When a team wants to spend less on a coding agent, the question is always the same: what does a cheaper model give up in quality, cost and time? MiniMax’s new M3.1 Flash was the latest model to put that question in front of us, so we ran it on the same long coding sessions, prompts and grading we use for every model, next to Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 and two versions of DeepSeek.
Short answer: MiniMax M3.1 Flash is good at quick coding questions and weaker and much slower on long agentic work. In our benchmark it scored 86% on one-shot coding questions, 4 points under Claude Opus 5.5. On long multi-turn agentic coding sessions it scored 73%, where Opus 5.5 scored 87% and Sonnet 5.5 scored 77%. A long session took 42.8 minutes and made 368 model requests.
These are our own runs on real coding work, not numbers the vendor reported.
MiniMax M3.1 Flash benchmark results
Scores are percentages of the maximum. A long agentic session is a multi-turn coding session on a real service: trace the code, work out why something fails, plan a fix, implement it, then audit it.
| Long agentic sessions | Score | Cost per session | Time per session |
|---|---|---|---|
| ClaudeOpus 5.5, medium (default) | 87% | $5.42 | 17.2 min |
| Open weightsDeepSeek V4.1 Flash 1017 | 77% | $0.14 | 15.4 min |
| ClaudeSonnet 5.5, medium (default) | 77% | $1.28 | 6.8 min |
| ClaudeHaiku 5.5, medium (default) | 76% | $0.69 | 8.1 min |
| NewMiniMax M3.1 Flash | 73% | about $2.60 | 42.8 min |
| Open weightsDeepSeek V4.1 Flash | 73% | $0.16 | 14.0 min |
Short sessions are one-shot asks, like explaining how a setting gets resolved, finding where a flag is used, or fixing a small bug.
| Short sessions | Score | Time per question |
|---|---|---|
| ClaudeOpus 5.5, medium (default) | 90% | 0.9 min |
| Open weightsDeepSeek V4.1 Flash 1017 | 90% | 2.0 min |
| NewMiniMax M3.1 Flash | 86% | 5.8 min |
| ClaudeHaiku 5.5, medium (default) | 79% |
Is MiniMax M3.1 Flash good for coding?
For quick, self-contained questions about a codebase, yes. It scored 86%, well above Haiku 5.5 at 79% and close to Opus 5.5 at 90%. You’ll wait close to six minutes for each answer, though.
For long, multi-step coding work it’s behind. At 73% it sits 14 points under Opus 5.5 and 3 to 4 points under Sonnet 5.5 and Haiku 5.5.
MiniMax M3.1 Flash vs Claude: where it wins and where it loses
The most interesting result is on the diagnosis step, where the model has to work out why something fails. M3.1 Flash scored 76% there, ahead of Sonnet 5.5 at 69%. It reads code carefully and follows failure paths further than most low-cost models we’ve run.
After that it stalls. Planning, implementing and auditing all came in between 66% and 68%.
| Turn in a long session | Score by model |
|---|---|
| Trace the code | 89%92%87%95% |
| Analyze the failure | 76%75%69%90% |
| Plan the fix | 68%81%81%81% |
| Implement | 68%68%73%85% |
| Audit | 66%70%73%82% |
MiniMax M3.1 FlashDeepSeek V4.1 Flash 1017Claude Sonnet 5.5Claude Opus 5.5
Planning is the weak spot, 13 points behind both Sonnet 5.5 and DeepSeek. Implementation is the biggest gap against Opus 5.5, at 17 points.
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash for coding
DeepSeek V4.1 Flash 1017 wins. It scored 77% on long sessions against 73% for MiniMax, and 90% on short questions against 86%. It also finished a long session in about a third of the time (15.4 minutes against 42.8). MiniMax’s one edge is diagnosis, 76% to 75%, which is a tie in practice.
If you’re picking a low-cost model for long agentic coding today, we’d pick DeepSeek V4.1 Flash 1017. We covered it in detail in our DeepSeek vs Claude for coding benchmark.
Why is MiniMax M3.1 Flash slow?
It does a lot of work for each answer. One long session with M3.1 Flash made 368 model requests. Opus 5.5 made 87 for the same work, and DeepSeek V4.1 Flash 1017 made about 240. Each M3.1 Flash session wrote around 290,000 output tokens, three times what Opus writes, and reread about 36 million tokens of conversation along the way.
Some of the 43 minutes may be launch-day serving. But a model that makes four times as many calls as Opus will be slow on any hardware.
How much does MiniMax M3.1 Flash cost?
About $2.60 per long session at MiniMax’s standard pricing. MiniMax hasn’t published a regular price for M3.1 Flash yet, so we priced its token counts at MiniMax M3’s current rates. That’s about twice Sonnet 5.5 at $1.28, and almost 20 times DeepSeek V4.1 Flash 1017 at $0.14. With this many requests per session, call volume will decide the bill more than the per-token rate.
How we tested
The benchmark is a set of agentic coding tasks taken from real work on our own services. Every model gets the same fixed prompts inside Claude Code with the version pinned, and every turn is graded against an answer key that was checked against the code. The judge is Claude Opus 5.5, scoring five dimensions. We ran 8 long sessions and 16 short ones for MiniMax M3.1 Flash, as-is, with Claude Code’s default settings and no extra prompting. We’ve found models score about the same in Claude Code, Codex and Cursor. Session-to-session noise is about 3 points, so treat 1 to 2 point gaps as ties.
FAQ
Is MiniMax M3.1 Flash better than Claude for coding?
Not on long agentic work. It scored 73% against 87% for Claude Opus 5.5 and 77% for Sonnet 5.5. On short one-shot questions it scored 86%, 4 points under Opus.
Is MiniMax M3.1 Flash better than DeepSeek for coding?
No. DeepSeek V4.1 Flash 1017 scored 77% on long sessions and 90% on short ones. MiniMax M3.1 Flash scored 73% and 86% and took about three times as long per long session.
How many tokens does MiniMax M3.1 Flash use?
A lot. One long session made 368 model requests and wrote about 290,000 output tokens, against 87 requests for Claude Opus 5.5.
How long does a MiniMax M3.1 Flash coding session take?
A long agentic session took 42.8 minutes on average, against 17.2 for Opus 5.5 and 6.8 for Sonnet 5.5. A short question took about 5.8 minutes.
What is MiniMax M3.1 Flash best for?
Quick questions about a codebase where you can wait a few minutes for the answer, and finding the cause of a bug.