MiniMax M3 has a 1M-token context window, but its official price has a step at 512K input tokens. If you run long agent sessions, that step matters more than the headline rate. Here is the math, using prices checked on 2026-10-29.
The official rates
MiniMax's pricing page (platform.minimax.io/docs/pricing/overview), standard tier, USD per 1M tokens, with the current "permanent 50% off" applied:
| Request size | Input | Cached input | Output |
|---|---|---|---|
| Up to 512K input tokens | $0.30 | $0.06 | $1.20 |
| Above 512K input tokens | $0.60 | $0.12 | $2.40 |
The higher row applies to the whole request once input crosses 512K, and the priority service tier costs 1.5x either row.
One request on each side of the line
An agent turn with 90% of the prompt cached and 8K output tokens:
- 510K input: 51K x $0.30 + 459K x $0.06 + 8K x $1.20 = about $0.052
- 520K input: 52K x $0.60 + 468K x $0.12 + 8K x $2.40 = about $0.107
About 2% more context, about twice the cost.
A month of traffic
100M input tokens and 10M output tokens, at different cache-hit rates:
| Cache hit | M3, all requests under 512K | M3, all requests over 512K | M3.1 Flash |
|---|---|---|---|
| 0% | $42.00 | $84.00 | $21.00 |
| 50% | $30.00 | $60.00 | $14.25 |
| 80% | $22.80 | $45.60 | $10.20 |
| 90% | $20.40 | $40.80 | $8.85 |
MiniMax M3.1 Flash, released on 2026-10-19, is listed at $0.15 / $0.015 / $0.60 with no long-context step on the same page. It is a different model, though: its thinking can't be turned off and reasoning_effort defaults to max, so reasoning tokens (billed as output) can eat much of the saving unless you set low or medium.
What I'd do
- Log input tokens per request and alert well below 512K, not at the 1M limit.
- Compact or summarize history before a session crosses the step.
- Re-check cached-input pricing on every provider you compare. Third-party hosts of M3 vary a lot, and many serve fp8 or fp4 weights while MiniMax's own weights are BF16, so compare like with like.
Disclosure: I work on IteraCompute, an OpenAI-compatible API for open models. We list MiniMax M3 at $0.30 input, $0.05 cached input and $1.19 output per 1M tokens (checked 2026-10-29). The model page has the current price: MiniMax M3 Price
Top comments (0)