DeepSeek’s latest model was reportedly tested at 90% lower cost while pushing out more than 237 tokens per second, according to a widely shared post on X.
The model in question is DeepSeek-V4.1-Flash, and it is quietly forcing an uncomfortable question onto every pricing team at the big AI labs: what exactly are customers paying extra for?
What DeepSeek actually shipped
DeepSeek launched V4.1-Flash on September 10, 2026. The model carries 552 billion total parameters. It uses a sparse mixture-of-experts design, or MoE, where only about 8 billion parameters are active when the model processes input and roughly 16 billion kick in when it generates output.
It also ships with a 1 million token context window, allowing the model to hold a very large amount of text in working memory at once, such as a sprawling codebase or a stack of lengthy documents.
Native image processing comes built in. The weights are open, released under the MIT license on Hugging Face, so developers can download and run the model themselves rather than renting access through an API.
On October 17, 2026, DeepSeek pushed an update, labeled V4.1-Flash-1017, under the same API. The refresh improved performance metrics and cut task runtimes, with one source reporting inference speeds above 237 tokens per second.
The benchmarks and the bill
On the DeepSWE v1.1 benchmark, a software engineering test, V4.1-Flash reportedly scored 74.2. Anthropic’s Claude Opus 5 came in at 74.0 on the same test. OpenAI’s GPT-5.6 Sol posted 73.0.
Independent tests show V4.1-Flash still trails leading rivals on some of the most complex reasoning tasks. Coding and agentic work are its strong suits, while the hardest multi-step logic problems remain territory where the closed labs hold an edge.
DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours. Those rates double during peak usage times.
Cached inputs get even cheaper. When a request reuses content the system has already seen, input costs can fall to between $0.003 and $0.006 per million tokens. That cache-hit pricing is approximately 30 times lower than certain frontier closed models.
Background: DeepSeek’s habit of undercutting
V4.1-Flash is positioned as the successor to DeepSeek-V4-Pro. The company has largely retired V4-Pro in favor of the new model, a shift that took effect from mid-September 2026.
The MoE architecture is central to that strategy. By activating only a small slice of the network for each request, the lab can offer a very large model without paying the full compute cost every time someone types a prompt.
What this means for the AI market
Because V4.1-Flash is available under the MIT license, enterprises can self-host it, avoiding API dependency entirely. That flexibility matters for companies with data residency rules or a general reluctance to send sensitive information to an outside provider.
Benchmarks measure specific tasks under specific conditions, and a lead of a few tenths of a point on DeepSWE v1.1 is not the same as dominance across real-world deployments. Peak-hour pricing doubles the headline rates, so actual costs depend heavily on when and how a team uses the model.
The things to watch next: whether independent testers confirm the 237 tokens per second figure across broader workloads, how quickly developers migrate production systems, and whether the major Western labs answer with lower prices or a sharper reasoning advantage.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.