground.
DeepSeek: the pictures pull the other way. Its weaker anatomy and different visual style weigh against an exact match, despite the close reasoning signature.
MiniMax: the family was right. Our four 27 September comparisons made an exact Flash match look unlikely. MiniMax calls Bunny an early version of M3.1 Flash, which fits a different checkpoint; the task instructions differ, so the drawings cannot prove it.
GPT 6.1 Sol: the closest drawings. The raccoon and beach scenes show the strongest resemblance; the selfie shares glossy eyes and reflective glass. The Mona Lisa differs more substantially. Similarity varies by subject, and the task instructions differ. MiniMax’s reveal shows that style resemblance was not shared identity.
Two dates, one codename. The 29 September drawings followed the behavior change our researchers observed that day. MiniMax has not said whether that change was a new stealth checkpoint.
04 PUT IT TO WORK
What can
Space Bunny use?
The stealth preview took screenshots and video, called your app’s functions and let you pick a reasoning effort.
Build your first request ↓Ask about what you can see.
Send a screenshot, diagram or video alongside your question. Space Bunny’s preview accepted text, images and video, and replied in text.
Connect it to your tools.
Your application can offer functions, such as finding a file or looking up a record. The model could request the right function and supply its arguments.
Try the release in OpenCode.
Choose MiniMax M3.1 Flash in OpenCode Go, or use the command builder below to start a coding task from your terminal.
Preview specifications and model details
The OpenRouter stealth preview supported JSON response format, without JSON Schema enforcement. MiniMax says M3.1 Flash weights will be released later; architecture, parameter count and training details remain undisclosed. MiniMax’s launch post ↗
DEMODOKOS FOUNDRY / MUSIC · VOICE · VIDEOAdvertisement
A STUDIO ON YOUR PC
You’re the
sound department.
Choose how each line is spoken. Add original music, expressive voices, translation and subtitles in the same local studio.
Play the scene0:00--:--
Scene 02 · Demodokos demo
Create with Demodokos Windows · Your GPU
05 THE LAUNCHPAD
Space Bunny is back
as M3.1 Flash.
The stealth routes have closed. MiniMax M3.1 Flash is on the MiniMax API at $0.15 per million input tokens and $0.60 per million output tokens.
MiniMax API
AvailableMiniMax-M3.1-Flash$0.15 input, $0.60 output and $0.015 cached input per million tokens. Flat pricing. MiniMax also offers M3.1 Flash in M Plan and MiniMax Code.
MiniMax pricing ↗OpenRouter
Stealth route endedstealth/space-bunny-alphaSpace Bunny hopped off OpenRouter on . On 19 October OpenRouter brought it back as MiniMax’s M3.1 Flash, free for 72 hours.
OpenRouter’s reveal post ↗OpenCode Go
Go planopencode-go/minimax-m3.1-flashSpace Bunny Free has left OpenCode. Its free MiniMax M3.1 Flash route has ended. OpenCode Go lists the model on a paid plan from $10 a month.
OpenCode Go ↗COMMAND BUILDER
Install OpenCode first. These examples use its CLI to reach MiniMax M3.1 Flash on a Go plan.
OpenCode CLI guide ↗opencode run --model opencode-go/minimax-m3.1-flash -- 'Write a Python function that validates SRT timestamp order. Include three unit tests.'Runs through OpenCode on your computer.
Access and prices checked . Read MiniMax’s launch post ↗
06 BEFORE YOU HOP
What to know
about Space Bunny.
What is Space Bunny Alpha? ＋
Space Bunny Alpha, also written space-bunny-alpha, was an anonymous AI model listed on OpenRouter on 23 September 2026. On 19 October 2026, MiniMax revealed it as an early version of MiniMax M3.1 Flash. The stealth preview accepted text, images and video and returned text, including code and tool requests. OpenCode listed it as Space Bunny Free.
Who made Space Bunny Alpha? ＋
MiniMax. Its 19 October 2026 launch post says: “Space Bunny 🐰, which processed more than 200 trillion tokens while operating anonymously, was an early version of MiniMax M3.1 Flash.” Our token-count study had already matched Space Bunny to MiniMax M3 on all 24 test phrases.
Is Space Bunny the same as MiniMax M3.1 Flash? ＋
It was an early version of it. MiniMax calls Space Bunny “an early version of MiniMax M3.1 Flash,” which it released on 19 October 2026. MiniMax has not said whether the stealth model matches the M3.1 Flash Preview that MiniMax Code listed on 27 September, or the final release. Our benchmark results, SVG outputs and speed measurements all come from the stealth model. See what we measured.
What are the space-bunny-alpha benchmark scores? ＋
Space Bunny Alpha scores 7.0/10 on AI BENCHY at high reasoning, 82.0% on our 60-question GPQA Diamond subset, 75% on MMLU-Pro in our evaluation and 46.1% on our 300-question subset of original HLE (approximate 95% CI: 40.4–51.8%). Browse the space-bunny-alpha AI BENCHY comparison, GPQA Diamond results, MMLU-Pro results and space-bunny-alpha HLE results for the model comparisons. All four were measured on the stealth model, MiniMax’s early M3.1 Flash.
Is Space Bunny Alpha good? ＋
Yes, especially for creative SVG work. Space Bunny Alpha’s SVG gallery shows expressive characters and thoughtful composition. In the SVG identity comparison, its faces, poses and material details are more developed in the portrait, selfie and jet-ski scenes. Its Max-reasoning beach rerun keeps the playful style of the earlier result. Explore the AI BENCHY and GPQA Diamond results for scored comparisons.
