On July 31, 2026, DeepSeek promoted V4-Flash-0731 to an official public API build: same 284B total / 13B active MoE as April's preview, with gains coming entirely from a fresh post-training pass. It now beats DeepSeek's own larger V4-Pro preview on agent benchmarks at roughly 1/36 to 1/179 of Claude Opus 4.8 list pricing. Flagship V4-Pro and the in-house agent framework Harness remain unreleased.
This piece is for developers on the DeepSeek API, agent pipeline owners, and buyers comparing Chinese open-weight economics. You get the April–August timeline, vendor pricing vs Artificial Analysis cost-per-task, CSA+HCA architecture notes, Harness sensitivity warnings, a Kimi K3 / GLM-5.2 / Qwen3.8-Max matrix, and a six-step migration checklist. By the end you should know whether to route to Flash or Pro today, which scores to trust, and how to retire legacy model aliases.
01 DeepSeek V4-Flash official timeline: preview to July 31 API beta
It is easy to read "DeepSeek V4 official" and assume a brand-new architecture dropped. It did not. The sequence:
- April 24, 2026: V4 preview ships two MIT open-weight MoE models — V4-Pro (1.6T / 49B active) and V4-Flash (284B / 13B), both with a 1M-token context window.
- July 24, 2026: Legacy aliases
deepseek-chatanddeepseek-reasonerretire; traffic routes to the V4 family. - July 27, 2026: Moonshot AI releases full open weights for Kimi K3 (2.8T total), tightening competitive pressure days before DeepSeek's update.
- July 31, 2026:
deepseek-v4-flashbecomes an official public API beta (build "0731"). Same architecture and parameter count — DeepSeek states gains came entirely from post-training. Open weights land on Hugging Face the same day. The changelog names DeepSeek Harness for the first time. API-only — consumer app and web chat are untouched. - As of August 5, 2026: Official V4-Pro remains unconfirmed. Some Chinese outlets cite unnamed sources for an August 10–20 GA window — not confirmed by DeepSeek; treat as rumor.
The jump came from post-training, not scaling — a 284B/13B model beating a 1.6T/49B sibling on agent tasks signals that late-2026 competition is shifting toward training quality, not parameter count alone.
02 DeepSeek V4-Flash pricing vs Kimi K3 and Qwen3.8-Max (August 2026)
All figures below are vendor-published list prices. DeepSeek has announced a future 2x peak-hour surcharge (9am–12pm and 2pm–6pm Beijing time) with no confirmed effective date yet.
| Model | Status | Total / active params | Input (miss / hit) | Output |
|---|---|---|---|---|
| DeepSeek-V4-Flash-0731 | Official (Jul 31) | 284B / 13B | $0.14 / $0.0028 | $0.28 |
| DeepSeek-V4-Pro | Preview only | 1.6T / 49B | $0.435 / $0.003625 | $0.87 |
| Kimi K3 | Open weights (Jul 27) | 2.8T / ~104B (community est.) | $3.00 / $0.30 | $15.00 |
| Qwen3.8-Max | API GA (Aug 2); weights pending | 2.4T / 95B | $2.00 / ~$0.17–0.25 | $6.00 |
Artificial Analysis Intelligence Index and average cost per task (independent firm, reported via financial outlets) tell a different story than vendor agent scores:
| Model | Intelligence Index | Avg. cost per task |
|---|---|---|
| DeepSeek-V4-Flash-0731 | 50 | $0.03 |
| Kimi K3 | 57 | $0.86 |
| GPT-5.6 Sol | 9+ points above Flash | $1.86 |
| Claude Fable 5 | 9+ points above Flash | $3.15 |
V4-Flash is not the highest-scoring model on the independent index, but per-task cost is roughly 1/29th of Kimi K3 and 1/105th of Claude Fable 5 — DeepSeek is optimizing for "good enough intelligence at a price nobody else can match."
03 How DeepSeek squeezed more from the same model: Harness and hybrid attention
Architecture unchanged — training data and post-training did the work. DeepSeek's technical report ("DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence") describes three structural choices carried from April:
- Hybrid attention (CSA + HCA): marketed as DSA sparse attention to cut compute and memory at long context.
