What shipped July 31 — and what still hasn't
On July 31, 2026, DeepSeek promoted V4-Flash-0731 from preview to an official API beta. The model ID, parameter count, and core architecture are unchanged from the April release — 284 billion total parameters, 13 billion active per token, 1M-token context. What moved is post-training quality: DeepSeek claims the Flash variant now beats V4-Pro preview on nine agentic and coding benchmarks, despite being the smaller sibling. V4-Pro's official GA release and the Harness agent framework DeepSeek used to generate those scores are still not publicly available as of August 5.
| Date | Event | Key detail |
|---|---|---|
| April 24 | V4 preview + MIT open weights | V4-Pro (1.6T/49B active) and V4-Flash (284B/13B active) ship together; CSA+HCA hybrid attention architecture disclosed |
| July 24 | Legacy API names retired | deepseek-chat and deepseek-reasoner permanently removed; migration to deepseek-v4-flash / deepseek-v4-pro required |
| July 27 | Kimi K3 weights open | Moonshot publishes full 2.8T MoE weights on Hugging Face, raising the open-weight bar days before Flash goes official |
| July 31 | V4-Flash-0731 official API beta | Same 284B/13B architecture; post-training gains only; Harness agent benchmarks first mentioned |
| August 3 | Qwen3.8-Max GA | Alibaba ships 2.4T/95B MoE API; competes on price and multimodal reach |
| August 5 (today) | Still unreleased | V4-Pro official GA, Harness framework, and independent benchmark reproductions |
Assuming Flash got bigger: V4-Flash-0731 is not a new parameter count — every benchmark gain is post-training, not architecture scale.
Treating Harness scores as reproducible: Terminal Bench 2.0 (82.7 vs V4-Pro preview 67.9) and other agent numbers were run through an unreleased Harness "minimal mode" — you cannot replicate them today.
Waiting for V4-Pro GA to decide: Pro official release is still pending; Flash may already cover your workload at a fraction of the cost.
Ignoring cache-hit pricing gaps: Early adopters report inconsistent Prompt Cache behavior; list $0.0028/M cache-hit input only applies when caching actually works.
Believing the August 10–20 GA rumor: Community chatter points to mid-August V4-Pro GA — DeepSeek has not confirmed any date.
V4-Flash pricing and specs vs Kimi K3, Qwen3.8-Max, and V4-Pro preview
| Spec | V4-Flash-0731 |
|---|---|
| Official API date | July 31, 2026 (beta) |
| Total / active parameters | 284B / 13B |
| Architecture change since preview | None — post-training only |
| Context window | 1M tokens |
| API pricing (per 1M tokens) | $0.14 input (cache miss) / $0.0028 input (cache hit) / $0.28 output |
| Artificial Analysis Intelligence Index | 50 |
| Artificial Analysis cost per task | ~$0.03 |
| Harness Terminal Bench 2.0 | 82.7 (vendor-run, unreleased Harness) |
| Open weights | MIT license; April preview weights on Hugging Face |
| Model | Lab | Total / active | Context | Price in (miss/hit) | Price out | AA Index | AA $/task |
|---|---|---|---|---|---|---|---|
| V4-Flash-0731 | DeepSeek | 284B / 13B | 1M | $0.14 / $0.0028 | $0.28 | 50 | $0.03 |
| V4-Pro preview | DeepSeek | 1.6T / 49B | 1M | $0.435 / $0.003625 | $0.87 | — | — |
| Kimi K3 | Moonshot AI | 2.8T / ~50B | ~1.05M | $3 / $0.30 | $15 | ~57 | $0.86 |
| Qwen3.8-Max | Alibaba | 2.4T / 95B | 1M | $2 / — | $6 | — | — |
| GPT-5.6 Sol | OpenAI | Closed | — | — | — | — | $1.86 |
| Claude Fable 5 | Anthropic | Closed | 1M | — | — | — | $3.15 |
At Artificial Analysis per-task pricing, V4-Flash costs $0.03 per Intelligence Index task — roughly 29× cheaper than Kimi K3 ($0.86), 62× cheaper than GPT-5.6 Sol ($1.86), and 105× cheaper than Claude Fable 5 ($3.15). The score gap is real; so is the price gap.
Architecture unchanged, Harness new: CSA+HCA, mHC, Muon, and the agent framework gap
Why Flash can beat Pro without more parameters. V4-Flash-0731 keeps the same hybrid attention stack DeepSeek disclosed in April: Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA), collectively branded DSA. At 1M-token context, DeepSeek claims inference FLOPs drop to 27% of V3.2 and KV-cache memory to 10% — vendor figures, not independently verified for the 0731 checkpoint. Manifold-constrained hyper-connections (mHC) stabilize the 43-layer residual stream; the Muon optimizer (Newton-Schulz orthogonalized gradients) drove faster post-training convergence. None of this changed on July 31 — only the SFT/RL recipe did.
