On Plotly's 43-question data analytics exam, MiniMax M3.1 Flash answered 41 questions correctly (95%) and found all 5 hidden patterns embedded in the data. At MiniMax's list prices, a complete exam cost $4.09, but the model's median time per question was 3 minutes, making it the slowest model on the leaderboard.
Tasks it can help assess: Data analytics, SQL, statistical-trap detection, data-quality checks, and analytical workflows that require proactively discovering patterns in data.
Tasks it should not be extrapolated to: General capabilities, accuracy on real business data, production stability, or speed and cost on other datasets or with other agent configurations.
Applicable model version: MiniMax M3.1 Flash, released on 2026-10-19; the article says Space Bunny Alpha was an earlier version of this model. The API model ID is not specified.
Test environment or client: Plotly's DataAnalyticsBench data analytics exam; the article says OpenRouter was offering M3.1 Flash for free at the time, while the cost was calculated using MiniMax's official list prices.
Reasoning tier and parameters: Not specified. The article does not disclose temperature, sampling, model snapshot, prompts, number of repetitions, harness, or tool configuration.
Plotly says it uses the same exam when new models are released and presents the results for 43 questions in the article. The dataset is a fictional wind-power operator's data warehouse containing 43 questions to answer. The exam remains private and is shared only with researchers to keep the answers out of training data. The questions include statistical traps, ambiguous questions, and requests that are "impossible to complete," and they embed 5 hidden patterns that the model must discover on its own.
The article compares model results on the same exam. MiniMax M3.1 Flash's 41/43 is tied with Claude Opus 5.5, Gemini 4 Argon, Muse Wonder, and GPT-6 Sol. The article says the next-highest scores from Chinese labs were achieved by GLM 5.3 Flash and Qwen 3.8 27B, with 38 questions correct each. It also says the complete leaderboard contains 33 models, but does not show all per-question results on this page.
| Metric | Space Bunny Alpha (September) | MiniMax M3.1 Flash (October) |
|---|---|---|
| Score | 38/43 (88%) | 41/43 (95%) |
| Hidden patterns found | 2/5 | 5/5 |
| Model calls per question | 15 | 45 |
| Tokens read per question | 0.7M | 4.2M |
| Tokens written per question | 8,500 | 32,000 |
| Median time per question | 36 seconds | 3 minutes |
| Complete exam cost (at MiniMax list prices) | $0.82 | $4.09 |
The article says M3.1 Flash made 3 times as many calls per question as Space Bunny, wrote nearly 4 times as many tokens, and that the additional reasoning work brought a higher score along with longer wait times and higher cost. M3.1 Flash also passed all the statistical traps, ambiguous questions, and impossible requests. For example, when determining whether windy weather caused SCADA alerts, it looked for a third factor that could plausibly explain both.
Claude Opus 5.5 and Gemini 4 Argon scored the same as M3.1 Flash; the article says their costs were approximately 3 times higher than M3.1 Flash's.
GPT-6 Sol answered 41 questions correctly at a cost of $2.61; GPT-6.1 Sol answered 43 questions correctly at the same cost as M3.1 Flash.
98% of M3.1 Flash's input hit the cache, calculated at MiniMax's cached-input price of one-tenth the normal input price. OpenRouter was offering the model for free at the time, so the article did not use an actual OpenRouter bill.
MiniMax official list prices: $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached-input tokens.
M3.1 Flash complete exam: 41/43, cost $4.09, approximately $0.10 per question; median time per question: 3 minutes.
Speed basis: The article says each model call ran at approximately 150 tokens/s; at MiniMax's claimed 170 tokens/s, writing 32,000 tokens alone would take approximately 3 minutes.
Two errors: One question asked for the highest 7-day rolling average of the daily energy from all wind turbines. The model's prose said it had divided by 7, but the SQL contained no division and treated the 1,320 GWh 7-day total as an average. On another question, it misclassified a normal pattern in the data as a data-pipeline error.
The article also says GPT-6.1 Sol, GPT-6 Astra, and Gemini 4 Argon found all 5 hidden patterns, while Space Bunny found only 2.
The conclusion supported by this test is that M3.1 Flash scored 41/43 and found all 5 hidden patterns on Plotly's private data analytics exam. Its main costs were a higher number of calls, greater token output, and longer wait times. The results are useful for comparing the tradeoff among analytics capability, cost, and speed on this exam.
The exam uses synthetic wind-power data and cannot represent real enterprise data. The page does not publish the questions, answers, complete prompts, model snapshot, reasoning tier, sampling parameters, tool permissions, harness, number of repetitions, or error range, so external readers cannot fully recalculate the score from the article alone. The cost is estimated from MiniMax's list prices and includes 98% cached input; it does not equal the actual bill for every account and request. Three minutes is the median per-question time on this exam and does not equal latency for other tasks or provider endpoints.
The article says MiniMax will release the M3.1 Flash weights later. As of the collection date, this remained a future plan and cannot be treated as evidence that the weights were already available.
For reproduction, fix the 43-question DataAnalyticsBench exam described in the original article, the model version, calling provider, cache-billing convention, and collection date. Record the number of model calls, input and output tokens, cache hits, elapsed time, and error type for each question separately. Because the exam and answers remain private, the 41/43 result cannot be independently reproduced from the public article alone; any new test should be labeled as a separate experiment.
MiniMax M3.1 Flash
