The DeepSeek V4 Flash update that landed today is the cheapest serious agent model getting dramatically better at the exact work builders use it for. Hacker News noticed, and so should anyone running agent loops on a budget.
DeepSeek moved V4-Flash-0731 into public beta as the official release of the API model, and the changelog reads like a direct answer to every complaint people had about the preview. Same architecture, same size, same price. Just re-post-trained, with agent benchmarks that DeepSeek says now blow past V4-Pro-Preview.
What DeepSeek V4 Flash 0731 changes
The headline numbers are the agent benchmarks. Terminal-Bench 2.1 sits at 82.7, which lands right next to models that cost dozens of times more. DeepSWE, the software engineering benchmark, hits 54.4. Cybergym scores 76.7. Toolathlon verified comes in at 70.3, and DSBench-Hard, DeepSeek’s internal hard-problem coding set, reaches 59.6.
For context, GPT-5.6 Terra scored 84.3 on Terminal-Bench 2.1 at launch and costs $2.50 input and $15 output per million tokens. DeepSeek’s flash model claims 82.7 on the same benchmark at $0.14 input and $0.28 output. That’s not a rounding error, it’s a pricing gap of roughly an order of magnitude for near-comparable agentic coding performance.
One honest caveat: these are DeepSeek’s own numbers, tested with its own harness in minimal mode at max effort. Independent evals will tell the real story. But the direction is consistent with every public benchmark release from this lab over the past year.
The Codex play is the interesting part
Beyond the benchmark bump, V4-Flash-0731 natively supports the Responses API format and is specifically adapted for Codex. That’s a deliberate move at the agentic coding crowd. OpenAI’s coding agents are the default entry point for a huge slice of developers, and DeepSeek V4 Flash is making itself a drop-in provider for that workflow with a fraction of the token cost.
The pricing docs back it up. Cache hits run $0.0028 per million tokens, which is nearly free for agent loops that reuse large system prompts or codebases. Artificial Analysis puts the blended rate around $0.06 per million tokens, against roughly $0.90 for GLM-5.2 and $0.70 for Kimi K2.6. For a recurring agent with a big stable prompt, that changes the economics of letting it run long.
What didn’t change
DeepSeek kept this update scoped tight. The V4-Pro API and the app and web models are untouched, and the official V4-Pro release follows soon. The peak-valley pricing structure from the June launch still applies, so off-peak traffic stays even cheaper, and the 2x peak-hour multiplier still punishes anyone who doesn’t schedule around it.
That scoped release is worth paying attention to. DeepSeek is iterating on the cheap model first, using it as the test bed for agent capability, then presumably carrying the lessons into the flagship. That’s the opposite order from every Western lab, which ships the big model and distills down.
What it means for builders
If you’re building agentic software, the takeaway is that the price-performance floor keeps dropping. A model that can hold its own on Terminal-Bench for $0.14 per million input tokens makes whole categories of automation viable that were previously too expensive to run at volume.
The broader pattern is the same one I called out when OpenAI slashed Luna pricing last week. Frontier pricing is in a deflationary spiral, and the Chinese labs are setting the pace at the bottom. OpenAI cut Luna by 80% and DeepSeek still undercuts it by a wide margin. Google and Anthropic are responding in their own ways, but nobody is closing the gap from the top.
The GLM 5.2 coverage on this site made the same point a month ago: the open-weight Chinese models are no longer a generation behind on the benchmarks that matter for actual work. DeepSeek V4 Flash 0731 is the strongest data point yet.
The reservation nobody is talking about
DeepSeek’s benchmarks keep improving, but the ecosystem around the model still lags. Tooling, integrations, and enterprise support are thin compared to the big labs. And for anyone building on the API, the peak-hour pricing and the changelog-driven iteration style mean you’re building on a moving target.
None of that changes the core fact: for agent workloads where cost matters, DeepSeek V4 Flash is now the benchmark for value. The update today made it stronger where it counts, and the price didn’t move. I’d expect the independent evals to confirm most of the claim, and I’d expect the V4 Pro release to be the next big story from this lab.
