DeepSeek V4.1 Flash vs MiniMax M3.1 Flash: Which Is Cheaper per Fixed Bug?

Two low-cost “Flash” models launched within two days of each other. DeepSeek V4.1 Flash 1017 arrived on October 17, 2026 with open MIT weights. MiniMax M3.1 Flash, the model OpenRouter users knew as Space Bunny Alpha, followed on October 19. On paper MiniMax is half the price per token. We wanted to know what that means per task.

So we ran both as coding agents on the same bug-fixing task, three runs each, next to Claude Haiku 5.5, GLM 5.3 Flash, and Claude Sonnet 5.5. For the other end of the price list, see our Claude Fable 5.1 benchmarks and pricing.

Which is better for coding, DeepSeek V4.1 Flash or MiniMax M3.1 Flash?

On our task, neither: both passed all 32 hidden tests in every run. DeepSeek was faster, 27 seconds and 14 steps against 48 seconds and 17 steps for MiniMax at low effort. On the two benchmarks both vendors publish, DeepSeek scores higher, though the harnesses may differ.

How we tested

We used the same task as in our Claude Sonnet 5.5 test: a small Python package with five planted bugs and two missing features (a settlement function and a command-line tool), plus 32 hidden tests the agent never sees. The prompt is one line: read the task file, complete it, run the tests until they pass. Each setup ran three times in Atomic Agent v0.6.9 through OpenRouter on October 19, 2026, with approvals off.

We pinned each model to one provider: Together for DeepSeek, which charges the same as DeepSeek’s peak price, MiniMax for MiniMax, and Z.AI for GLM. MiniMax ran twice, once at its default max effort and once with OpenRouter’s reasoning.effort: low. Everything else used default settings.

Results

Three runs per setup, averages shown. Token counts are the provider’s native counts from OpenRouter’s generation records. MiniMax ran on its free OpenRouter route, so its cost is estimated from those token counts at MiniMax’s list price; all other costs are what OpenRouter billed.

Every setup solved the task every time, so this table doesn’t rank intelligence. It shows how much work each model does to get there, and that work is what you pay for.

Limits of this test. One small, well-specified Python task, three runs per setup, on one day. It shows cost and speed on a task both models can solve. It can’t tell you which model handles hard, long, messy work better. For that, the vendor benchmarks below are the only data so far, and neither vendor’s numbers are independent.

Which is cheaper per task?

The one that’s cheaper at the moment you run it. At low effort both models spent almost the same tokens on our task, on average about 316,000 (DeepSeek) and 376,000 (MiniMax) input tokens per run, most of it cached, and 5,500 to 6,400 output. So the bill follows the price list: MiniMax ~$0.012, DeepSeek $0.016 at peak, DeepSeek ~$0.008 off-peak.

Here is the arithmetic. We priced each model’s actual token counts at each price list:

Prices per million tokens for uncached input, cached input, and output. Bold marks the price each model actually runs at.

The headline “half the price” is true for uncached input and output. It isn’t true for cached input, where MiniMax charges $0.015 per million against DeepSeek’s $0.006 at peak and $0.003 off-peak. In an agent loop the conversation is resent on every step, so 93 to 94% of input tokens were cache hits for both models, and cached input made up about 44% of MiniMax’s estimated low-effort bill.

DeepSeek’s peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. That’s about 21% of the week. The rest, including weekends, is billed at the off-peak rate on DeepSeek’s own API. Through OpenRouter every listed price is flat, including DeepSeek’s own endpoint there; Together and Fireworks charge DeepSeek’s peak rate around the clock. The off-peak discount needs DeepSeek’s own API.

At scale, for 1,000 tasks like ours: about $8 for DeepSeek off-peak, $12 for MiniMax at low effort, $16 for DeepSeek at peak, and $86 for MiniMax at its default max effort.

Which is faster?

DeepSeek. It finished our task in 26 to 28 seconds per run. MiniMax at low effort took 32 to 62 seconds, and at its default max effort 341 to 771 seconds. The effort setting matters more for MiniMax than the choice between the two models.

MiniMax quotes fast generation. It says M3.1 Flash reaches “200+ TPS” with a median above 170 tokens per second. On our task, though, wall-clock time was driven by the number of steps and how long the model reasoned before each one, not by raw generation speed. MiniMax at max effort produced about 14 times more output than at low effort, most of it reasoning.

Which one gets stuck in loops?

