MiniMax M3.1 Flash is a new option for coding and business agents. Its launch results put it ahead of DeepSeek V4.1 Flash-1017 on Automation-Bench, while DeepSeek leads on DeepSWE v1.1. MiniMax's flat input and output prices match DeepSeek's off-peak rates; DeepSeek has cheaper cache reads. Choose a trial workload around those differences before changing your default model.
DeepSeek released V4.1 Flash-1017 on October 17. MiniMax followed with M3.1 Flash on October 19. Together, the releases give developers more room to test lower-cost models on work they previously assigned to expensive ones.
The useful question is which tasks they can finish at an acceptable cost. A strong score on software engineering does not automatically predict success across office applications, and a lower token price does not always produce a lower bill.
Where does M3.1 Flash lead, and where does DeepSeek lead?
The reported results show a split: M3.1 Flash scores 62.8 on Automation-Bench v1.0.6 against 60.5 for DeepSeek V4.1 Flash-1017. On DeepSWE v1.1, DeepSeek scores 77.0 against MiniMax's 73.7.
| Benchmark | MiniMax M3.1 Flash | DeepSeek V4.1 Flash-1017 | Evaluation notes |
|---|---|---|---|
| Automation-Bench v1.0.6 | 62.8 | 60.5 | 600 public business-workflow tasks |
| DeepSWE v1.1 | 73.7% | 77.0% | Mini-SWE-agent 2.4.5; maximum context 512K |
| Terminal Bench 4.0 | 38.1 | 42.3 | Codex 0.149.1 harness |
| BankerToolBench | 76.2 | Not reported | OpenCode 1.17.18; MiniMax-M2.7 judge |
| LOCA-Bench 256K | 60.3 | Not reported | Official benchmark implementation |
These are reported results, not SaveMyToken measurements. Unreported results do not mean zero.
That makes both models worth testing for different jobs. Start with M3.1 Flash when the task spans business tools; keep DeepSeek in the comparison for repository changes and terminal work. The results support a shortlist, rather than a universal ranking of coding value.
View the complete launch table and methodology
What does the computer-use result actually show?
MiniMax reports 61.7 on OSWorld 2.0 and a 29.6% pass rate on Agents' Last Exam. Its ALE chart estimates a $111 total run cost for M3.1 Flash at max effort, using Claude Code, across the 152-task ALE-V1 Overall set.
The chart gives a reason to try it on a budget. It also shows that several models complete a larger share of tasks. A team choosing an agent for a long workflow still needs to measure failed attempts and the time spent repairing them.
Keep the DeepSeek versions separate here. The OSWorld comparison lists 54.3 for the earlier DeepSeek V4.1 Flash, with an internally reproduced result. It leaves the 1017 column blank. Those figures cannot establish a computer-use win over the October 17 release.
Is MiniMax M3.1 Flash cheaper than DeepSeek Flash?
For standard input and output, MiniMax matches DeepSeek's off-peak price and costs half as much as DeepSeek at peak times. For cached input, DeepSeek is cheaper in both periods.
| USD per 1 million tokens | MiniMax M3.1 Flash | DeepSeek Flash, off-peak | DeepSeek Flash, peak |
|---|---|---|---|
| Input, cache miss | $0.15 | $0.15 | $0.30 |
| Input, cache hit | $0.015 | $0.003 | $0.006 |
| Output | $0.60 | $0.60 | $1.20 |
USD API prices as of October 19, 2026. DeepSeek peak hours: 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. These rates do not describe subscription allowances or SaveMyToken plan prices.
A cache-heavy workload changes the comparison. Consider one month with 1 billion input tokens and 100 million output tokens. Hold token volumes and cache-hit rates equal so that the table isolates the listed prices.
| Input cache-hit rate | MiniMax monthly total | DeepSeek off-peak total | DeepSeek peak total |
|---|---|---|---|
| 0% | $210.00 | $210.00 | $420.00 |
| 90% | $88.50 | $77.70 | $155.40 |
| 95% | $81.75 | $70.35 | $140.70 |
Illustrative arithmetic, not a measured workload. Each total includes the assumed input and output; it excludes tool charges, taxes and other fees. The DeepSeek columns place all usage in the named period.
Monthly token cost = input millions × [(1 − hit rate) × input rate + hit rate × cache-read rate] + output millions × output rate
At peak prices, DeepSeek's input-only bill becomes lower once the cache-hit rate exceeds about 94.34%. Output shifts that crossover because MiniMax charges half the peak output rate. In the 95% example above, DeepSeek's cached input advantage still leaves its peak total higher. Compare the complete bill before treating either price as the winner.
How fast is it, and what happened to Space Bunny?
MiniMax identifies Space Bunny as an early M3.1 Flash version. It reports output above 200 tokens per second in everyday use, a median above 170, and nearly three times faster online inference after increasing sparsity. Those are vendor figures, rather than a latency guarantee for every request.
Output throughput is only part of waiting time. For an agent, record the delay before the first useful response, time spent thinking and calling tools, and elapsed time until the task passes its acceptance check.
Which model ID and settings should developers use?
Explore our Coding Plan to try MiniMax and DeepSeek Flash models.
The API model ID is MiniMax-M3.1-Flash. It supports a 1,000,000-token context window and text, image and video inputs. Thinking depth accepts low, medium, high, xhigh and max, with max as the default.
Use https://api.minimax.io/anthropic for Anthropic-compatible clients and https://api.minimax.io/v1 for OpenAI-compatible clients. DeepSeek's current API alias is deepseek-flash, serving V4.1 Flash-1017.
MiniMax announces access through its API, M Plan and MiniMax Code. Model weights are planned for a later release. Treat API access, subscription inclusion and downloadable weights as separate checks.
What do the creative demos add?
Reported creative examples include a folding-city web demo made through Codex, with a fold slider, night mode and a movable 3D view. Developer AJ also reportedly generated a pagoda demo in one API call with about 50,000 tokens and no tools or framework.
These examples show the kinds of visual tasks people are trying. They do not establish a success rate, code quality or the amount of manual review needed. SaveMyToken has not reproduced either demo.
What should you test before switching models?
Use a small batch of work you already know how to judge. Include a repository fix and a workflow across the tools you use. Give both models the same files, tool access and completion criteria.
- Record the model version, harness and effort setting. Try a lower effort level for simple instructions before accepting the default.
- Count accepted completions, including retries. Keep a note of any manual repair.
- Capture actual cache hits, input and output tokens, tool charges and total elapsed time.
- Compare cost per accepted result, then decide which tasks to route to each model.
M3.1 Flash has a promising case for business automation and predictable token rates. DeepSeek remains competitive for coding, with especially low cache-read prices. Your own batch can show whether either advantage survives the rest of the workflow.
Frequently asked questions
Is MiniMax M3.1 Flash cheaper than DeepSeek V4.1 Flash-1017?
Its standard input and output rates match DeepSeek's off-peak rates and are half DeepSeek's peak rates. DeepSeek charges less for cached input, so the total depends on cache hits, output volume and the time of use.
Is M3.1 Flash the Space Bunny model?
Space Bunny was an early version of M3.1 Flash. That does not establish that every anonymous preview used the final release.
Can I download M3.1 Flash weights now?
Model weights are planned for a later release. M3.1 Flash is available through the MiniMax API, M Plan and MiniMax Code.