Artificial Analysis Independent Benchmark: A Snapshot of MiniMax-M3.1-Flash's Intelligence, Speed, and Cost

One-sentence takeaway

The current Artificial Analysis page gives MiniMax-M3.1-Flash the following snapshot: an Intelligence Index v4.3.2 score of 43, an output speed of 93.1 tokens/s, a TTFT of 1.39 s, and an Intelligence Index task cost of $0.22; the model ranks fourth among 173 comparable models for intelligence and 94th among 151 models for speed. These results are a snapshot under a specific evaluation version, reasoning-model version, and MiniMax API measurement conditions.

Use cases

• Tasks this can help assess: Intelligence, output speed, time to first token, task cost, and output token usage under the same Artificial Analysis evaluation methodology.

• Tasks this cannot be extrapolated to: Inferring all-task performance, other reasoning levels, other API providers, or real-world business latency and billing from the composite score alone.

• Applicable model version: The MiniMax-M3.1-Flash reasoning version listed on the page; the page lists its release date as 2026-10-19.

• Test environment or client: Independent testing by Artificial Analysis on dedicated hardware; the speed and pricing notes are based on the MiniMax API.

• Reasoning level and parameters: The page lists Reasoning: Yes but does not disclose sampling parameters, reasoning level, prompt, random seed, or number of repetitions.

Evaluation method

1. Intelligence Index v4.3.2: The page says this index contains 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. It defines the index as an independent measurement, with higher scores being better.

2. Speed: Measured in output tokens per second (tokens/s); the dataset description identifies the results as independent testing by Artificial Analysis on dedicated hardware. The FAQ also gives TTFT, the number of seconds until the first answer token is received, and says that thinking time for reasoning models is included in TTFT.

3. Cost: Cost per Intelligence Index task is the weighted-average dollar cost of running one Intelligence Index task, broken down by token type; Time per Intelligence Index Task on the page is the weighted-average decoding time per task, excluding TTFT and additional overhead.

4. Pricing and snapshot: Input and output prices are stated per million tokens, and the page displays pricing for the MiniMax API; the page's data and rankings change as evaluations, prices, and the model set are updated.

Key results

The page summary describes this model as leading in intelligence, moderately priced, below average in speed, and prone to longer outputs; this summarizes the models listed on the page and is not a conclusion about all tasks.

Raw data

• Model identifier: The Artificial Analysis page slug is minimax-m3-1-flash, and the page name is MiniMax-M3.1-Flash. The page does not display the provider's API model ID; this article therefore does not add an API ID that does not appear on the page as a fact.

• Model specifications: The page lists Reasoning: Yes; input supports text and image, while output supports text; the context window is 1M tokens.

• Index composition: The 10 evaluation names in Intelligence Index v4.3.2 are listed under “Evaluation method.” The page's model summary does not provide the prompt for each item, per-question traces, number of repetitions, or error range.

• Speed definition: Output speed is measured in tokens/s; TTFT is the number of seconds until the first answer token is received and includes thinking time for reasoning models.

• Cost definition: Task cost is the weighted-average USD cost of running an Intelligence Index task; the page also provides a cost breakdown by token type, including input, output, cache hit, and cache write tokens.

• Pricing: The page FAQ explicitly states $0.15/1M input tokens and $0.60/1M output tokens; using a 7:2:1 cache hit/input/output mix gives $0.10/1M tokens, and the page warns that pricing may change with the provider.

• Measurement method: The page's JSON-LD dataset description labels the Intelligence, Speed, Cost per Task, Output Tokens, and Latency results as independent testing by Artificial Analysis on dedicated hardware.

Conclusions and limitations

• In the page snapshot collected for this article, MiniMax-M3.1-Flash has a relatively high composite intelligence score and ranking; its output speed of 93.1 tokens/s is below the 114.2 tokens/s median given for the similar-price tier, while its TTFT of 1.39 s is below that tier's 1.66 s median.

• The reported $0.22/Intelligence Index task is a weighted average under the institution's evaluation task distribution and cannot be treated as the cost of any arbitrary business request. Actual cost depends on input and output length, cache hits, retries, and provider pricing.

• The 180M output tokens are related to the page's description of the model as “very verbose”; this is the aggregate token usage generated by the index evaluation and does not mean that every real task will produce the same length of output.

• The Intelligence Index aggregates 10 evaluations. The page does not provide the model's complete score for each sub-evaluation, so a score of 43 cannot establish that it necessarily leads in any particular capability.

• This page is a dynamic leaderboard. The ranking denominators, price medians, and comparison model set may change; the figures in this article represent only the page snapshot collected on 2026-10-21.

• The page does not disclose the complete prompts, sampling parameters, number of repetitions, error ranges, hardware configuration, or provider deployment details, so external readers cannot fully recompute the Intelligence Index from it.

Reproduction notes

1. Open https://artificialanalysis.ai/models/minimax-m3-1-flash and confirm that the page canonical, model name, and page slug all correspond to MiniMax-M3.1-Flash.

2. Record the collection date, model release date, Intelligence Index version, its 10 component evaluations, reasoning flag, context window, and input/output modalities.

3. Record the Intelligence Index, speed, TTFT, task cost, output token usage, pricing, and rankings shown in the page summary, preserving the units and displayed precision.

4. For any later comparison, fix the same page version, model configuration, provider, and pricing date; do not directly combine a dynamic leaderboard snapshot with scores from another harness or version.