MiniMax M3.1 Flash stands out with its low-cost performance

Summary

MiniMax has launched M3.1 Flash, a 428-billion parameter Mixture-of-Experts (MoE) model. Priced at $0.15 per million input tokens, the model scored 62.2 on Automation-Bench, outperforming Opus 5 and GPT-5.6 Sol. However, it yielded competitive yet inconsistent results in coding benchmarks.

Why it matters

M3.1 Flash delivers near-frontier agentic performance at a cost significantly below the median of $0.24 per million input tokens, making it an attractive option for budget-conscious developers. The model’s 1-million-token context window and native multimodal input support (text, image, video) expand practical use cases. Conversely, it is noted to operate at below-average speeds and produce longer responses. Coding benchmarks like SWE Refactor Bench (23.0 points vs. Opus 5’s 31.0) reveal that the model is not superior across all domains.

Highlights

• Scored 62.2 on Automation-Bench at $0.15 per million input tokens, surpassing Opus 5 (50.3) and GPT-5.6 Sol (45.8).
• Mixed coding results: Achieved 73.7 on DeepSWE, but scored only 23.0 on SWE Refactor Bench compared to Opus 5’s 31.0.
• Logical reasoning remains consistently effective and offers five distinct effort levels. Sending thinking.type=disabled or effort=now parameters triggers errors, breaking existing integrations.

Critical take

The headline claims a 'high hallucination rate,' yet the article provides no data on hallucination metrics, rendering the claim entirely baseless. Furthermore, all benchmark scores are self-reported by MiniMax. The company’s own data contains internal inconsistencies (62.8 vs. 62.2 on Automation-Bench). The only independent cross-check is a single comparison against Kimi K3’s specification table.