Is Space Bunny Alpha an OpenAI model? ＋
No. MiniMax made it. Our 30 September assessment proposed a router built around GPT-6.1 Sol, based on Harmony-style channel names and compressed developer turns in our probes. MiniMax’s reveal supersedes that assessment.
Did Space Bunny Alpha change after DevDay? ＋
Our researchers saw a change. Before DevDay, Bunny’s visual style closely resembled GPT-6 Sol’s. Within hours of OpenAI releasing GPT-6.1 Sol, Bunny’s behavior changed abruptly and its visual style became strikingly similar to Sol 6.1’s. We inferred a reasoning-core upgrade; MiniMax’s reveal supersedes that inference. MiniMax has not said whether it switched stealth checkpoints that day. See the dated timeline.
What do Juice and summary reveal about Space Bunny Alpha? ＋
Less than we thought. Our 30 September probe returned Harmony’s analysis, commentary and final channels, a fourth summary channel and a # Juice marker. We read # Juice as an OpenAI-specific reasoning-budget marker and made these clues central to our GPT router assessment. MiniMax’s reveal supersedes that reading; its statement does not explain why the markers appeared.
Does the 83.8 DeepSeek match identify Space Bunny Alpha? ＋
No. It measures similarity in visible reasoning habits across 48 prompts: DeepSeek V4 Pro scored 83.8 and V4.1 Flash 82.6. Space Bunny was MiniMax’s early M3.1 Flash, so the match shows shared habits, not shared identity. See the DeepSeek evidence.
Did the tokenizer study point to MiniMax? ＋
Yes. Our token-count comparison matched MiniMax M3 on all 24 reported input-token increments; GPT 6.1 Sol matched 7. From 30 September we gave runtime clues more weight. MiniMax’s reveal shows the token counts pointed the right way.
Was Space Bunny Alpha a router? ＋
MiniMax names one model: Space Bunny was an early version of MiniMax M3.1 Flash. Our earlier router assessment, with GPT reasoning, MiniMax-type input counts and a DeepSeek-style reasoning display, is superseded.
Why does Space Bunny Alpha forget its own name? ＋
In our 30 September probe, earlier conversation turns appeared as compressed developer messages, and these summaries dropped the Space Bunny name within a few turns. That explains why later self-identification could differ from the opening turns.
Is Space Bunny Alpha free? ＋
Not as Space Bunny: its stealth routes have closed. MiniMax M3.1 Flash costs $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens on the MiniMax API. On 19 October OpenRouter brought it back free for 72 hours, a window that has closed, and OpenCode’s free route has ended. OpenCode Go lists it on a paid plan.
What is Space Bunny Alpha’s context window? ＋
The OpenRouter stealth preview listed 1,000,000 context tokens and a separate maximum-completion ceiling of 524,288 tokens. Reasoning counted toward completion tokens. These are the preview’s listed capacities; check MiniMax’s documentation for M3.1 Flash limits.
Where can I use Space Bunny Alpha? ＋
Use its release, MiniMax M3.1 Flash, through the MiniMax API (MiniMax-M3.1-Flash), M Plan, MiniMax Code or an OpenCode Go plan. The access guide has commands and links.
Can Space Bunny Alpha generate audio? ＋
Space Bunny Alpha returns text, including code. Demodokos is a separate local studio for music, speech and dubbing. Listen to the demo.
Can I turn off Space Bunny Alpha’s reasoning? ＋
The OpenRouter stealth preview required reasoning and offered low, medium, high, xhigh and max efforts, with max as its default. Hiding reasoning text did not disable reasoning computation.
Why do Space Bunny Alpha speed benchmarks differ? ＋
The speed summary shows OpenRouter’s median dashboard figures from 24 September 2026. The expandable OpenCode comparison uses TokenDyno’s 24-hour averages from a fixed writing task. Different routes, workloads and timing definitions produce different numbers; each result keeps its source and window attached. All of them measure the stealth model on its preview routes.
THE FLIGHT LOG
Trace the signal.
01 / REVEALMiniMaxM3.1 Flash launch post ↗02 / PRICINGMiniMax APIM3.1 Flash rates ↗03 / SPEEDTokenDynoService tests & methodology ↗04 / OUR RESEARCHTokenizer study24 phrases · 28 models · pointed to MiniMax ↓
MiniMax revealed Space Bunny as an early version of MiniMax M3.1 Flash. Our 24/24 token-count match with MiniMax M3 pointed the right way; the GPT router assessment is superseded. M3.1 Flash is on the MiniMax API, OpenRouter brought it back free for 72 hours, and Space Bunny Free left OpenCode.
Free access continued on OpenCode for a few more days. OpenCode is sponsoring the extension while Bunny is tuned ahead of its reveal. OpenRouter access has ended.
OpenRouter announced removal for 5 October 2026. OpenCode remained listed, with provisional access.
Our assessment: a model fusion router with a GPT-6.1 Sol or Astra core. Harmony-style channel and context-compression findings led the evidence. Superseded by MiniMax’s 19 October reveal.
Before DevDay, Bunny’s visual style closely resembled GPT-6 Sol’s. Within hours of Sol 6.1’s release, its behavior changed abruptly and its visual style became strikingly similar to Sol 6.1’s. We read it as a reasoning-core upgrade, an inference the reveal superseded. Updated SVG benchmark results for Bunny and GPT 6.1 Sol. The earlier Bunny count test matches MiniMax on 24/24 phrases and Sol on 7/24. DeepSeek leads the earlier reasoning study. OpenCode’s preview was extended until at least 5 October.
Our M3.1 Flash SVG comparison weighed against an exact match. MiniMax later called Bunny an early version of M3.1 Flash.
OpenAI / GPT-OSS and MiniMax identity clues, model specifications, speed measurements and five SVG experiments.