DeepSeek's free V4.1-Flash-1017 beats paid giants on agentic coding tests
DeepSeek just dropped a massive open-weight model update that smokes GPT-5.6 Sol on coding benchmarks for a fraction of the cost.
Models · Source: Dealroom News
What happened
DeepSeek released V4.1-Flash-1017 on Saturday, a massive post-training refresh of its 552-billion parameter model. The company dropped the weights on Hugging Face under a permissive MIT license. The update claims higher scores across all 11 benchmarks DeepSeek tracks, while keeping API prices completely unchanged. The biggest leaps happened in agentic coding and automation. Terminal-Bench 4.0 jumped 36 percent, and ProgramBench rose 43 percent.
The model now goes head-to-head with the most expensive proprietary giants on the market. According to DeepSeek, the new Flash model beats OpenAI's GPT-5.6 Sol on nine out of eleven tests. It also scored 77 percent on the DeepSWE 1.1 benchmark, edging out Anthropic's Opus 5.0 which scored 74 percent. DeepSeek retained its core architecture, including Causal Encoder-Decoder and FP4 Quantization, but heavily tweaked the model for agentic coding environments.
DeepSeek achieved these gains through extreme output condensation. Using aggressive reinforcement learning, the post-training team penalized the model for rambling and rewarded it for finding the shortest path to a correct answer. The model is now hardwired to stop vomiting paragraphs of explanation and just give you the exact code. Shorter answers mean fewer output tokens. This directly equals faster speeds and significantly cheaper costs for developers running continuous loops.
Key facts
- 552 billion — Backbone parameters in the V4.1-Flash architecture
- $0.15 — Cost per million uncached input tokens off-peak
- 77% — DeepSeek's score on the DeepSWE 1.1 coding benchmark
- 890 bytes — KV cache size per token
- 510.3 GB — Total size of the MIT-licensed model weights
Why it matters
The AI race is fundamentally shifting. It is no longer just a vanity contest over who can build the smartest, most expensive model. It is about who can deliver the most efficient, cost-effective toolkit for developers to actually build products. DeepSeek is giving away flagship-level coding performance for $0.15 per million input tokens off-peak. That completely changes the unit economics of building AI agents. You can now run heavy, iterative coding loops where an agent tests code, fails, and rewrites it, without burning through your startup runway.
The second-order effect here is the democratization of hardware-heavy workflows. DeepSeek managed to compress the KV cache to just 890 bytes per token. That means a full one-million-token context window takes less than a gigabyte of cache in GPU memory. When you combine extreme memory efficiency with MIT-licensed open weights, you remove the gatekeepers. Paid giants like OpenAI and Anthropic will have a very hard time justifying their premium API pricing when developers can get this level of agentic performance for absolute pennies. DeepSeek's post-training lead called it the era of democratized wish-driven coding, and the math backs him up.
For builders
Run massive context windows cheaply
A full one-million-token prompt costs just $0.15 off-peak on the DeepSeek API. If you hit the cache, that price plummets to $0.003. Founders building document-heavy applications or complex agent loops win massive margins here.
Leverage open MIT weights for commercial use
The 510 gigabyte model weights are completely free and MIT-licensed. You can download all 48 files, host them on your own infrastructure, and build commercial products without asking anyone for permission or paying enterprise licensing fees.
Build faster automated coding agents
The model is explicitly trained to stop rambling and output exact code immediately. Developers building automated coding agents save money on expensive output tokens and get much faster execution times during iterative testing loops.
My take
DeepSeek is playing a brilliant game. While OpenAI and Anthropic fight over who has the absolute smartest model, DeepSeek is quietly giving developers the cheapest, most efficient toolkit to actually build things. Wish-driven coding is here, and it does not care about premium API tiers.
Original reporting: Dealroom News. This is my rewrite and opinion.