Short answer
MiniMax M3 (MiniMax-M3) is the model MiniMax released on June 1, 2026, built for coding and agent work, with a context window of up to 1 million tokens, native image and video input, and open weights. MiniMax's standard price is $0.30 per million input tokens and $1.20 per million output tokens (listed as a "permanent 50% off" the $0.60 / $2.40 base price), with cache reads at $0.06. On TeamoRouter it is available under the same model ID at half of MiniMax's price at the time of writing: $0.15 input / $0.60 output, cache reads $0.03, on one key and one balance shared with Claude, GPT and Gemini.
Specs at a glance
From MiniMax's website and platform docs:
| MiniMax M3 | |
|---|---|
| Model ID | MiniMax-M3 |
| Released | June 1, 2026 |
| Context | Up to 1M tokens (512K guaranteed minimum) |
| Attention | MSA (MiniMax Sparse Attention); MiniMax says per-token compute at 1M context is 1/20 of the previous generation |
| Input | Text, images and video; MiniMax says it can operate a desktop computer |
| Thinking | Can be switched on or off; both modes cost the same per token |
| Open weights | Published on Hugging Face on June 2 (MiniMaxAI/MiniMax-M3, about 427B parameters) |
Pricing
Per 1 million tokens, standard tier:
| MiniMax | TeamoRouter | |
|---|---|---|
| Input (≤512K tokens) | $0.30 | $0.15 |
| Output (≤512K tokens) | $1.20 | $0.60 |
| Cache read | $0.06 | $0.03 |
| Requests with more than 512K input tokens | $0.60 input, $2.40 output | $0.30 input, $1.20 output |
MiniMax also has a priority tier (service_tier=priority, scheduled first under high load) at $0.45 / $1.80 up to 512K. TeamoRouter prices are at the time of writing; the current numbers are on the pricing page.
MiniMax also sells a monthly Token Plan; at launch it was Plus $20/month for about 1.7B tokens, Max $50 for about 5.1B and Ultra $120 for about 9.8B. If MiniMax is the only model you use and you use a lot of it, the official plan is far cheaper than pay-as-you-go, so buy that instead.
When TeamoRouter makes sense: you pay as you go below Token Plan volumes, where TeamoRouter's per-token price is half of MiniMax's; you already use Claude, GPT or DeepSeek and want MiniMax on the same key and bill; you want to switch models in Claude Code, Cline or Cherry Studio without opening another account; or you want to pay by card or crypto.
MiniMax's published benchmarks
These are MiniMax's own numbers from launch. The comparison models at the time were Claude Opus 4.7 and GPT-5.5; both labs have shipped a newer generation since.
| Benchmark | MiniMax M3 |
|---|---|
| SWE-Bench Pro | 59.0% |
| Terminal-Bench 2.1 | 66.0% |
| MCP Atlas | 74.2% |
| BrowseComp | 83.5 (MiniMax lists Opus 4.7 at 79.3) |
MiniMax also showed two long-running tasks: reproducing an ICLR 2025 Outstanding Paper on its own over nearly 12 hours, and optimizing an FP8 matrix-multiply kernel on Hopper GPUs over about 24 hours (147 benchmark submissions, 1,959 tool calls, hardware utilization from 7.6% to 71.3%). These are vendor tests, so compare it on your own tasks.
Call it through TeamoRouter
from openai import OpenAI
client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
model="MiniMax-M3",
messages=[{"role": "user", "content": "Find the race condition in this code: ..."}],
)
print(resp.choices[0].message.content)
Full reference: API integration.
Paying for usage
No subscription: you top up a balance from $5 and pay only for the tokens you use, by card or with USDT, USDC or ETH from any exchange app (how to pay with crypto).
FAQ
How does M3 relate to M3.1 Flash? M3.1 Flash is a smaller model MiniMax released on October 19, priced at $0.15 / $0.60, cheaper and faster; M3 is the flagship. TeamoRouter offers both: M3.1 Flash is minimax-m3.1-flash at $0.075 / $0.30 at the time of writing, also half of MiniMax's price.
Does thinking mode cost more per token? No, both modes have the same price; thinking uses more output tokens.
Do I need a MiniMax account? Not with TeamoRouter. One key covers MiniMax, Claude, GPT, Gemini and DeepSeek.