MiniMax’s Free-to-Paid Funnel Is Finally Visible in the Usage Data

Free models are excellent at generating attention and terrible at proving demand. Remove the price, add prominent placement on a routing platform, and almost any competent model can manufacture an impressive usage spike. MiniMax now has a more interesting result: people appear to be paying after the free trial.

In OpenRouter’s latest weekly rankings, the paid MiniMax M3.1 Flash endpoint entered the top 20 at No. 20 with 2.31 trillion tokens processed and reported growth above 999%. Its free sibling remains far higher, at No. 6 with 8.71 trillion tokens and 75% growth. Together, the two endpoints handled roughly 11.02 trillion tokens during the trailing week. If OpenRouter combined variants, that total would rank fourth, although keeping free and paid endpoints separate is the correct choice because they represent different operational and economic products.

The headline is not that MiniMax won a benchmark. OpenRouter’s ranking is explicitly a usage chart, counting prompt and completion tokens from non-private requests. It does not measure correctness, active developers, revenue, request count or customer retention. The useful signal is narrower: a paid version of a model with a heavily used free tier has accumulated enough billable traffic to sit alongside established OpenAI, Anthropic and DeepSeek models in a crowded leaderboard.

The conversion signal matters more than the debut rank

Free endpoints commonly produce spectacular launch curves. They are attractive for experiments, hobby projects, evaluation harnesses and workloads that can tolerate rate limits or changing availability. Paid traffic is harder to dismiss. Somebody must decide that the model’s output is valuable enough to attach a budget, and production systems tend to expose weaknesses that disappear in a five-minute playground test: tool-call drift, tail latency, malformed JSON, retry storms and inconsistent behavior under long context.

That makes the 2.31 trillion-token paid footprint evidence of conversion, not proof of quality. The free endpoint still carries nearly four times the volume. MiniMax has not shown that M3.1 Flash is the best model in its class; it has shown that its acquisition funnel is moving a meaningful quantity of work from zero-cost sampling into billable execution. For an API model business, that is a more consequential milestone than another benchmark screenshot.

There is also a concentration clue hiding in the numbers. OpenRouter’s wider market-share table gives MiniMax about 2.0% of text requests, while the token leaderboard makes M3.1 Flash look much larger. The likely explanation is not universal developer mindshare but long-running, token-heavy workloads. A small number of agent loops, repository-scale coding sessions or large-context document jobs can consume enormous token volumes without representing broad adoption.

This is the first important caveat for anyone reading rankings as market share: tokens measure throughput, not customers. A model can climb because it is useful for difficult long jobs, because it is verbose, because its tokenizer produces more tokens for the same text, or because agents are taking too many steps. Three of those explanations are commercially interesting; one is simply inefficiency. The chart cannot tell us which is dominant.

A million-token window is a capability, not an architecture

MiniMax’s appeal is easy to understand. The model supports tool use, structured responses, streaming and configurable reasoning through OpenRouter. A hands-on MindStudio evaluation reports a one-million-token context window and adjustable thinking depth. In that test, M3.1 Flash quickly found a subtle inclusive-date error in a six-join DB2 query, returned valid JSON for a multimodal drone task and assembled a playable Godot and Blender game in about 30 minutes. It also misplaced a power line and chose a poor camera angle. That is a familiar frontier-model profile: impressive breadth, followed by the sort of small mistake that becomes expensive when nobody checks the work.

The million-token context claim deserves particular skepticism from system designers. A large window makes it possible to send an entire repository, but possibility is not a recommendation. Full-context prompts increase latency, cost and retrieval noise, while making it harder to understand why the model used one file and ignored another. For many teams, an indexed search step plus a smaller, well-selected context remains cheaper, faster and easier to debug.

The practical test is straightforward: run the same representative tasks with full context and with a boring retrieval pipeline. Measure success rate and total cost per completed task, not whether the model accepted the prompt. If both approaches solve the work at similar rates, keep the simpler and cheaper system. Context-window size belongs in a capacity plan, not on a trophy shelf.

Deployment constraints matter too. OpenRouter currently says EU in-region routing is unavailable for the model. For teams with residency commitments, regulated data or customer contracts that specify processing geography, that is not a footnote. It is a disqualifier until the routing option changes. A model that wins an internal coding bake-off but cannot satisfy the data-processing agreement has lost the only benchmark procurement cares about.

Benchmark the bill, not the token price

Community reaction has been appropriately mixed. A Hacker News discussion attracted 364 points and 156 comments in contemporaneous tracking. Some users connected the model to the free stealth release “Space Bunny” and said it had not impressed them; another comparison preferred DeepSeek V4.1 Flash 1017 for coding. A separate discussion focused on hidden reasoning tokens and argued that cost and latency are more useful than raw output-token totals. These are anecdotes, not a controlled evaluation, but they point toward the right purchasing question.

Engineers considering M3.1 Flash should not migrate because it reached No. 20. They should add it to a task-level bake-off against DeepSeek V4.1 Flash, Claude Haiku 5.5 and whatever model already serves production. Use real prompts and tools from the application, including the ugly cases that cause retries. Record task success, tool-call validity, p50 and p95 latency, retry rate, input and output volume, reasoning usage where observable, and total cost per accepted result.

Cost per million tokens is useful for invoices and weak for architecture. Two models can charge the same token rate while one needs twice as many agent steps, emits longer answers or silently spends more on reasoning. Conversely, an expensive model can be cheaper per completed task if it succeeds on the first attempt. The unit that matters is not the token. It is the resolved support ticket, merged patch, correctly classified document or verified research result.

Teams should also test the paid endpoint directly. Free-tier routing, queueing and rate limits can create a very different operational profile, so a successful free trial does not establish production latency or availability. Pin versions where possible, maintain a fallback, cap agent steps, and log provider and model identifiers with every run. A leaderboard is a discovery mechanism; observability is what turns a discovered model into a service you can trust.

MiniMax has crossed a meaningful line. It did not merely attract bargain hunters with a zero-dollar endpoint; it generated enough paid usage to become visible beside the industry’s largest names. But the honest interpretation stops there. The next durable signal is whether paid M3.1 Flash remains in the top 20 after launch traffic normalizes, and whether MiniMax’s share of requests rises alongside token volume. If tokens keep climbing while request share stalls, this may be a strong specialist for enormous jobs—or an expensive participant in very long loops.

Either result would be useful. What would not be useful is treating token volume as a quality score. MiniMax has earned a place in the evaluation queue, not an automatic merge to production.

Sources: OpenRouter AI Model Rankings, MiniMax M3.1 Flash on OpenRouter, MindStudio hands-on test, Hacker News discussion