On the Parallel Search (fast) search-efficiency leaderboard, MiniMax M3.1 Flash scored 68.8 at a cost of $43.8 per 1,000 tasks, close to DeepSeek V4.1 Flash (1017)'s 68.9, at about 35% of the latter's cost. The result depends on the Parallel Search tool and cannot be treated as a general measure of the model API's capabilities.
Tasks it can help assess: Web research using Parallel Search (fast), and trade-offs between answer quality and cost.
Tasks it should not be extrapolated to: The model's standalone capabilities outside Parallel Search; general question answering; the cost, latency, or production stability of other search tools or API endpoints.
Applicable model version: MiniMax M3.1 Flash; no more specific model snapshot or API ID was disclosed.
Test environment or client: Parallel Web Systems' Search Efficiency Leaderboard; the post explicitly says it used Parallel Search (fast).
Reasoning tier and parameters: Not specified.
The original post defines the evaluation as a comparison of search-intelligence scores and costs: cost is the amount required to complete 1,000 tasks, and the chart's x-axis uses a logarithmic scale. Full-color dots represent the top 10 search-intelligence scores and the Pareto frontier. A supplemental post in the same thread defines “efficiency” as how cost-effectively a model obtains good answers relative to the other models on the leaderboard, and says that M3.1 Flash costs less than half as much as the second-best model.
The attached chart is titled Search Efficiency Leaderboard, and its header shows 10 of 24 · 48 models. The post does not explain the relationship between these counts, the task set, sample size, evaluator, aggregation formula, or number of repetitions, so the scores cannot be recomputed from this post alone.
| Model | Cost per 1,000 tasks | Search Intelligence Score | Leaderboard position |
|---|---|---|---|
| MiniMax M3.1 Flash | $43.8 | 68.8 | 1st |
| Hy4 Preview | $93.2 | 64.6 | 2nd |
| Pareto 26.10 | $116 | 68.6 | 3rd |
| DeepSeek V4.1 Flash (1017) | $125 | 68.9 | 4th |
| GPT-6.1 Sol | $130 | 70.4 | 6th |
The original post says that M3.1 Flash and DeepSeek V4.1 Flash (1017) deliver similar web-research performance, with the former costing roughly one-third as much. The exact values in the attached image give 125 ÷ 43.8 ≈ 2.85; M3.1 Flash also costs less than half as much as the second-place Hy4 Preview (43.8 < 93.2 ÷ 2).
M3.1 Flash: 68.8 points, $43.8 per 1,000 tasks.
DeepSeek V4.1 Flash (1017): 68.9 points, $125 per 1,000 tasks.
Other readable values in the attached image: Pareto 26.10 scored 68.6 at $116; Gemini 3.7 Flash scored 66.8 at $130; Mistral Large 4.0 scored 64.4 at $136; Muse Spark 1.3 scored 66.1 at $142; Muse Spark Open scored 67.2 at $171; Gemini 3.8 Flash scored 65.1 at $177.
Converting the screenshot's units, M3.1 Flash costs approximately $43.8 ÷ 1,000 = $0.0438/task; this is an arithmetic conversion, not a separately disclosed billing figure from the post.
The limited conclusion supported by this source is that, in this snapshot of the Parallel Search (fast) leaderboard, MiniMax M3.1 Flash ranked 1st with a score of 68.8 at $43.8 per 1,000 tasks; its score was nearly identical to DeepSeek V4.1 Flash (1017)'s, while its cost was substantially lower.
The original post does not disclose the search questions, sample size, prompts, answer evaluator, model snapshot, API ID, reasoning tier, sampling parameters, tool version, concurrency or retries, token composition, caching, region, cost-accounting details, margin of error, or confidence interval. Parallel Search is part of the comparison environment, and the provider's tool may affect answer quality and cost. Therefore, the leaderboard cannot be extrapolated to a general ranking of M3.1 Flash on its native API, with other search tools, or on arbitrary business tasks.
Open https://x.com/p0/status/2112544138662981825, click “Show original” on the post detail page, and verify the body and timestamp.
Review the original post, supplemental posts in the same thread, and the attached image. Record the costs and scores for M3.1 Flash, DeepSeek V4.1 Flash (1017), and the other competitors shown in Search Efficiency Leaderboard; retain the original 10 of 24 · 48 models text without explaining what it means.
To run an independent test, obtain Parallel Search's task set, scoring rules, model snapshot, tool version, sampling details, and cost breakdown, and label the results as a separate experiment.
MiniMax M3.1 Flash
