DeepSeek released the official public beta of V4-Flash-0731 on July 31, and the benchmark numbers are worth a second look. Same 284B MoE architecture, same $0.14 per million input tokens, and a post-training rerun that pushed agent benchmark scores past DeepSeek's own V4-Pro-Preview on every metric the company published.
What Actually Changed
Nothing about the architecture or scale changed. DeepSeek ran a new round of post-training, the phase that shapes how a pre-trained model uses its knowledge, not what knowledge it holds. The underlying structure (284B total parameters, 13B active per token, with CSA and HCA sparse attention layers) is identical to the Flash-Preview build. DeepSeek just worked the behavioral layer on top again.
That alone shouldn't be remarkable. Except for how much it moved the needle.
The Benchmark Story
Here are the numbers DeepSeek published for Flash-0731 vs Flash-Preview vs V4-Pro-Preview:
- DeepSWE: 7.3 (Flash-Preview) to 54.4 (0731). That's a 645% jump. V4-Pro-Preview sits lower.
- Cybergym: 38.7 to 76.7.
- Terminal Bench 2.1: 61.8 (Flash-Preview) to 82.7 (0731), vs 72.1 for V4-Pro-Preview.
- DSBench-FullStack: 37.0 to 68.7.
All nine agent and coding benchmarks DeepSeek published follow the same pattern: Flash-0731 beats Flash-Preview by a wide margin, and beats V4-Pro-Preview too. That last part is the unusual bit.
I haven't run these myself yet. A DeepSWE jump from 7.3 to 54.4 is large enough that I'd want third-party replication before treating it as ground truth. That said, the direction is consistent across multiple independent reports. And separately reported SWE-bench numbers put Flash-Max at 79.0% versus Pro-Max at 80.6%, so the gap between the two tiers is real and narrow.
What This Means for Your Routing Logic
If you've been routing model calls with something like "use Flash for simple tasks, Pro for complex agent work," that heuristic is now wrong for DeepSeek.
The price difference makes this more than academic. V4-Flash is $0.14/$0.28 per million input/output tokens. V4-Pro runs at $0.435/$0.87. You're paying 3x more for a model that Flash-0731 now beats on the benchmarks that matter most for agents. If your routing logic sends planning loops, tool-calling chains, or multi-step code tasks to Pro, you're spending 3x for less performance on those specific workloads.
The practical move: run your agent evals against Flash-0731 this week. Not the synthetic benchmarks, your actual production tasks. If they hold, you can drop Pro from your routing tier for agent workloads.
One real caveat: instruction-following on broad knowledge and long-document reasoning may still favor Pro. The benchmark gap there is tighter, and 79.0% vs 80.6% on SWE-bench is a real difference even if it's narrow.
Zero Migration, Free Upgrade
Here's the part easy to miss. If you're already calling deepseek-v4-flash on the API, you're already on 0731. DeepSeek updated the model behind that API alias in place: same endpoint, same pricing, no version migration required. Your API calls from yesterday are now going to a model that scores 645% higher on DeepSWE than the model from the day before.
That's good if your tasks got harder to solve overnight. It's also a reason to rerun your regression evals now, because model behavior changed even if your code didn't.
I've hit this with other providers too. Model-in-place updates on stable aliases are common, but this one moved things far enough that treating it like a minor patch seems wrong. Log the model alias explicitly in your call records. Rerun evals any time a provider touches a stable alias. That applies here as much as anywhere.
What to Do This Week
- Run your current agent tasks against Flash-0731 and compare outputs to your existing baseline.
- If V4-Pro is in your routing stack, benchmark Flash-0731 against it on your actual workloads before assuming Pro is still earning its 3x price premium.
- Log the model alias explicitly. "deepseek-v4-flash" now means something different than it did 48 hours ago.
The bigger point is this: post-training is increasingly where the real capability gains happen. Same architecture, very different behavior, depending on how the model was fine-tuned and reinforced after pre-training. Architecture announcements alone don't tell the whole story anymore.