Prices as of 2026-07-14 · rev 179affb· unknown is shown as unknown, never $0.00
Capability fingerprint
independent benchmarks, grouped by domain · higher is better · not our measurementReasoning
peer median · percentile ranks this model among distinct comparable models (deduped — one entry per model, not per provider) scored on that benchmark · exact model / family proxy per row.
Source: Epoch AI — “AI Benchmarking Hub” (epoch.ai), CC-BY (Epoch AI). Some rows carry an upstream leaderboard (shown per row); Aider & Terminal-Bench are additionally Apache-2.0. Each benchmark measures a different thing on a different scale — rows are not comparable across domains, and a score is not a guarantee of quality on your task.
Capabilities Index
Epoch AI's cross-model capability re-fit · one comparable axis · not our measurementThe dot is Minimax M3; the whisker is its 95% CI; the firm tick is the peer median; faint ticks are every other scored model. Unlike a raw benchmark, the ECI axis is comparable across models — but it's a relative re-fit scale, so a model's exact position can shift when Epoch recomputes it.
Source: Epoch AI — Capabilities Index & Notable-Models data (epoch.ai), CC-BY (Epoch AI). The ECI is Epoch's index, not TokenTriage's; per-fact confidence labels are Epoch's own.
Human preference (LMArena)
blind pairwise human votes · Elo · style-controlled · not a correctness scoreminimax-m3statistically tied with 16 models overallEach dot is Minimax M3's Arena score in that category; the whisker is the 95% CI; the firm tick is the peer median across distinct models; faint ticks are the field. This is human preference — which answer people pick in a blind A/B — not a correctness or capability score, and adjacent ranks routinely overlap within their intervals.
Source: LMArena (LMArena Leaderboard (lmarena.ai) — lmarena-ai/leaderboard-dataset on Hugging Face, CC BY 4.0.). Ratings are the style-controlled estimates (length/formatting confound removed), as of 2026-07-27.
Also worth considering
cheaper models that score about as well — pick the benchmark that matters to you1 cheaper model scores within 3 points of Minimax M3 on SimpleBench.
| Model | SimpleBench | Cost / req | Savings |
|---|---|---|---|
| Deepseek V4 Pro | 53%+18 | $0.00087 | −3% |
Deduped to one entry per model (cheapest offering), priced at the reference 1K-in / 500-out request. Filtered only on the benchmark score — a model with an unknown capability is never silently excluded. A lower price is only cheaper if quality holds on your task — that's what routing proves.
Pricing
| Input · fresh prompt tokens | $0.30 / M |
|---|---|
| Output · generated tokens | $1.20 / M |
| Cache read · cached-input hit | $0.06 / M |
| Cache write · cache creation | unknown |
| Reasoning · thinking tokens | unknown |
Context-length tiers
| Base (≤ 512K) | $0.30 / $1.20 per M |
|---|---|
| Above 512K | $0.60 / $2.40 per M |
When total input exceeds the threshold, the tier rate applies to the whole request.
Cross the 512K line by a single token and the whole request reprices — every token, not just those above the line. Assumes 500 output tokens (the cliff's position depends on input alone). Snapshot rates; providers do change tier structure.
Reasoning tokens: no separate rate is published. Providers typically bill thinking tokens at the output rate ($1.20 / M) — which is what the estimator assumes, so an extended-thinking workload isn't silently undercounted.
Model info
- Provider
- Minimax
- Catalog key
- minimax/minimax/MiniMax-M3
- Identifier
- Floating alias (may re-point over time)
- Type
- Chat
- Max input
- 1,000,000 tokens
- Max output
- 128,000 tokens
- Released
- 2026-06-12
- Knowledge cutoff
- unknown
- Weights
- Open weights
- Price snapshot
- 2026-07-14 · rev 179affb
Pricing from TokenTriage's resolved price artifact (TokenTriage resolved price artifact (internal/pricing/data/prices.json)); capability + model metadata is third-party (LiteLLM model_prices_and_context_window.json, models.dev). Unknown means unstated, not absent.
Capabilities
4 documented as supportedInput & output
API behaviour
Agents & tools
✓ supported ·✕ not supported ·? not documented ·⚠ sources disagree. A models.dev tag means the LiteLLM catalog was silent and an independent source supplied the value (lower confidence); ? is never silently turned into a confident ✕.
Open weights
verified Hugging Face repo · facts from the model's safetensors index- Parameters
- 427.0B
- Architecture
- MiniMaxM3SparseForConditionalGeneration
- Native precision
- BF16
Model
- Native size
- 854 GB
Weights footprint
Memory to hold the weights — not VRAM. Real serving needs extra headroom for activations and the KV cache, which grows with context length and batch size.
- License
otherCommercial use — conditional- Repository
MiniMaxAI/MiniMax-M3
License & source
Source: Hugging Face Hub — facts read from the model repo as of 2026-07-30. Each fact carries the model's own repo license (shown above); TokenTriage neither hosts nor relicenses the weights.
Estimate a request
disjoint token buckets · input = fresh (uncached) tokensHow output length drives the bill
at a 1K-token prompt · cost scales linearly with outputOutput tokens overtake input spend at ~250 tokens for this model — past that, generation is your bill, whatever the headline input price says. Computed by the same cost engine as the estimate above.
Route to Minimax M3
OpenAI-compatible · one base URLcurl https://YOUR-TOKENTRIAGE-HOST/v1/chat/completions \ -H "Authorization: Bearer $TOKENTRIAGE_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "minimax/minimax/MiniMax-M3", "messages": [{"role": "user", "content": "Hello"}] }'from openai import OpenAI client = OpenAI( base_url="https://YOUR-TOKENTRIAGE-HOST/v1", api_key="YOUR_TOKENTRIAGE_KEY", ) resp = client.chat.completions.create( model="minimax/minimax/MiniMax-M3", messages=[{"role": "user", "content": "Hello"}], ) print(resp.choices[0].message.content)
Point your existing OpenAI client at your TokenTriage host and pass the catalog key as the model. TokenTriage attributes the real cost per request and can route to a cheaper model once it's proven just as good on your traffic.
Data & API
machine-readable · rev 179affbPull this model's pricing, capabilities, benchmarks and cross-source provenance programmatically — every field carries the snapshot rev, and unknown is null, never 0. Free to reuse with attribution: benchmarks & Capabilities Index © Epoch AI (CC-BY (Epoch AI)); human-preference Elo © LMArena (CC-BY-4.0); capability & price cross-check via models.dev v2 (MIT) · Portkey (MIT) · TrueFoundry (MIT); open-weights facts via Hugging Face (per-repo license).
A price is a guess until it meets your traffic.
This page tells you what Minimax M3 charges per token. TokenTriage tells you what it costs on your real requests — and cuts the bill only after proving a cheaper model is just as good.