If you run a coding agent, the number that should reach you first is $0.14 per million input tokens — because on July 31 DeepSeek released a budget model at that price and claimed it now beats its own flagship on every agent benchmark it lists. The claim is vendor-stated and unreproduced. The price is not. That gap is the whole story, and it tells you exactly what to do next.
The short version: DeepSeek shipped V4-Flash-0731 — an official public-beta API and MIT-licensed weights. It is not a new model. Same 284B-parameter mixture-of-experts, 13B active per token, 1M-token context. What changed is a fresh post-training pass, and DeepSeek says that pass alone pushed the cheap Flash build past its V4-Pro-Preview on all nine agent and coding-agent benchmarks. Headline: DeepSWE from 7.3 to 54.4. Every one of those scores is run on DeepSeek's own harness, which wasn't public on release day. Treat the benchmarks as a hypothesis; treat the price as fact.
What actually shipped#
This is the part that's easy to verify and easy to under-read. DeepSeek did not scale the model, change the architecture, or add parameters. The July 31 changelog and the Hugging Face model card both say the structure and size are unchanged from the preview: a 284B/13B-active MoE, 1M context, MIT license. The release is a re-post-training — the phase that shapes how a model uses what it already knows, not what it knows.
Two operational upgrades came with it, and these are the ones you can act on today:
- Responses API + Codex. The 0731 build natively speaks the OpenAI Responses API format and is adapted for Codex. If you've already wired an agent to that surface, pointing it at
deepseek-v4-flashis a base-URL-and-model-name change, the same shape as migrating off the retired deepseek-chat/reasoner IDs last week. - Budget-tier pricing at agent scale. $0.14 / 1M input on a cache miss, $0.0028 on a cache hit, $0.28 / 1M output, 2,500 concurrency. For a long-running agent loop that re-reads the same context every turn, the cache-hit price is the line that moves your bill.
The benchmark claim — and why you should half-believe it#
Here is DeepSeek's own table. It is genuinely striking, and it is genuinely un-refereed.
The re-post-trained Flash reportedly beats V4-Pro-Preview on all nine listed benchmarks: Terminal-Bench 2.1 82.7 (vs 72.1), DeepSWE 54.4 (vs 7.3 — the ~645% jump doing the headline work), DSBench-FullStack 68.7 (vs 37.0), plus Cybergym 76.7, NL2Repo 54.2, Toolathlon 70.3, and DSBench-Hard 59.6. On Agents' Last Exam it posts 25.2, which DeepSeek places roughly level with Opus-4.8's 25.7.
A budget model matching a frontier model on an agent benchmark isn't a leaderboard result until someone who doesn't sell the model runs the test.
Now the caveats, because they're load-bearing. Every figure is vendor-run on DeepSeek's own harness, and that harness was not public as of July 31 — the company said it's coming "soon." A 7.3-to-54.4 jump on DeepSWE is exactly the kind of move that can be real or can be an artifact of a changed test harness, a friendlier scaffold, or a baseline chosen to flatter. You cannot tell which from a vendor table, and neither can we. This is the same discipline we walked through in How to Read an Agent-Memory Benchmark: when the lab that trained the model also built and ran the eval, the score is a claim about the harness as much as the model. We turned that into a five-question checklist — how to read a vendor's agent-benchmark table before you believe it — worked on exactly these DeepSeek numbers.
What a founder should do this week#
Do not swap production on a press-release table. Do run the bake-off, because the price makes it cheap to find out:
- Pull two or three tasks that mirror your real agent workload — not DeepSWE, your repo, your tool-call chain. Vendor benchmarks rarely predict your failure modes.
- Compare cost-per-completed-task, not headline score. A model that's 10% weaker at a third of the price can still win on the metric your budget cares about, the same lens we used for Qwen3-7-Flash vs Gemini 3.6 Flash.
- Watch for third-party reproduction before you standardize. When DeepSeek's harness ships and independent runs land, the DeepSWE number either holds or it doesn't. Standardize after that, not before.
The pattern here is the one this desk keeps flagging in the cheap-model floor: open-weight labs are competing on post-training and price, not just parameter count — and the cheap coding model that lands "second only to Opus" is becoming a monthly event. That's great news for your token bill and a standing reason to keep your own eval harness warm. The models will keep claiming they beat the flagship. Your eval is the only thing that gets to decide whether they beat your flagship.
