Somewhere in the pricing tables making the rounds this year sits a number that doesn’t belong: MiniMax M2.7 at $0.0002 per million tokens, for both input and output. Read that again. Two hundredths of a cent, for a million tokens. Deepseek V4 Flash, itself marketed as an aggressive price disruptor, charges 700 times more for input tokens alone. If the number is real, it’s not a competitive price. It’s a different category of thing entirely, closer to free than to cheap.
Numbers that look like typos are usually typos. So before anyone reroutes a production workload chasing it, the number deserves an actual look: where it comes from, whether the model behind it can back it up, and what a price that low would actually mean if it held.
What MiniMax M2.7 can do, independent of price
Set the pricing aside for a moment, because the benchmark story isn’t in dispute the way the price is. MiniMax M2.7 posts a coding score in the 82nd percentile, with 59 percent on SWE-Bench Pro and 66 percent on Terminal-Bench, plus a 74.2 percent result on the MCP Atlas agent benchmark and an intelligence score in the 93rd percentile. Those are legitimate, competitive numbers against models that cost meaningfully more to run. Nobody serious is disputing that MiniMax built a strong model. The question is entirely about the price attached to using it.
That distinction matters, because it rules out the laziest explanation. This isn’t a case of a mediocre model padded out with a suspiciously cheap price tag to get attention. The model earns its benchmark scores on its own merits. The pricing is the anomaly, not the product.
Three explanations, and which one is most likely
There are basically three ways a price this far outside the norm ends up published.
It’s a decimal error somewhere in the pipeline. Pricing aggregators scrape and reformat numbers from primary sources, and a misplaced decimal point turns $0.02 into $0.0002 without anyone noticing until someone tries to reconcile an invoice. This is the most mundane explanation and, given how far outside the pattern the number sits (not double or half of a plausible price, but roughly two orders of magnitude off) it’s a genuinely plausible one. A price that’s 700 times cheaper than the next-cheapest competitor is more consistent with a transcription error than with a real business decision.
It’s a promotional or introductory rate being reported as the standing price. Companies routinely launch a new model with an aggressive limited-time rate to drive adoption, then revert to something closer to market rate once usage ramps up. If that’s what happened here, the number was accurate at the moment it was captured and stale by the time anyone reads this. This publication has already flagged one real instance of exactly this pattern elsewhere: Claude Sonnet 5’s introductory pricing runs through August 31, 2026, after which the standard rate is 50 percent higher. Promotional pricing that later gets treated as the permanent number is a known failure mode in this industry, not a hypothetical one.
It reflects a different billing model than a simple per-token price. Some platforms report an API rate that’s technically accurate but structurally misleading, for example a rate that only applies inside a bundled subscription tier, or one that’s the marginal cost after a separate flat fee, or a rate quoted before volume-based markups apply at production scale. MiniMax also offers free and paid agent-platform plans that sidestep the API entirely, and if a subscription-tier number got mislabeled as the raw API rate somewhere in the reporting chain, the discrepancy would look exactly like what we’re seeing.
None of these three explanations require MiniMax to be acting in bad faith. All three are consistent with a genuinely capable model attached to a pricing figure that shouldn’t be trusted without a second source.
What would have to be true for the number to be real
Play it out honestly: suppose the $0.0002 rate is real, sustained, and not a promotional artifact. What would that actually mean?
It would mean MiniMax is pricing meaningfully below its own infrastructure cost, at least in the near term, which is a viable strategy for buying market share fast but not one that survives indefinitely without either a much larger revenue stream elsewhere or a price correction once the land grab is over. Aggressive loss-leader pricing to establish a foothold against entrenched competitors is a real strategy real companies run. It’s also, by its nature, temporary. Nobody subsidizes inference at a 99.9 percent discount forever.
It would also mean the economics of running an agent swarm change dramatically for anyone who builds on it early, right up until the pricing normalizes and whatever was built around the assumption of near-zero marginal cost needs to be re-architected around a real number. That’s the actual risk of over-trusting an anomalous rate: not that the model turns out to be bad, but that a system gets designed around a cost structure that quietly stops existing.
The honest verdict
Treat the $0.0002 figure as unverified until it’s confirmed directly against MiniMax’s own current, primary pricing documentation, not a third-party aggregator. That’s not a dismissal of the model. The benchmark numbers are real and worth taking seriously on their own terms. It’s a statement about how far outside the normal range this specific price sits, and how many mundane, non-conspiratorial explanations exist for exactly that pattern.
If you’re evaluating MiniMax M2.7 for a real workload, the responsible move is straightforward: pull the current rate from MiniMax’s own billing documentation before you build a cost model around a number you found in a comparison table. If the rate holds up, you’ve found a genuinely remarkable deal on a genuinely strong model. If it doesn’t, you’ve saved yourself from architecting a production system around a number that was never going to last. Either way, the model’s actual capability isn’t the part in question here. The price is.
Sources: MiniMax M2.7 API pricing (pricepertoken.com) and MiniMax M3 benchmarks and pricing (minimax-ai.chat); Deepseek V4 API pricing for comparison (benchlm.ai); Anthropic Claude API pricing documentation for the Sonnet 5 introductory-pricing precedent. Pricing figures reported by third-party aggregators as of August 2026; this article recommends independent verification against MiniMax’s primary documentation before any production use, and does not assert the $0.0002 rate is confirmed accurate.
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com
