InclusionAI's Ling 3.0 Flash costs $0.02 per million input tokens. Qwen 3.7 Flash costs $0.03. At those rates, processing the entire text of every book you have ever read costs less than a coffee. The bottom of the market has fallen far enough that cost has stopped being the constraint on most high-volume work, and almost nobody has updated their instincts.
What a cent actually buys now
Prices per million tokens, verified against live listings in mid-August 2026:
- InclusionAI Ling 3.0 Flash: $0.02 in, $0.06 out. 262K context.
- Qwen 3.7 Flash: $0.03 in, $0.13 out. 1M context, with image and video input.
- Nvidia Nemotron 3.5 Lightning: $0.08 in, $0.20 out. 262K context.
- Poolside Laguna S 2.1: $0.09 in, $0.18 out. 1M context.
- Upstage Solar Pro 4: $0.09 in, $0.36 out. 524K context.
For scale, GPT-5.6 Sol launched at $5 per million input tokens. Ling 3.0 Flash is 250 times cheaper on input. That is not a discount, it is a different category of thing.
What these models are genuinely good at
Small, cheap models have improved considerably and are still not frontier models. The honest framing is that they are excellent at bounded tasks and unreliable at open-ended ones.
Work they handle well: classification and routing, extraction into a fixed schema, summarising individual documents, straightforward translation, normalising messy text, first-pass filtering before something more expensive looks at what survives.
Work they handle badly: multi-step reasoning, anything requiring the model to notice it lacks information, nuanced judgement, long-horizon coherence, and tasks where being subtly wrong is expensive and hard to detect.
The failure mode matters more than the failure rate. A cheap model tends to fail confidently and plausibly, which is the worst combination if nothing downstream checks it.
The pattern that works: cascade
The most effective use of this price spread is not choosing one model. It is running a cheap model over everything and escalating the hard cases.
Filter a hundred thousand documents with a model at $0.03 per million tokens, have it flag the ambiguous ones, and send only those to a frontier model. If 5% escalate, you have bought frontier-quality outcomes on the cases that needed it at roughly a twentieth of the cost of running everything through the expensive model.
Two things make or break this. The cheap model needs to be reliable at recognising its own uncertainty, which some are and many are not. And you need to measure the escalation rate rather than assume it, because a cascade that escalates 60% of cases is just an expensive pipeline with extra steps.
Why the instinct lags the prices
Most people's sense of what AI costs was formed when a long prompt to a good model cost real money and you thought before sending it. That instinct is now wrong by two orders of magnitude at the bottom of the market, and it quietly stops useful work from being attempted.
The fix is to actually price your workload rather than estimating it from memory. Multiply your token volume by the current rate for a few candidate models, including the cheap ones you dismissed. The answer is frequently small enough to change what you decide to build.
Then test whether the cheap model is good enough for the specific job, which is the only question that matters and the one benchmarks answer least well. Running the same work through a cheap model and an expensive one side by side takes minutes and settles it. It is the core of what we built Deepest to do, and it regularly shows that the expensive default was not buying anything.
Frequently Asked Questions
Are these models safe to use in production?
For bounded tasks with validation downstream, yes, and plenty of production systems run on them. The requirement is that something checks the output: a schema, a test, a threshold, or a more capable model reviewing flagged cases. Cheap models without verification in front of users is where this goes wrong.
Why are they so much cheaper?
Mostly size. These are small models, and inference cost scales with parameters. Some are also priced competitively to win adoption. We would rather not speculate about individual providers' margins, which are not public.
How do I tell if a cheap model is good enough for my task?
Take twenty real examples, run them through the cheap model and your current one, and compare the outputs directly. Twenty is usually enough to see the pattern. If you cannot tell which is which, you have been overpaying; if the gap is obvious, you now know exactly what the extra money buys.