Harness is the missing piece. DeepSeek first mentioned Harness on July 31 alongside the Flash official release. It is an agent orchestration framework — not yet public — that ran every headline agent benchmark in the launch materials. Terminal Bench 2.0 scored 82.7 for V4-Flash-0731 vs 67.9 for V4-Pro preview, both through Harness "minimal mode." Without the framework, those numbers are vendor claims tied to an unreleased toolchain. DeepSeek positions Harness as the runtime that unlocks Flash's post-training gains in real agent workflows — but until it ships, you are buying API access to a model whose best scores depend on software you cannot run.
How this compares to open-weight rivals. Kimi K3 shipped full weights July 27 with MoonEP and FlashKDA infrastructure — reproducible on your own hardware. Qwen3.8-Max GA'd August 3 with a lower API price ($2/$6) but no weights yet. V4-Flash sits in the middle: MIT-licensed weights from April, official API beta from July 31, and agent scores that require unreleased Harness.
Note: Every agent benchmark in the July 31 launch deck was Harness-run. No neutral platform has reproduced Terminal Bench 2.0 or the nine-benchmark sweep as of publication.
Six-step checklist: API access, caching, and pre-migration due diligence
Confirm model ID: Use deepseek-v4-flash (or the dated deepseek-v4-flash-0731 if your dashboard shows it) — legacy deepseek-chat was retired July 24.
Enable Prompt Cache: Reuse system prompts and tool definitions to target the $0.0028/M cache-hit input rate; verify hits in your billing dashboard — early reports of inconsistent cache behavior are worth watching.
Route by task complexity: Flash for routing, classification, and high-frequency calls; escalate to V4-Pro preview only when Flash fails on your own workload tests.
Do not migrate on Harness benchmarks alone: Run your real repos, long-context jobs, and agent loops against Flash before cutting over from Kimi K3 or Claude.
Compare total cost, not list $/M: Artificial Analysis per-task cost ($0.03) is more useful than raw token pricing when workloads include thinking tokens and tool round-trips.
Plan agent hosting separately: Harness is unreleased; for 24/7 OpenClaw or custom agent runners, see the help center.
from openai import OpenAI
client = OpenAI(
api_key="your_deepseek_api_key",
base_url="https://api.deepseek.com"
)
response = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[
{"role": "system", "content": "You are a coding agent. Reuse this prefix for cache hits."},
{"role": "user", "content": "Refactor this Swift module for async/await..."}
],
extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)Controversies, the kill line, three hard numbers, and what to do next
Chinese AI Twitter coined zhǎn shā xiàn — literally "kill line" — for the price-performance threshold where a cheaper model makes paying for a frontier closed API irrational. V4-Flash's $0.03/task Artificial Analysis figure is the latest candidate. Community nicknames for DeepSeek leadership also flipped around the July 31 launch: founder Liang Wenfeng was briefly called "Liang Baikai" (roughly "Liang who defeats everyone") before skeptics countered with "Liang Sheng" ("Liang who loses") once Harness dependency and cache issues surfaced. Neither label is official; both capture the hype cycle.
Funding rumors circulated the same week — unverified claims about a new DeepSeek financing round. More concrete: Chinese chip stocks barely moved on July 31, a sharp contrast to the January 2025 DeepSeek-R1 rally that wiped billions off Nvidia's market cap. Markets appear to be pricing V4-Flash as an incremental post-training upgrade, not a paradigm shift.
82.7 / 67.9: Terminal Bench 2.0 via unreleased Harness — Flash beats Pro preview by 14.8 points, but neither score is independently reproducible.
$0.03 / 105×: Artificial Analysis per-task cost vs Claude Fable 5 ($3.15) — the "100× cheaper" headline is directionally right on cost, not on capability parity.
27% / 10%: Vendor-claimed 1M-context FLOPs and KV-cache vs V3.2 — architecture efficiency unchanged since April; July 31 gains are training, not silicon.
Caution: Harness-dependent benchmarks, inconsistent cache hits, unconfirmed August GA dates, and unverified funding rumors all qualify the July 31 narrative. Test your own workload before treating Flash as a Claude replacement.
Stacking alternatives: closed API only makes long-horizon agent cost hard to control at Fable 5 pricing; waiting for V4-Pro GA delays a model that may not beat Flash on your tasks; running agent loops on a personal Mac hits sleep, network, and queue limits. For iOS CI/CD, multi-model A/B testing against Kimi K3 and Qwen3.8-Max, and production-grade agent automation, KVMNODE dedicated Mac Mini M4 cloud rental is usually the better fit: exclusive Apple Silicon, sudo access, multi-region nodes, daily/weekly/monthly billing. See the pricing page and order page.
Data as of August 5, 2026 · Sources: DeepSeek API pricing and July 31 announcement, Artificial Analysis Intelligence Index, Moonshot Kimi K3 release, Alibaba Qwen3.8-Max GA materials