DeepSeek's Budget Flash Model Just Beat Its Own Flagship on Every Agent Benchmark

DeepSeek quietly shipped the official V4-Flash API on July 31, and the headline is unusual: the cheaper model now beats the flagship.

DeepSeek-V4-Flash-0731 scores higher than V4-Pro-Preview on all nine agent and coding benchmarks the company published. Terminal Bench 2.1 sits at 82.7 versus Pro’s 72.1. DeepSWE jumped from 7.3 in the preview to 54.4 in the official release — a 645% improvement from the same architecture through post-training alone. Cybergym went from 38.7 to 76.7. NL2Repo from 39.4 to 54.2.

The kicker: nothing changed structurally. Same 284B MoE with 13B parameters activated per token. Same model string. Same endpoint. No code changes required — existing integrations just pick up the new behavior.

What actually changed

The gains are entirely from a fresh post-training pass targeting agent behavior. DeepSeek describes it as retraining for tool use, step-by-step planning, and multi-turn task recovery. The result is a model that completes agentic loops more reliably, recovers from partial failures, and follows tool schemas more precisely.

For coding specifically, the new V4-Flash adds native support for the Responses API format and has been adapted for Codex. That means it works with Codex without a translation layer. DeepSeek says support for V4-Pro in Codex is coming “early August.”

The pricing math

V4-Flash costs $0.14 per million input tokens (cache miss), $0.0028 on a cache hit, and $0.28 per million output tokens. V4-Pro-Preview runs at $0.87 per million output tokens. For typical agentic workloads with long outputs and moderate input reuse, Flash costs roughly one-third as much.

If the benchmark numbers hold in real deployments, most teams would be leaving money on the table by continuing to route to Pro by default.

Open weights, no gating

The model is MIT-licensed and available as open weights. Self-hosting needs around 110GB of memory at 3-bit quantization, or a 4×GB300 node for full precision. That’s within reach for mid-size engineering teams running their own inference.

DeepSeek made a point of noting that only V4-Flash changed on July 31. V4-Pro and the web app are untouched. “An official V4-Pro release is promised as soon as possible,” per the release post, but no date was given.

Re-run your benchmarks

One explicit caution from DeepSeek: the post-training changes affect agent behavior in ways that may not be obvious from static tests. If you’ve built evaluation suites against V4-Flash-Preview, you should re-run them. The model string is the same, but the behavior can shift.

For teams building coding agents or autonomous workflows on top of DeepSeek, this is a meaningful moment. A model that was the cost-saving fallback option now leads on the benchmarks that actually matter for agentic use.

Sources: DeepSeek V4-Flash GA post · MarkTechPost coverage · TechTimes