Flash Models Compared: Gemini 3.8, DeepSeek V4.1, Qwen3.8-Omni and MiMo V2.6

Four cheap models in twenty days

Between 2 and 22 September 2026, four labs shipped a model into the same slot: fast, cheap, high volume, aimed at the work that sits underneath an application rather than the work that impresses on a benchmark chart.

They are not equivalent products, and the published prices differ by roughly eight times on a realistic workload. Two of them carry billing conditions that a single per-token figure cannot express at all.

How we compared

Every price below was read from the provider's own pricing page or model page on 29 September 2026 and linked once from each model's heading. We have not run these models against a benchmark harness, and we make no claims about relative quality. Where a specification is reported by a third party rather than the lab, we say so.

The published rates

Sources: MiMo V2.6 Flash, Qwen3.8-Omni-Flash on Alibaba Cloud Model Studio, DeepSeek pricing, Gemini API pricing.

What a month actually costs

Take a workload of 10 million input tokens and 2 million output tokens a month, which is a modest production feature rather than a toy:

Those are arithmetic from the table above, not vendor estimates. The headline is not that Gemini is expensive in absolute terms; at fifteen dollars a month nobody is filing a variance report. It is that the ratio holds as you scale, and at a hundred times this volume the same choice is fifteen hundred dollars against two hundred.

The two billing traps

Gemini's price is introductory and Google says so plainly. The pricing page carries the rate "through December 31, 2026" and the replacement rate "starting January 1, 2027" in the same cell. Comparison articles quoting $0.75 without that condition are describing a price with a deadline attached.

DeepSeek charges by the clock. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays, and everything doubles inside them. For a US-facing application most traffic falls outside those windows, so the effective rate is near the off-peak figure. For a European product, the 06:00 to 10:00 UTC window is the morning rush. The same model can be the cheapest or the second most expensive here depending on where your users wake up.

What you get beyond the price

Cheap is a property; capability is a different one.

• Qwen3.8-Omni-Flash accepts text, images, audio and video, and returns text, with a 1 million token context window, according to TechNode's report of Alibaba's announcement. If your pipeline currently transcribes audio before sending it to a model, this collapses two steps into one. Alibaba also lists a realtime variant with separate audio rates of $0.93 input and $1.87 output per million tokens.
• MiMo V2.6 Flash publishes a 1 million token context window on its model page. Third-party listings describe it as a mixture-of-experts model with 309 billion total parameters and 15 billion active per token, released under an MIT licence; we could not confirm those figures on Xiaomi's own page, so treat them as reported rather than verified.
• DeepSeek V4.1 Flash is the volume workhorse of the group, and its cache-hit price of $0.003 per million tokens off-peak is the lowest number in this article by an order of magnitude.
• Gemini 3.8 Flash is the only fast-tier release here from a US lab, and the only one that arrives inside an enterprise cloud with the contractual apparatus that implies.

Who is missing, and why that matters

Neither OpenAI nor Anthropic shipped a fast-tier model in September 2026. OpenAI's cheapest current option is GPT-5.6 Luna, released in July 2026, and Anthropic's is Claude Haiku 4.5, from October 2025.

That is worth stating because a reader scanning release trackers could reasonably conclude the entire cheap tier refreshed this month. It did not. Three quarters of this cohort came from Chinese labs, and the competitive pressure on price is coming from there rather than from San Francisco. Our frontier model comparison covers the other end of the same market.

Is the cheapest model the right default?

Not automatically, and the reason is not quality.

Three of these four models come from Chinese labs. For a hobby project that is irrelevant. For a company with data residency obligations, a customer contract naming processing locations, or a security review to pass, it is the first question rather than the last one. Alibaba prices this model from its Singapore deployment; DeepSeek and Xiaomi publish from China. None of that makes a model unsafe to use, and we are not suggesting it does. It makes the choice a procurement decision as much as an engineering one, and the person who signs it should know which jurisdiction they are signing into.

The second consideration is operational. A price that doubles on a calendar date and a price that doubles at 06:00 UTC are both fine if you know about them, and both nasty if you discover them in an invoice. If you route across several providers, a gateway makes switching a configuration change rather than a migration, and our guide to AI evals covers how to check that a cheaper model still does the job before you move traffic onto it.

Which should you choose?

If cost per token is the deciding factor: MiMo V2.6 Flash, at $0.14 and $0.28. Confirm it can do your task at acceptable quality first, because there is no independent benchmark in this article and the cheapest model is only cheap if it does not need three attempts.

If your input includes audio or video: Qwen3.8-Omni-Flash. It is the only model here that takes those natively, and folding a transcription step into the model call usually saves more than the token price difference.

If your traffic is concentrated outside 01:00 to 10:00 UTC on weekdays: DeepSeek V4.1 Flash, where you will mostly pay off-peak rates, with cache hits at $0.003 per million tokens if your prompts share a stable prefix.

If you need a US provider, enterprise contracts or Google Cloud adjacency: Gemini 3.8 Flash, budgeted at the January rates of $1.50 and $7.50 rather than today's introductory figures.

If you are choosing on benchmark charts: be careful. Every model here was released this month, most published benchmark comparisons are vendor-run, and fast-tier models are unusually sensitive to prompt format and task shape. Measure on your own workload. Our comparison of inference providers covers the hosting side of the same question, and the open-source model guide covers running weights yourself.

Conclusion

The fast tier is now a price war, and the interesting part is not who is cheapest this week. It is that two of the four prices in this article are conditional: one on a calendar date, one on the hour of the day. A single number in a comparison table cannot carry either condition, which is why most tables covering this cohort are already wrong.

Work out your monthly token volume, apply the rates above to it, and then apply them again as they will be on 1 January. That second number is the one to budget from.

This comparison is an editorial synthesis of vendor pricing pages and model documentation read on 29 September 2026 and linked inline. We did not run benchmarks, measure latency, or receive briefings from these labs. Specifications reported only by third parties are labelled as such. Prices in this category move quickly and one of them changes on a known date; verify before committing spend.