DeepSeek's cheapest model now beats its own flagship on coding benchmarks, DeepSeek says

DeepSeek's own release notes make an unusual claim: its cheap model now beats its expensive one. DeepSeek-V4-Flash-0731, published to Hugging Face and moved into API public beta on July 31, "outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count," according to the model card. Flash activates 13 billion of its 284 billion parameters per token; Pro activates roughly 49 billion of 1.6 trillion. On paper, that's the smaller model beating the bigger one at the bigger one's own job.

DeepSeek is careful about what changed to get there. The company describes the release as "a training and post-training upgrade, not a new model" — same architecture, same parameter count, same 1-million-token context window as the April preview. Every gain came from retraining on a stronger post-training pipeline, weighted toward agentic and tool-use work rather than raw knowledge.

What moved, and by how much

DeepSWE — a benchmark for autonomous software-engineering agents — moved the most, from 7.3 to 54.4, roughly seven and a half times over. Terminal-Bench 2.1 climbed from 61.8 to 82.7, Cybergym, a cybersecurity-exercise benchmark, doubled from 38.7 to 76.7, and NL2Repo — translating a natural-language request into a working code change across a real repository — rose from 39.4 to 54.2. These are all DeepSeek's own reported numbers, run on DeepSeek's own suite, and they measure a different thing than the independent index below: agentic task completion rather than general reasoning and knowledge. Worth separating the two, because they tell different parts of the story and DeepSeek's release post doesn't distinguish them for the reader.

Where it sits against the field

That last bar is the interesting one. Independently, DeepSeek's own V4 Pro still outscores the newly retrained Flash by six points on the Intelligence Index — the opposite of what DeepSeek's agentic benchmarks show. Both things can be true at once: Pro hasn't been through the same post-training pass Flash just got, so its score reflects an older checkpoint, and the two evaluations are measuring different capabilities in the first place — one broad reasoning and knowledge, the other narrow agentic task completion. Flash beating Pro is a real, specific, reproducible result on DeepSeek's own agentic suite. It is not evidence that Flash is now the stronger model overall, and DeepSeek's release material doesn't claim that either.

On price, the gap to comparable models is real and large. V4-Flash-0731 runs $0.14 per million input tokens and $0.28 per million output tokens, with a 98% discount to $0.0028 on cached input — the unchanged part of a repeated prompt, like a system message or tool schema, billed at a fraction of a cent once it's been seen. That undercuts GPT-5.6 Luna's $0.20/$1.20 by a wide margin at the same 51-vs-50 index tier, and it is roughly a fortieth of Claude Fable 5's $10/$50 top-tier pricing — not a fair comparison of capability, but a real comparison of what a high-volume workload actually costs to run on each.

This isn't Flash's first moment as the default. When DeepSeek retired its legacy deepseek-chat and deepseek-reasoner API aliases on July 24, every call against them started routing permanently to V4-Flash — closing out the migration from V3.2 without any user having to change their code. This retraining lands one week later on the exact model line that migration just made the default, which means the upgrade reaches DeepSeek's existing API traffic immediately rather than waiting on anyone to opt in.

The honest limits: officechai's and digitalapplied's benchmark tables comparing V4-Flash-0731 to Claude Opus 4.8 are their own test runs, not Artificial Analysis's independent index, and DeepSeek's headline comparisons to Pro are DeepSeek's own suite — treat both as vendor- and reporter-run numbers pending independent confirmation, the same caveat that applies to every self-reported benchmark on this beat. Self-hosting the MIT-licensed weights takes roughly 110GB of memory at 3-bit quantization, or a single four-GPU GB300 node at full precision; the API-only 0731 build is what's live today, with the downloadable Hugging Face weights still reflecting the April preview checkpoint.

The release also exposes three reasoning_effort levels — low, high, and max — a knob DeepSeek added so callers can trade latency and cost against depth of reasoning on a per-request basis, rather than picking one fixed tradeoff for an entire deployment. Recommended maximum output climbs to 384,000 tokens at the high and max settings, which matters for agentic workloads that write long plans or large code diffs before finishing. None of the benchmark figures above specify which effort level DeepSeek used to generate them, which is itself worth flagging: a max-effort score and a low-effort score from the same model can differ by a wide margin, and the release materials don't say which one produced the numbers in the tables.

Zoom out and this is one release in a pattern, not an isolated event: Moonshot AI's Kimi K3 went fully open-weight on Hugging Face on July 27, and Z.ai's GLM-5.2 already undercuts most closed frontier models on price. Three Chinese labs have now shipped open-weight or heavily-discounted models within the same fortnight, each pitched on some combination of near-frontier capability and a fraction of the cost — the shape of a price war that started with DeepSeek's own V3/R1 releases and has only broadened since. What's different about this specific release is that the discount got deeper without the model getting smaller or the license getting more restrictive; DeepSeek retrained the same weights it already had and passed the entire gain through as capability, not as a new pricing tier.