While the frontier labs shipped Claude Fable 5.1 and GPT-6 Astra in the first week of September, a quieter thing happened underneath. Qwen 3.8 Omni Flash takes text, images, audio and video for $0.15 per million input tokens. Xiaomi's MiMo V2.6 Flash does the same four modalities for $0.14. Multimodal stopped being a premium feature and became a commodity, and almost nobody noticed.
What the budget tier looks like now
Here is what you can buy at the cheap end as of mid-September 2026, all verified against live catalogue pricing rather than launch announcements.
- Qwen 3.8 Omni Flash: $0.15 in, $0.47 out. Text, image, audio and video in. 1M context.
- Xiaomi MiMo V2.6 Flash: $0.14 in, $0.28 out. Same four modalities. 1M context.
- DeepSeek V4.1 Flash: $0.10 in, $0.50 out. Text and image. 1M context, and a 943,000-token maximum output.
- GLM 5.3 Flash: $0.15 in, $0.50 out. Text, image and video. A 1.3M context window, the largest here, and the same enormous output ceiling.
- Xiaomi MiMo V2.6 Pro: $0.43 in, $0.87 out. Four modalities, a step up in capability.
For comparison, GPT-6 Astra is $10 in and $50 out. The gap between the cheapest capable multimodal model and the frontier is about seventy times on input.
Audio and video input is the real story
Text and image inputs have been widely available and reasonably cheap for a while. Audio and video are newer at this price point, and they change what is economically sensible to build.
Transcribing and summarising a meeting recording, pulling structure out of a screen capture, indexing a video archive: these were all possible before and all expensive enough that you thought carefully before running them at scale. At $0.14 per million input tokens, the calculation changes.
The caveat is that "accepts video" is not a capability claim. It tells you the endpoint will take the input. It does not tell you the model is good at it, and quality across these models varies far more than the price does. Test on your own material before committing.
Where the cheap tier still falls down
It would be dishonest to present this as the frontier being obsolete. The budget models are markedly weaker on long multi-step reasoning, on hard coding problems, and on tasks where getting it subtly wrong is expensive.
There is also a practical trap. Several of these are reasoning models, which means they spend output tokens thinking before answering. Probe one with a tight output limit and it can return nothing at all, because the reasoning consumed the entire budget. We hit exactly this while testing five of these models and briefly concluded they were broken; given room, all five answered normally. Do not let a badly configured smoke test talk you out of a good model.
How to use the spread properly
The useful pattern is not "pick the cheap one" or "pick the best one". It is routing: cheap models for the high-volume, well-specified work, expensive models for the hard, ambiguous, high-stakes work, and a deliberate decision about which is which.
Most people never do this, because working out where the line falls for their own tasks means running the same work through several models and comparing. That is tedious across separate accounts and separate bills. It is the whole reason Deepest exists: one prompt, several models, answers side by side, so you can see for yourself whether the seventy-times-cheaper option is actually worse at the thing you need.
Our experience is that for a surprising share of everyday work, it is not.
Frequently Asked Questions
Is a $0.14 model really usable for production work?
For well-defined tasks, frequently yes: classification, extraction, summarisation, routine drafting. For open-ended reasoning or anything where a subtle error is costly, test it hard against a frontier model before committing. The honest answer depends on your task, and the only way to find it is to run both.
Why is the output limit so much larger on the cheap models?
We do not have a stated reason from the providers and would rather not guess at one. What we can say is that it is published and consistent across several of them, and that it is genuinely useful for bulk generation where chunking would otherwise cost you coherence.
Does Deepest support the audio and video inputs?
Deepest handles text, images and document attachments today. Audio and video input are supported by several of these models at the API level, and are on our list rather than shipped. We would rather say what is true now than describe a roadmap as a feature.