- Manifold-Constrained Hyper-Connections (mHC): an enhancement over standard residual paths.
- Muon optimizer: used for faster convergence and training stability.
Vendor-reported efficiency at 1M context: V4-Pro needs only 27% of V3.2 per-token FLOPs and 10% of KV cache footprint — not independently reproduced yet.
Harness debut. July 31 marked the first official mention of DeepSeek Harness, an in-house agent execution framework positioned against Claude Code. Every agent benchmark DeepSeek published (Terminal Bench 2.0, Toolathlon, etc.) used Harness minimal mode (not yet public) at max effort, top_p 0.95, temperature 1.0. The changelog warns scores are "extremely sensitive to harness choice" — worth taking literally.
21st Century Business Herald, citing overseas developer feedback, reported low input cache-hit rates and occasional safety-classifier timeouts on the official build — a reminder that compute budgets still cap what post-training alone can fix.
04 V4-Flash vs Kimi K3 vs Qwen3.8-Max: six-step rollout checklist
This release lands in the densest stretch of Chinese open-weight launches this year. For agent and batch workloads, the buying question is no longer "who tops the leaderboard" but "who keeps bills low at an acceptable intelligence floor."
- Retire legacy aliases: Since July 24,
deepseek-chatanddeepseek-reasonerare gone — route todeepseek-v4-flashordeepseek-v4-pro(preview). - Confirm your surface: Build 0731 is API-only; app and web chat may lag — treat API changelog as production truth.
- Keep SDK compatibility: OpenAI ChatCompletions and Anthropic-shaped clients need no code changes;
deepseek-v4-flashpoints at the new official build automatically. - Separate benchmark classes: SWE-bench Verified and similar third-party suites carry more weight than Terminal Bench 2.0 scores tied to unreleased Harness.
- A/B on your own repos: Compare Flash, Kimi K3, and Qwen3.8-Max on real agent workflows — success rate and token spend — not vendor slide decks alone.
- Track V4-Pro and Harness GA: Official text is only "as soon as possible." Any August 10–20 date you see online is rumor until DeepSeek's changelog confirms it.
Verify pricing and model names against DeepSeek's official API documentation after publication:
05 Benchmark caveats, the Chinese "kill line," and agent infrastructure
- Terminal Bench 2.0 (vendor-reported): V4-Flash-0731 at 82.7 vs V4-Pro preview 67.9 — measured with unreleased Harness minimal mode; treat as framework-specific, not portable to Claude Code or Cursor.
- Price multiples vs Claude Opus 4.8 (per 21st Century Business Herald): roughly 36x cheaper on cache-miss input, 179x on cache-hit input, 89x on output — vendor list prices, not an audit.
- "Kill line" (斩杀线): Chinese developer slang for DeepSeek's "good-enough capability plus rock-bottom price" bar — competitors that neither clearly beat it on quality nor undercut it on cost risk fading from routing tables.
FAQ
- Is DeepSeek V4 open source? Yes — V4-Flash-0731 ships as MIT weights on Hugging Face.
- Flash or Pro for daily use? Flash-0731 for bulk and agent pipelines; Pro preview only if you need deeper reasoning and budget is loose — re-evaluate when official Pro ships.
- Is official V4-Pro live? Not as of August 5 — only Flash-0731 is official, API-only.
- Can I trust the benchmark numbers? Partially — widely adopted third-party suites yes; Harness-dependent agent scores need independent reproduction.
Relying entirely on cloud API calls leaves you exposed to latency, billing spikes, and routing churn. Running open weights locally hits unified memory ceilings, 24/7 stability, and multi-node scaling walls. For teams that need full macOS to run Claude Code, OpenCode, OpenClaw, and Xcode CI/CD with always-on agent nodes, CALMVPS bare-metal Mac Mini M4 rental is usually the better production fit: dedicated Apple Silicon, six-region nodes, 120-second delivery. See CALMVPS pricing.
Verify latest pricing, benchmarks, and V4-Pro/Harness status before shipping — figures current as of August 5, 2026.