The Vibe Father
Model review
MiniMax M3, released June 1, 2026 by MiniMax, is the cheapest model we'd let anywhere near a real codebase: $0.30 per million input tokens, $1.20 per million output. At that price the usual move is to stop reading — sub-dollar models have historically been where good code goes to get refactored into worse code. But M3 posts an 82.1-adjacent LiveCodeBench score of 82.2, carries a full 1M-token context window, and streams at a perfectly usable 95 tokens per second. That combination — actually-frontier-adjacent contest coding at pocket-change prices — is why it's on our leaderboard and why it gets a full review instead of a footnote.
It also arrives with two large blank spaces where benchmarks should be, and those blanks shape everything we're about to recommend.
The numbers
Our Vibe Coding Index weights SWE-bench Verified 40%, Terminal-Bench 30%, LiveCodeBench 30%. M3 has published exactly one of the three:
| Model | SWE-bench Verified | LiveCodeBench | $ in / M | $ out / M |
|---|---|---|---|---|
 Claude Sonnet 5 | 85.2 | 82.4 | 3.00 | 15.00 |
 MiniMax M3 | not yet published | 82.2 | 0.30 | 1.20 |
 Kimi K2.7 Code | not yet published | 82.1 | 0.95 | 4.00 |
 DeepSeek V4 Pro | 77.6 | 87.5 | 0.435 | 0.87 |
The comparison that jumps out: M3's 82.2 LiveCodeBench is 0.2 points behind Claude Sonnet 5 — a model that costs ten times as much on input and twelve and a half times as much on output. On contest-style coding, those two models are functionally tied. The comparison that keeps us honest: SWE-bench Verified and Terminal-Bench are not yet published, which means real-repo issue resolution and agentic shell work — 70% of our index weighting — are unknowns. And the in-tier fight is genuinely awkward for M3: DeepSeek V4 Pro costs slightly more on input, slightly less on output, and has a published 77.6 SWE-bench plus a 5.3-point LiveCodeBench lead. On published evidence, DeepSeek is the stronger buy. M3's counters are speed (95 tok/s vs DeepSeek's sluggish 64) and the lowest input price on the board.
The price math
A heavy month — 50 million input tokens, 10 million output — runs 50 × $0.30 + 10 × $1.20 = $27. Twenty-seven dollars. The same month on Claude Sonnet 5 is $300; on Claude Opus 4.8, $500; even the celebrated-for-cheapness DeepSeek V4 Pro comes to $30.45. M3 is, by a hair, the cheapest heavy month in our entire cheapest-models roundup. At this price, entire categories of use stop requiring justification: run it on every log dump, every "summarize this dependency," every throwaway prototype, every bulk transformation. The meter is effectively off.
Its best seat on the team
M3's seat is budget builder behind a strict reviewer — the same seat as DeepSeek V4 Pro, with a sharper version of the same rule. The published evidence says M3 can genuinely code: 82.2 LiveCodeBench is Sonnet-5-grade problem solving. The unpublished evidence — all of the real-repo and agentic benchmarks — says you should not yet trust it to navigate a large codebase or run long unattended sessions. So the workflow writes itself: M3 drafts in volume, on well-scoped tickets, with the 1M context loaded with everything it needs; a frontier model from our overall rankings reviews before merge; a verification gate runs the real build and real tests before anything is called done. Because M3's tokens are nearly free, the reviewer's cost is your model bill, and the total still embarrasses any single-flagship setup. Its 95 tok/s also makes it a credible cheap scout — quicker than DeepSeek for read-and-summarize work — though the pure speed seats still belong to the models in our fastest-models guide.
There's also a category of work M3 unlocks that more expensive models quietly discourage: speculative volume. Generate three different implementations of the same feature and diff them. Ask for property-based tests on every module, not just the scary ones. Run a full-repo "what looks wrong here" sweep weekly. On a $15-per-million-output flagship, each of those is a line item someone questions; at $1.20, they're rounding errors. The teams getting the most out of the budget tier aren't the ones replacing their flagship — they're the ones adding a layer of cheap, gated, always-on model labor underneath it that simply didn't exist in the budget before.
Verdict
MiniMax M3 is the most interesting $27 in AI coding right now: contest-coding ability that ties a $300-a-month model, at a price where experimentation is free. It has not yet earned trust with real repositories, because nobody has published the numbers that would earn it — when SWE-bench Verified and Terminal-Bench land, this review gets a follow-up. Who should skip it: anyone wanting one model to run unsupervised end-to-end — the missing benchmarks make that a gamble, and DeepSeek V4 Pro is the better-evidenced budget pick for repo work today; teams without a review-and-gate workflow, who will pay the discount back in cleanup; and anyone who needs open weights specifically, where our open-weight guide covers the field. For gated bulk work at the absolute floor of the price sheet, though, M3 actually codes — and that's new for $0.30.