Vals AI Task Benchmarks and Cost: A Capability Snapshot of MiniMax-M3.1-Flash

One-sentence takeaway

The current Vals AI page shows MiniMax-M3.1-Flash with an overall Vals Index score of 53.78% ± 1.07, ranked 15/23, and also lists seven component benchmarks. The page's update note says that the evaluation used Vals Index v2.2, the official MiniMax API, temperature 1, top_p 0.95, and max compute effort. The overall score describes only a snapshot of that index and cannot replace any individual task score.

Use cases

• Tasks it can help assess: Compare the published component results for agent, coding, finance, legal, tax, and terminal tasks under the same Vals AI page conventions, together with the cost and latency recorded on that page.

• Tasks it should not be extrapolated to: Treating the overall Vals Index score as general-purpose accuracy; extrapolating any component score to all business performance; directly combining different page update times, ranking denominators, or cost labels.

• Applicable model version: The page identifies it as MiniMax-M3.1-Flash, with a publication date of 2026-10-20; the page describes it as a MiniMax open-weight model.

• Test environment or client: The update note says the evaluation ran through the official MiniMax API, with MiniMax as the default provider. The page notes that some benchmarks may use different providers and parameters, but does not break this down for each benchmark.

• Reasoning tier and parameters: The update note discloses temperature 1, top_p 0.95, compute effort max, and 262k max output tokens; the hyperparameter panel separately shows top_k DEFAULT and max output tokens 524,288, as discussed in the limitations.

Evaluation method

The Vals AI update entry states that Vals Index v2.2 and its components were run on MiniMax-M3.1-Flash. The current page lists Vals Index, EMB v1.1, Finance Agent v2.1, Legal Research Bench, Tax Agent Bench, Vibe Code Bench v1.1, HLAB-Vals-1.1, and Terminal-Bench 4.0. The page reports accuracy as percentages and displays a ± value after most rows; it does not disclose the questions, sample size, or method used to calculate uncertainty for each item.

The current model card reports an overall Vals Index accuracy of 53.78% ± 1.07, a rank of 15/23, a Vals Index test cost of $0.660, and latency of 43 min 18 s. The page's update note instead gives 15/22, $4.46 per test, and an open-weight model rank of 5/11. Both sets of values are visible on the same page and may correspond to different update times or display conventions; this article preserves them separately rather than merging them into one figure.

Key results

Overall index

The update note additionally states: Vals Index 53.78%, rank 15/22, $4.46 per test, and open-weight model rank 5/11. This is text from the page's 2026-10-20 update and cannot override the current card or be added to it or converted into it.

Specific task benchmarks

The page's update text also gives comparisons with MiniMax-M3: Vibe Code Bench at 73.94% versus 47.57%, Tax Agent Bench at 32.14% versus 20.59%, EMB v1.1 at 67.84% versus 59.75%, Finance Agent v2.1 at 58.83% versus 54.34%, and Legal Research Bench at 34.13% versus 29.81%. These are score comparisons for the corresponding tasks, not differences in the overall index.

Raw data

• Model information: MiniMax-M3.1-Flash; developer: MiniMax; open-weight; 1M context window; model card maximum output: 524K tokens; the input modality is shown as supporting text and images.

• Pricing: The model card shows input/output token costs of $0.15/$0.60; the update note states $0.15 per million input tokens and $0.60 per million output tokens.

• Runtime parameters: The update note specifies the official MiniMax API, temperature 1, top_p 0.95, compute effort max, and 262k max output tokens. The expanded hyperparameter panel shows MiniMax as the default provider, temperature 1, top P 0.95, top K DEFAULT, max output tokens 524,288, and compute effort MAX.

• Version: The update note explicitly says Vals Index v2.2; component names and versions follow the titles visible on the page, such as EMB v1.1, Finance Agent v2.1, Vibe Code Bench v1.1, and HLAB-Vals-1.1.

• Page dynamism: The current model card shows Vals Index 15/23, while the update text shows 15/22; the ranking denominator, model set, and cost display may change as the page is updated. This article treats the content visible on the page on 2026-10-21 as a snapshot.

Conclusions and limitations

The limited conclusion supported by this page is that, in the public Vals AI snapshot, MiniMax-M3.1-Flash has its highest score on Vibe Code Bench v1.1 (73.94%), followed by EMB v1.1 at 67.84% and Finance Agent v2.1 at 58.83%; HLAB-Vals-1.1 is 7.50%, and Terminal-Bench 4.0 is 25.25%. These are component results for different tasks and cannot be combined into a single business success rate.

The page does not disclose the prompts for individual questions, task counts, random seeds, number of repeats, scoring rules, hardware and concurrency settings, the API or model snapshot, the actual provider and parameters for each benchmark, or the uncertainty calculation method. It also does not explain the difference between $0.660 and $4.46 per test. The page data can therefore be used to record relative positions and reproduction parameters within the same Vals snapshot, but cannot support a complete recalculation of costs or reproduction of all scores.

Reproduction notes

For reproduction, fix the original URL, collection date, MiniMax API provider, model identifier, Vals Index v2.2, temperature 1, top_p 0.95, and compute effort max, and record separately the update note's 262k max output tokens and the hyperparameter panel's 524,288. Record the question count, version, provider, parameters, elapsed time, input and output tokens, and cost for each benchmark separately. If the page still shows two sets of rankings or costs at the same time, retain their labels and snapshot times instead of deriving missing data.