Where the free tier really is, and what you give up to use it

The finding first: the free tier is real, but it is not where the best measured value lives.

In the current catalogue, 21 of 396 models are free to call on input price. That sounds generous until you ask the question that matters in production: what measured quality and throughput do you get, and what are you giving up versus the cheapest paid tier? On that question, free looks less like a durable price class and more like a sampling lane. You get room to test, often very large context windows, and sometimes decent speed. What you usually do not get is the strongest measured quality-per-dollar story, or even complete measurement coverage.

The first thing the free tier does offer is breadth of access to large windows. Several free models sit at 1.0M context, including Thinking Machines Inkling Small (free), Thinking Machines Inkling (free), MiniMax M3 (free), Google’s Lyria preview models, and NVIDIA’s free Nemotron variants. Others come in at 512K, 262K, 256K, and 200K. If your immediate problem is, “Can I fit this document set into one request?” free can already answer yes more often than it could a year ago.

That matters. Context is one of the few things a free model can offer that is immediately useful even before you trust its answers. Teams use free endpoints to test prompt structure, chunking strategy, retrieval payload size, and failure modes under long inputs. For that, a million-token free model is not a gimmick. It is a subsidy.

But production selection is not won on context alone, because a long prompt that produces weak output is still a bad deal.

The cheapest measured paid tier is now low enough to make “just use free” a lazy answer. Ling-3.0-flash sits at $0.021 per million input tokens with measured quality 37.8 and output throughput 371 tokens per second. Upstage Solar Pro 4 comes in at $0.03/M with quality 41.6. OpenAI’s gpt-oss-120b is $0.037/M with quality 24.1 and 179 tokens per second. NVIDIA Nemotron 3.5 Lightning is $0.08/M with quality 23.6 and 283 tokens per second. DeepSeek V4 Flash 0423 is $0.081/M with quality 51.8 and 135 tokens per second. Those are not premium prices. They are floor prices.

And that is the problem for the free tier. Once the paid floor is measured in hundredths of a dollar per million input tokens, free no longer competes mainly against expensive frontier models. It competes against very cheap, benchmarked, production-shaped alternatives.

The chart is one-sided on purpose. The catalogue can tell us there are 21 free models, and it can show the measured floor of paid models very clearly. What it does not show with the same completeness is a rich measured set for the entire free cohort. That absence is itself part of the answer.

What do you give up to use free? First, you give up certainty. The free set includes many preview, router, and promotional entries. Some are likely there to seed adoption, demonstrate capability, or fill traffic opportunistically rather than to anchor a long-term production contract. If you are standardising a service, “free today” is not the same thing as “stable commercial tier tomorrow.”

Second, you give up comparability. The cheapest paid models are easy to compare because they are measured on quality and, often, on throughput. The free cohort is less consistently benchmarked in the way an operator needs. In practice that means more of your own evaluation burden survives. You can absolutely save money on calls, but you spend more of that savings in uncertainty: prompt testing, fallback design, rate-limit handling, and deciding whether a free preview is good enough to trust on real traffic.

Third, you often give up the cleanest speed-quality bargain. A paid model at two or three cents per million input tokens is close enough to free that throughput and consistency start to dominate the decision. Ling-3.0-flash at 371 tokens per second is not merely cheap; it is cheap enough that latency-sensitive workloads should take it seriously on speed alone. Nemotron 3.5 Lightning at 283 tokens per second makes a similar case. Free is hardest to defend once the paid alternative is both cheap and fast.

So where is the free tier really? It is in three places.

It is in experimentation: trying prompts, tools, formats, and long-context payloads without creating a bill worth discussing.

It is in overflow capacity: a place to route non-critical workloads, internal demos, and evaluation traffic when you do not need hard guarantees.

And it is in market-making: providers using zero-priced access to get distribution for new or less-established models.

What it is not, at least from the catalogue as measured today, is the obvious home of best value. The best value frontier is already crowded by paid models that are cheap enough to make the word “paid” almost irrelevant and benchmarked enough to make it operationally meaningful.

If you are choosing for production, the practical rule is simple. Start with free when the question is exploratory: can this model family handle the context length, format, or task shape at all? Move to the cheapest measured paid tier as soon as the question becomes operational: what can I trust for known quality, known speed, and repeatable cost? The surprise in the market now is not that some models are free. It is how little you have to pay before free stops being the bargain that matters.

Both projects are on GitHub and PyPI. Install them.