MiniMax at max effort, in our runs. Atomic Agent’s loop detector logged a “no-progress loop” in all three MiniMax max-effort runs: the model kept repeating the same shell and test commands. It fired in none of the DeepSeek runs and none of the MiniMax low-effort runs.

The same warning also appeared in two of three runs for MiniMax M3 and for GLM 5.3 Flash, so it’s a sign of long reasoning in general rather than a defect unique to M3.1 Flash. The fix on our task was simple: lower the effort.

We didn’t measure tool-call error rates reliably in this test, so we’re not reporting them. OpenRouter showed a 3.16% tool-call error rate for the stealth version of MiniMax in its first days, according to CellCog’s snapshot; we have no comparable figure for DeepSeek.

What do the vendor benchmarks say?

Only two benchmarks appear in both vendors’ own published numbers. DeepSeek reports 77.0 on DeepSWE v1.1 and 35.4 on Agents’ Last Exam; MiniMax reports 73.7 and 29.6. DeepSeek names its DeepSWE harness (Mini-SWE-V2); MiniMax doesn’t, so the comparison isn’t exact.

Vendor-published, from DeepSeek’s Hugging Face model card and MiniMax’s launch post. Not our tests.

Each vendor also reports benchmarks the other doesn’t. DeepSeek lists 42.3 on Terminal-Bench 4.0 and 60.5 on Automation-Bench v1.0.6. MiniMax lists 61.7 on OSWorld V2, 76.2 on BankerToolBench, and 82% on CyberGym, and says it ranks among the top performers on AutomationBench-AA but gives no score, so it can’t be compared with DeepSeek’s 60.5 on Automation-Bench v1.0.6. If you use third-party comparison tables, check which model version each number refers to.

Can you run either one locally?

Not on a typical machine. DeepSeek V4.1 Flash 1017 has open MIT weights today, but at 552B parameters community builds run to hundreds of gigabytes. MiniMax M3.1 Flash has no public weights yet; MiniMax says they’re coming, and CellCog reports October 23.

• DeepSeek. A 2-bit community build for the ds4 engine is a 341 GiB file that streams from SSD on a 128 GB Mac, with up to 20 tokens per second reported by its author. Mainline llama.cpp and local Ollama don’t support the model yet. Details are in our DeepSeek V4.1 Flash 1017 post.
• MiniMax. M3.1 Flash has about 428B total parameters, the same as M3, according to MiniMax’s model page. Judging by M3, a 4-bit build needs over 200 GB of memory.

For an agent that runs fully on hardware you already own, a mid-size open model is the practical choice; we tested one in Qwen 3.8 27B as a local agent.

Which should you use?

For routine coding agent work like our task, either model is a cheap default, and the cheaper one depends on when you run. Pick by the constraint that matters most to you:

• Lowest cost for batch jobs: DeepSeek V4.1 Flash, off-peak. About $0.008 per task in our test, and off-peak covers about 79% of the week: everything outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.
• Lowest cost during peak hours: MiniMax M3.1 Flash at low effort. About $0.012 per task. Set the effort yourself; the default max cost about 7 times more on our task.
• Fastest turnaround: DeepSeek V4.1 Flash. 27 seconds per task, the quickest of the two.
• Open weights today: DeepSeek V4.1 Flash. MIT-licensed and already on Hugging Face.
• Free to try right now: MiniMax M3.1 Flash on OpenRouter until about October 22, or DeepSeek through OpenCode’s free tier inside the OpenCode app.

You don’t have to choose once. In Atomic Agent, both models are a /model switch apart through OpenRouter, and DeepSeek also has a direct preset for its own API with off-peak pricing. That’s how we produced the table above: same tools, same prompt, different model. If you are still picking the agent itself, our open-source Claude Code alternatives list and OpenCode vs Claude Code comparison cover the main options.

How to reproduce this test

Pick a small task from your own codebase that has a clear finish line, ideally a failing test suite the agent has to make pass. Run it three times per model with the same prompt, and compare time, steps, and cost rather than a single run.

In Atomic Agent:

curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell, or download the desktop app. Connect OpenRouter in the first-run setup, then switch between deepseek/deepseek-v4.1-flash-1017 and minimax/minimax-m3.1-flash:free with /model. To run MiniMax at low effort, add "extraBody": { "reasoning": { "effort": "low" } } to the OpenRouter entry under llm.providers in ~/.atomic-agent/config.json. OpenRouter’s Logs page lists each request with its tokens and cost, so you can total a run yourself. On the free MiniMax route the cost shows as $0, so multiply the token counts by the list price.