DeepSeek shipped the official public beta of its V4-Flash API on Friday under the build name V4-Flash-0731, and the interesting part isn’t a new model — it’s the same one. Identical architecture, identical parameter count, identical price. DeepSeek reran the post-training, and the budget model now scores higher than the company’s own V4-Pro preview on all nine agent and coding benchmarks it published. Flash activates roughly 13 billion parameters per token. Pro activates about 49 billion.
The published numbers: Terminal Bench 2.1 at 82.7, against 72.1 for V4-Pro-Preview and 61.8 for the earlier Flash preview — and 85.0 for Anthropic’s Opus-4.8, per DeepSeek’s own comparison table. Agents’ Last Exam at 25.2 versus Opus-4.8’s 25.7, a gap of half a point. Toolathlon (verified) 70.3, Cybergym 76.7, NL2Repo 54.2. The eye-catcher is DeepSWE, the GitHub-issue-resolution test, which went from 7.3 on the preview to 54.4 on this build. Nothing in the base model moved: 284 billion total parameters, 13 billion active, a one-million-token context window, MIT license.
Price is where this lands hardest. V4-Flash stays at $0.14 per million input tokens and $0.28 per million output. V4-Pro sits at $0.435 and $0.87. A coding pipeline burning 100 million output tokens a month runs roughly $28 on Flash against roughly $87 on Pro — and Flash now benchmarks better at the agent work. That is the same squeeze OpenAI answered last week with an 80% cut to its cheapest tier, arriving from the other direction.
Our take: The story here is that capability arrived without capital. No bigger cluster, no new architecture, no fresh pre-training run — just another pass of reinforcement learning against a verifiable reward, because code either compiles and passes tests or it doesn’t. That is the cheapest kind of progress there is, and it is available to anyone holding open weights. Two practical consequences for anyone building on these APIs. First, the model behind your endpoint changed on Friday whether you asked for it or not — same name, same key, different behavior — so if you are not pinning builds and running your own regression suite, you are shipping a vendor’s Tuesday to your customers. Post-training gains are lopsided; some tasks improve and some quietly regress, and no public benchmark will tell you which yours is. Second, stop pricing budget tiers as permanently inferior. The gap between cheap and flagship is now a post-training decision, not a physics one.
The number to distrust
That 645% DeepSWE jump is self-reported, measured with DeepSeek’s own harness, which the company says it will release “soon.” Nobody outside can replicate it yet. For calibration: V4-Pro posted 80.6% on SWE-bench Verified under a vendor-controlled harness, then scored 8% pass@1 on an independent DeepSWE run by yage.ai — against 70% for GPT-5.5 and 54% for Claude Opus 4.7. DeepSWE is the tighter test, with a verifier that false-positives at 0.3% versus roughly 8.5% for SWE-bench, so 54.4 on it means more than 80.6 on the other. It still isn’t independent. DSBench-Hard and DSBench-FullStack are explicitly internal sets. We have watched this exact gap open before, and it usually closes downward.
What to watch
- An independent re-run. yage.ai or a comparable lab putting 0731 through its own harness is the only number that settles this. Until then, treat 54.4 as a claim.
- Your effective cost, not your token cost. Thinking mode is on by default and reasoning tokens bill at the output rate even though you never see them. Measure dollars per completed task before you migrate anything.
- Which build you are actually getting. At announcement, only DeepSeek’s own endpoint served 0731. Traffic routed through aggregators may still be hitting the April preview — the model that scored 7.3.
- V4-Pro’s official release. DeepSeek says it follows “as soon as possible.” If the same re-post-training runs on 49 billion active parameters, the ceiling moves again — from a company that told investors to stand down on a $74 billion round a week ago.
- Where the data sits. The hosted API routes prompts to servers in China, subject to the 2017 National Intelligence Law. The MIT license means self-hosting removes that; it does not remove the engineering bill.