MiniMax M3.1 Flash halves token prices, but agent task costs tell a different story

MiniMax unveiled M3.1 Flash on Monday with claims of a 70 percent reduction in agent operating costs. The per-token pricing has dropped to half its previous level, but whether that translates to lower bills for developers hinges on a question the company hasn't fully addressed: how many tokens does an agent actually need to complete a task?

The new model's rate card shows $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. MiniMax also reports inference speeds nearly three times faster than M3. The 70 percent savings figure deserves scrutiny, though—cheaper per-token rates only reduce bills if agents also require fewer iterations to finish work.

Research into token consumption during agentic coding tasks reveals that running the same task multiple times can consume up to 30 times as many tokens, with input tokens dominating the total. This variability makes it difficult to validate MiniMax's savings claim without knowing which specific workloads the company tested.

Artificial Analysis, an independent benchmarking firm, measured average cost per task at $0.22 for M3.1 Flash compared to $0.51 for M3—a 57 percent reduction. While this falls short of MiniMax's 70 percent claim, the discrepancy likely reflects different test workloads rather than conflicting data.

Coding and computer use

On the DeepSWE v1.1 benchmark, M3.1 Flash scores 73.7 percent, marginally below the 74 percent MiniMax cites for Claude Opus 5 and one point ahead of GPT-5.6 Sol at 73. VentureBeat's testing places DeepSeek V4.1 Flash 1017 at 77 on the same scale. Anthropic's newer Opus 5.5 does not appear in that comparison, so teams benchmarking against current Opus versions should run their own evaluations.

On Artificial Analysis's Intelligence Index, M3.1 Flash achieves a score of 42, tied with GLM-5.3 Flash, surpassing Gemini 3.8 Flash at 41 but trailing DeepSeek V4.1 Flash 1017 at 45.

Computer use is where the model looks weakest.

On the OSWorld V2 benchmark, M3.1 Flash scores 61.7, which MiniMax characterizes as leading among Chinese models. However, VentureBeat notes the model falls behind the strongest offerings from Anthropic and OpenAI in this category.

MiniMax reports an 82 percent score on CyberGym and states that internal security specialists rated the model as strong for code auditing, patching, penetration testing, malware analysis, and reverse engineering. The company's launch announcement does not disclose how these evaluations were conducted, so independent verification would be prudent before relying on these security assessments.

Confident answers, uncertain accuracy

Artificial Analysis's AA-Omniscience benchmark reveals a notable strength and weakness. M3.1 Flash answered 56 percent of factual questions correctly, outperforming DeepSeek V4.1 Flash at 43 percent. However, the model attempted nearly every question presented to it, with approximately 93 percent of incorrect responses being wrong answers rather than refusals to answer.

The model's overall Omniscience score was 16, which Artificial Analysis characterized as high accuracy paired with a high hallucination rate. External reviewers reached similar conclusions: The Decoder noted that M3.1 Flash attempts all test questions without declining when uncertain, suggesting that organizations deploying the model for knowledge-intensive applications should implement verification layers, retrieval systems, guardrails, or human review processes.

When DeepSeek's discount applies

DeepSeek represents the closest pricing comparison. Its V4.1 Flash model matches MiniMax's $0.15 input and $0.60 output rates during off-peak hours. Peak pricing applies from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays, doubling those rates. DeepSeek's cached input tokens cost $0.003 per million, compared to MiniMax's $0.015.

Teams operating in New York or San Francisco during business hours typically fall within DeepSeek's off-peak window. Overnight batch processing, however, would incur peak pricing. Since research shows input tokens drive most agent expenses, the cache rate differential may prove more significant than headline pricing.

Space Bunny and what's still unknown

Space Bunny, a previously unidentified model that processed over 200 trillion tokens, was later confirmed by MiniMax to be an early version of M3.1 Flash. The released model supports a one-million-token context window, enabling large codebases to fit within a single prompt. Developers who tested the preview should not assume the shipped version behaves identically, as models often change between preview and release stages.

Model weights are reportedly arriving on October 23, though MiniMax's official announcement only states "later." The license terms remain unpublished, a critical detail for anyone considering deployment on private infrastructure. A more capable Pro variant is also in development, with no release timeline announced.

A stronger Pro model is also on the way, with no date yet.