DeepSeek V4.1 Flash: Frontier AI for Pennies 2026

A near-frontier, million-token, MIT-licensed model just landed at $0.15 per million tokens, and it resets what "cheap enough to build on" means.

On September 10, 2026, DeepSeek shipped V4.1 Flash: a 552B-parameter mixture-of-experts model that activates only 8B to 16B parameters per token, reads a 1 million token context window, understands images natively, and ships its weights under the MIT license - DeepSeek model card. The list price is $0.15 per 1M input and $0.60 per 1M output tokens during off-peak hours - DeepSeek pricing.

For anyone building a product on top of a language model, that second number is the whole story. Gemini 3.8 Flash, Google's newest budget tier, lists at $0.75 / $3.75 and is scheduled to double to $1.50 / $7.50 on January 1, 2027 - Google pricing. A frontier flagship like Claude Opus 5 runs $5 / $25 - Anthropic pricing. DeepSeek just undercut the cheap tier by roughly 5x and a premium flagship by more than 30x, while handing you the weights to run the thing yourself.

1. What DeepSeek actually shipped

The architecture is the reason the price is even possible. V4.1 Flash is a sparse MoE on a new Causal Encoder-Decoder design (a 40-layer stack of 20 encoder plus 20 decoder layers) that carries 384 routed experts but fires only 6 of them per token - DeepSeek model card. Because so little of the network activates on any given token, the serving cost per token collapses. DeepSeek also shrank the KV cache to about 890 bytes per token, roughly a quarter of the previous Flash generation, which is what makes a million-token window affordable to actually use rather than just advertise.

Capability is genuinely strong for the tier, though not frontier. Independent testing by Artificial Analysis places V4.1 Flash at 40 on its Intelligence Index, sixth among open-weight models and well above the open-weight median of 18, while the closed frontier leaders still sit clearly ahead - Artificial Analysis. DeepSeek's own scorecard is higher still, reporting 90.9 on GPQA Diamond and a 3471 Codeforces rating - DeepSeek changelog, though vendor benchmarks always flatter the vendor. The honest read: this is not the smartest model you can buy, but it is comfortably the smartest model at this price, and for the bulk of production work that is the trade that matters.

The pricing has one wrinkle worth knowing before you budget: DeepSeek runs peak and off-peak rates, with peak hours from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and cached input as low as $0.003 per 1M tokens - DeepSeek pricing. Even at the peak $0.30 / $1.20, it is still the cheapest capable model on this list. The old default Flash endpoint now serves V4.1 automatically, so callers on deepseek-flash inherit the new model and the new floor without changing a line of code.

2. Why the price floor just moved (and what it does not do)

Strip this down to first principles and the pattern is old. Intelligence is becoming a commodity input, and commodity inputs trend toward their marginal cost of production. What is new is that DeepSeek attached an MIT license to a near-frontier model, which removes the last soft floor under the price: if the API ever gets expensive, you can run the open weights on your own hardware, so the vendor's margin is no longer something you are forced to pay. A closed model can hold a premium only as long as nothing open is good enough. Something open is now good enough.

The competitive pressure is already visible, including on DeepSeek itself. The company floated folding its pricier V4 Pro tier into Flash economics, then reversed course and kept V4 Pro running at its existing prices past September 14 - Yotta Labs. Google, meanwhile, is moving the other way: Gemini 3.8 Flash doubles in January, and it "works harder" by design, spending more thinking tokens, which are billed at the output rate. In one independent run it burned 120M output tokens to finish a benchmark suite against a 71M median for comparable models - eesel AI.

That last point is the counter-narrative, and it matters: sticker price is not total cost. A cheaper per-token rate that triples your token count is not cheaper. And a 40 on the intelligence index is not a 53: for genuinely hard reasoning, a frontier flagship that reaches the answer in fewer, better steps can still be the cheaper path to a correct result. Cheap intelligence does not delete the frontier. It sorts work into tiers.

3. What a founder actually does with $0.15 tokens

If intelligence is the cheap input, the value moves to whoever turns it into outcomes: a shipped feature, a booked customer, a company that runs itself. The practical move is not "switch everything to the cheapest model." It is model routing: put the bulk of the work (drafting, extraction, classification, chat, first-pass code) on V4.1 Flash, and reserve a frontier model for the hardest slice where a wrong answer is expensive. We break the mechanics down in our guide to cutting agent costs with model routing, and the companion piece on pricing your product to beat token costs.

For a non-technical founder, the shift shows up as a lower cost to operate, not a model picker to babysit. A platform like Founden builds and runs a whole company from one conversation (website, app, billing, admin), and every one of those steps is a stack of model calls; a 5x cheaper bulk tier means the same autonomous company costs a fraction of what it did a quarter ago. The open weights also change the risk math: an MIT model cannot be deprecated out from under you or repriced the way a closed endpoint can, which is exactly the kind of dependency worth planning around when you choose the model you build on and design the AI-native company stack.

The one habit to build now is measuring cost per task, not cost per token, and turning the effort dial down wherever a cheaper model clears the bar. DeepSeek did not just release a model. It reset the number every builder plugs into their unit economics, and the founders who win the next two quarters are the ones who re-run that math this week.

This dispatch reflects the AI model market as of September 2026. Model prices, licenses, and rankings change fast, so verify the current rates on each provider's own pricing page before you commit a budget.