MiniMax's M3.1 Flash Halves Token Prices, but Real Savings Depend on Agent Efficiency

MiniMax unveiled M3.1 Flash on Monday with claims of a 70 percent reduction in agent costs. Token pricing dropped to half its previous levels, yet the full picture of savings depends on how efficiently agents complete tasks—a metric the company has not disclosed its methodology for measuring.

The pricing structure lists $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, its predecessor. The 70 percent claim requires scrutiny; lower rates alone provide value only if agents also require fewer iterations to complete work.

Research into token consumption during agentic coding tasks revealed that identical tasks can consume up to 30 times more tokens across repeated runs, with input tokens representing the majority of usage. This variability makes it hard to validate MiniMax's 70 percent savings claim without knowing which specific workloads the company tested.

Artificial Analysis, an independent benchmarking firm, measured the average cost per Intelligence Index task at $0.22 for M3.1 Flash, compared to $0.51 for M3—a 57 percent reduction. While this trails MiniMax's stated 70 percent, the two figures reflect different test workloads.

Coding and computer use

On the DeepSWE v1.1 benchmark, M3.1 Flash achieved 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. According to VentureBeat, DeepSeek V4.1 Flash 1017 scored 77 on the same test. Anthropic's newer Opus 5.5 does not appear in that comparison, so teams evaluating against current Opus versions should run their own assessments. On Artificial Analysis's Intelligence Index, M3.1 Flash tied at 42 with GLM-5.3 Flash, ahead of Gemini 3.8 Flash at 41 but behind DeepSeek V4.1 Flash 1017 at 45.

Computer use is where the model looks weakest. The model scored 61.7 on OSWorld V2, which MiniMax claims leads Chinese models, though VentureBeat notes it lags behind the strongest offerings from Anthropic and OpenAI.

MiniMax reported 82 percent on CyberGym and stated that internal security experts assessed the model as strong in code auditing, patching, penetration testing, malware analysis, and reverse engineering. The company's launch announcement does not specify 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 revealed a notable finding: M3.1 Flash answered 56 percent of factual questions correctly, outperforming DeepSeek V4.1 Flash at 43 percent. However, the model attempted nearly all questions, with approximately 93 percent of incorrect responses being wrong answers rather than refusals to answer. Its overall Omniscience score was 16, reflecting what Artificial Analysis described as high accuracy paired with a high hallucination rate.

Independent analysts drew comparable conclusions. The Decoder observed that the model attempted every test question without declining when uncertain, suggesting that organizations deploying M3.1 Flash for knowledge-intensive applications should consider adding verification layers, retrieval mechanisms, guardrails, or human review.

When DeepSeek's discount applies

DeepSeek emerges as the natural competitor. Its V4.1 Flash pricing matches MiniMax's $0.15 input and $0.60 output rates during off-peak periods. Peak hours—01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays—double those rates. DeepSeek's cached input pricing stands at $0.003 per million tokens, substantially lower than MiniMax's $0.015.

For teams operating in New York or San Francisco during business hours, most requests fall within DeepSeek's off-peak window. Overnight batch processing, however, would incur peak pricing. Given that the earlier study found input tokens drive the majority of agent expenses, the cache rate differential may prove more significant than headline pricing.

Space Bunny and what's still unknown

Space Bunny, an unidentified model that processed more than 200 trillion tokens, was eventually confirmed by MiniMax to be an early iteration of M3.1 Flash. The released version supports a one-million-token context window, allowing large codebases to fit within a single prompt. Those who tested the preview should not assume its behavior matches the final release, as models often change between preview and production versions.

Model weights are expected to arrive on October 23, according to reports, though MiniMax's official announcement only states "later." The license remains unpublished—a critical detail for anyone considering local deployment. A more capable Pro variant is also in development, with no timeline announced. A stronger Pro model is also on the way, with no date yet.