Models
MiniMax and DeepSeek: Flash models slash the price of agentic coding
Within three days, DeepSeek and then MiniMax shipped "Flash" models that chase frontier coding scores for a few cents per task.
Between October 17 and 19, 2026, two Chinese labs released "Flash" models built for coding agents, with the same pitch: a much lower cost per task. Their list prices are now identical at off-peak rates, and their coding scores are closing in on frontier models.
MiniMax M3.1 Flash (Oct 19)
- Price: $0.15 per million input tokens, $0.60 output, $0.015 cached input (MiniMax).
- Coding: 73.7% on DeepSWE v1.1, according to MiniMax.
- Efficiency: nearly 3x faster inference and 70% lower usage cost on agentic tasks versus its predecessor, according to MiniMax, which claims more than 200 output tokens per second.
- Other uses: 61.7 on OSWorld V2 (computer use) and 82% on CyberGym (vulnerability discovery), again according to MiniMax.
- Availability: MiniMax API, M Plan and MiniMax Code; weights will be released later. An anonymous early version, "Space Bunny", had already processed more than 200 trillion tokens.
See the MiniMax M3.1 Flash page.
DeepSeek V4.1-Flash-1017 (Oct 17)
- Coding: 77.0% on DeepSWE v1.1, up from 74.2% for the previous version, and 42.3 on Terminal-Bench 4.0, up from 31.2 (Hugging Face). In DeepSeek's table, Claude Opus 5 scores 74.0% and GPT-5.6 Sol 73.0% on DeepSWE.
- Price: $0.15 per million input tokens and $0.60 output at off-peak rates, double at peak hours, with a 1-million-token context (DeepSeek).
- Cost per task: just under $0.20 per DeepSWE task on average, against about $0.50 for the previous version, per DeepSeek's published chart (The Decoder).
- Open: MIT-licensed weights on Hugging Face and ModelScope. DeepSeek is also opening its Search and Fetch APIs at $0.75 per 1,000 calls; calls made through DeepSeek Harness, Claude Code or Codex stay free until October 31 at 16:00 UTC (DeepSeek).
See the DeepSeek V4.1-Flash-1017 page.
What independent tests say
Artificial Analysis gives MiniMax M3.1 Flash 43 points on its Intelligence Index, against 29 for M3 (Artificial Analysis, page checked on October 20), and 45 to DeepSeek V4.1-Flash-1017 in Max mode (changelog). The same firm, however, measures 104.7 output tokens per second for MiniMax, far from the 200 claimed. More importantly, on the AA-Omniscience knowledge test, M3.1 Flash answers 56% of questions correctly, but when it does not know, it makes up an answer 93% of the time instead of admitting uncertainty (VentureBeat).
What it changes for developers
- Route tasks: send simple, well-tested tasks to a Flash model and keep a pricier model for ambiguous cases.
- Track cost per task, not just price per token: a verbose model spends more for the same result.
- Plug in an open agent: tools such as OpenCode accept these models through their APIs.
The limits
Most scores come from the vendors, under their own test conditions. In the enterprise, Chinese open models still account for only 1% of API usage, according to Menlo Ventures (see our article): compliance, data hosting and reliability still need checking. MiniMax has not yet given the licence terms for its future open weights, so self-hosting it remains an open question.
Most of these scores come from the vendors themselves: wait for independent tests before betting production on them.
Spotted an error in this article? Report it · Our corrections policy
