What shipped on July 31
DeepSeek’s V4 Flash — the cost-efficient sibling of the flagship V4 Pro — graduated from preview to official public-beta release on 31 July 2026, per DeepSeek’s API change log (the OrcaRouter guide reproduces the release notes verbatim). The build, marked -0731, doesn’t change the shape of the model: it is still a mixture-of-experts design — a large pool of specialists where only a fraction fire on any given token, which is what keeps inference fast. What changed is post-training — extra fine-tuning that DeepSeek says materially improved the model’s agentic, coding, and tool-calling capabilities. Two practical additions travel with the upgrade: native support for OpenAI’s newer agent-friendly request format, and a Codex-flavoured tuning pass. The model still speaks the usual chat and message interfaces, so existing integrations don’t need rewriting.
How much better, on paper
70.3Toolathlon (verified) on the new V4 Flash build — up from 51.8 on the April preview. DeepSeek’s own harness.
DeepSeek’s reported numbers, run on its own harness in max-effort mode (DeepSeek’s API change log, via OrcaRouter), show the gains concentrated on the agentic and coding axes — exactly the workloads a Flash-tier model is meant to carry. Tool-calling reliability and end-to-end agent automation both climbed sharply, and the new build adds several coding benchmarks the preview didn’t report. Independent context: Artificial Analysis scored the April preview (2026-06-25) at around 56.2 AA Coding, 40.3 AA Intelligence, and 50.0 AA Math — a coding-leaning generalist above the open-model median. Vendor-harness numbers run optimistic; the full benchmark list is in the box, and you should run your own evals before committing production traffic.
Where V4 Flash sits in the DeepSeek ladder
DeepSeek’s V4 line is a two-rung ladder. V4 Pro is the flagship — heavier, slower, more expensive, for the hardest reasoning. V4 Flash is the efficiency tier: same shape as a frontier model, far less active compute per token, a fraction of the price. After this upgrade, Flash is also a substantially better agent. The fact that DeepSeek invested in Responses-API support and a Codex adaptation for Flash — not Pro — signals where the company expects the volume to run: everyday agentic and coding traffic on Flash, hardest reasoning reserved for Pro. That mirrors the wider cheap-default-plus-flagship pattern across the industry — Anthropic’s Fable versus Opus, OpenAI’s GPT-5.6 Luna versus Sol, Google’s Gemini Flash versus Pro — and is why routing by difficulty beats defaulting to the biggest model every time.
For a deeper look at the earlier cost story on V4 Flash, see DeepSeek Flash breaks the agent cost curve. On running it on your own kit, DeepSeek V4 Flash reaches older Nvidia cards and Codex speeds up DeepSeek V4 Flash locally are the adjacent reads.
How to try it this afternoon
V4 Flash is API-served, not a desktop model. Three places to reach it today:
- DeepSeek API directly — model id
deepseek-v4-flash; Responses API, ChatCompletions, and Anthropic-style interfaces all work. - Fireworks AI — serverless at $0.14 / $0.028 / $0.28 per 1M tokens (input / cached input / output), with on-demand deployment on dedicated GPUs available (Fireworks). Model path
accounts/fireworks/models/deepseek-v4-flash; the Python client, REST API, and OpenAI’s Python client all work. - OrcaRouter — one OpenAI-compatible endpoint (
deepseek/deepseek-v4-flash) at 0% markup, alongside 185+ other models so you can benchmark V4 Flash against GPT-5.6 Luna, Gemini 3.6 Flash, GLM-5.2, Qwen 3.8 and Kimi K3 in one place (OrcaRouter).
A short, honest test plan: pick a coding-agent task you already run — a 200-line refactor or a test-generation job — and point your existing harness at deepseek-v4-flash (it’s OpenAI-compatible, so it’s a config change, not a rewrite). Cap reasoning effort and tool iterations so you don’t pay for runaway loops, turn on prompt caching if your system prompt is stable, and compare both the output and the bill against your current model. At ten million tokens a month on a 70/30 input/output split, V4 Flash runs around £1.50–£2 — roughly a third of what the same workload would cost on V4 Pro.
The right way to read this release: the cheap option got sharper at exactly the work you’d route to it for. That’s a swap worth making behind an agent harness, not a model you’d pick for the hardest single reasoning call. Pro is still the right tool for that.
Sources & quotes
Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. How we verify →
