DeepSeek's V4-Flash Benchmarks Jumped Without a Bigger Model

DeepSeek pushed an official update to its V4-Flash API into public beta on July 31, 2026, and the DeepSeek V4-Flash benchmarks that came with it are the kind of result that's easy to misread. The headline isn't a bigger model. It's the same model, retrained, and DeepSeek says it now exceeds its own larger flagship on the agent and coding benchmarks it publishes.

The build is called V4-Flash-0731. According to DeepSeek's own release notes and its API changelog, the model keeps the exact same architecture as the earlier V4-Flash-Preview: a 284-billion-parameter mixture-of-experts design with 13 billion parameters active per token, and the same one-million-token context window. Nothing about the model's shape changed. What changed was the training that came after it. DeepSeek describes the update as a fresh post-training pass, and states plainly that the gains are "alignment and agent-behaviour gains, not a new base model."

The DeepSeek V4-Flash benchmarks, before and after

The numbers are the part worth sitting with, and DeepSeek's own release table gives a clean before-and-after against its prior V4-Flash build, not the Pro model. On Terminal Bench 2.1, a benchmark for complex command-line agent work, the April preview build scored 56.9. The retrained 0731 build scores 82.7, a jump of nearly 26 points with no change to model size. On Cybergym, a security-focused agent benchmark, the earlier build scored 38.7 against the new build's 76.7, essentially double. DeepSeek states that this same retrained build now exceeds V4-Pro-Preview, its own larger sibling, across the nine agent and coding benchmarks it has published: Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon, Agent Last Exam, Automation Bench, and two DSBench variants. That claim is worth flagging precisely: DeepSeek states it in its own release notes, but has not published V4-Pro-Preview's matching scores on any of these nine benchmarks, so there's no public table where a reader can check the comparison number against number.

Pricing didn't move either. The API still runs $0.14 per million input tokens on a cache miss and $0.28 per million output tokens, the same rate as before the update. Nothing about calling the model changed: the model string is still deepseek-v4-flash, and existing integrations pick up the new weights automatically. DeepSeek also added native support for the Responses API format and adapted the model specifically for Codex, so the update carries a couple of integration changes alongside the benchmark jump.

Why post-training beat a bigger model

For anyone building on top of large language models, the interesting move here isn't the specific scores. It's the mechanism. DeepSeek had two ways to make V4-Flash more capable at agent tasks: scale it up, or retrain the same model with better post-training. It chose the second, and by its own account the retrained model now outperforms the larger one it might otherwise have had to build to compete.

This matters because "bigger model" and "better model" get treated as synonyms more often than they should. A model's raw parameter count says almost nothing about how well it's been aligned to actually finish a multi-step agent task, hold state across a long tool-use sequence, or recover from an error mid-task, and those are exactly the skills that post-training targets. DeepSeek's claim points directly at this: the smaller, cheaper V4-Flash reportedly now beats the larger, more expensive V4-Pro-Preview on the tasks that matter most to anyone running agents in production, coding and terminal work.

If you're building or operating AI agents, the practical takeaway is to stop assuming the model with the bigger name or the higher price tag is the better pick for your workload. Test the smaller, cheaper option on your actual task before assuming the flagship is worth the premium. That's the same lesson AIOS Guide has covered before with agent benchmarks more broadly: published scores describe a narrow test, not your workload, and the gap between the two can be wide enough to change which model you should actually use.

The vendor-benchmark caveat

One honest caveat belongs in any read of these numbers: seven of the nine benchmarks DeepSeek published are vendor-reported, run on DeepSeek's own harness, with no independent reproduction as of this writing. That doesn't make them false, but it does mean they should be read as DeepSeek's own account of its model's performance rather than an audited result.

The one piece of outside corroboration comes from Artificial Analysis, an independent model-benchmarking firm, which measured V4-Flash-0731 at 50 on its own Intelligence Index, roughly ten points above the earlier V4-Flash build, and recorded a jump from an Elo of 1189 to 1559 on its GDPval-AA v2 agentic benchmark. That's a different benchmark suite than DeepSeek's own, but it points the same direction: a real, independently measured improvement, even if the exact scale of DeepSeek's headline numbers hasn't been reproduced by a third party yet.

What this means if you're choosing a model

Three things worth carrying forward if you're picking a model for agent or coding work right now. First, DeepSeek V4-Flash benchmarks are a live example of open-weight, non-US labs closing the gap with frontier proprietary models through technique rather than raw scale, a pattern that's shown up repeatedly this year as smaller open models get more capable per dollar and per gigabyte of memory. Second, a retrain is cheaper and faster to ship than a new architecture, so expect other labs to run the same play: watch for benchmark jumps on existing model names, not just new releases, since those quiet updates can change your model math without a headline event. Third, treat any single benchmark suite, DeepSeek's or anyone else's, as one data point. Run your own agent's actual tasks against both the model you're using now and any candidate that claims to beat it before you switch, since the benchmarks that matter are the ones that match what you're actually asking the model to do.

Sources: DeepSeek API changelog, July 31, 2026; DeepSeek V4-Flash GA benchmark announcement; Artificial Analysis, DeepSeek V4-Flash-0731 Intelligence Index analysis.