Open-weight LLM August 5, 2026 ~20 min read DeepSeek V4 Harness

Is DeepSeek's New Model Really 100x Cheaper Than Claude?
Inside the V4-Flash Benchmarks

July 31, 2026 · V4-Flash-0731 official · Post-training only · Harness teased · Pro still undated

DeepSeek V4 Flash 0731 official open-weight model benchmarks concept

Summary: On July 31, DeepSeek promoted deepseek-v4-flash to an official public API build (tag 0731) with the same 284B-parameter architecture as April's preview—gains came entirely from a fresh post-training pass. Agent scores now rival Claude Opus 4.8 at roughly 1/36 to 1/179 of the price (vendor list rates). Flagship V4-Pro and in-house agent framework Harness remain unreleased. This piece covers the timeline, pricing tables, architecture notes, Chinese open-weight rivalry, benchmark caveats, the "kill line" pricing narrative, and a five-step developer checklist.

01

Four pain points before you switch

  1. 01

    Legacy API names are dead: deepseek-chat and deepseek-reasoner retired July 24—unmigrated production breaks immediately

  2. 02

    "Official" ≠ new architecture: 0731 is a re-trained Flash, not a bigger model—wrong expectations skew routing decisions

  3. 03

    Agent scores are Harness-dependent: Terminal Bench 2.0 figures used DeepSeek's unreleased minimal-mode Harness—don't port them blindly to Claude Code or Cursor

  4. 04

    Pro and Harness still missing: Only Flash is official, API-only; app and web chat untouched; August 10–20 GA rumors are unconfirmed

02

Timeline: April preview to July "official"

DateEvent
Apr 24, 2026V4 preview + MIT open weights: V4-Pro (1.6T/49B active) and V4-Flash (284B/13B active), 1M context
Jul 24, 2026Legacy aliases retired; all traffic routes to V4 family names
Jul 27, 2026Moonshot AI ships Kimi K3 open weights (2.8T total)—direct competitive pressure
Jul 31, 2026V4-Flash-0731 official API beta; weights on Hugging Face; changelog names Harness "coming soon"; API-only
As of Aug 5, 2026Official V4-Pro still "as soon as possible"; media August 10–20 window not confirmed by DeepSeek
03

Core numbers (vendor-published pricing)

ModelStatusTotal / activeContextInput $/M (miss / hit)Output $/MLicense
V4-Flash-0731Official284B / 13B1M$0.14 / $0.0028$0.28MIT
V4-ProPreview only1.6T / 49B1M$0.435 / $0.003625$0.87MIT
Kimi K3Open weights Jul 272.8T / ~104B est.~1.05M$3.00 / $0.30$15.00Modified MIT
GLM-5.2Open Jun 2026~744B / ~40B1MNot verified hereMIT
Qwen3.8-MaxAPI GA Aug 22.4T / 95B1M$2.00 / ~$0.17–0.25$6.00Weights pending

DeepSeek announced a future 2× peak-hour surcharge (Beijing 9–12, 14–18) with no effective date yet. Figures as of August 5, 2026.

04

How performance moved without scaling up

Same architecture, new post-training

V4-Flash-0731 is identical in size and structure to April's preview. DeepSeek attributes agent benchmark jumps entirely to re-run post-training—a 284B/13B model now beats the family's 1.6T/49B preview on multiple agentic tasks, underscoring that late-2026 competition is as much about training data and methods as raw parameter count.

Hybrid attention, mHC, Muon

The technical report describes CSA + HCA hybrid attention ("DSA"), manifold-constrained hyper-connections (mHC), and the Muon optimizer. Vendor claims at 1M tokens: V4-Pro needs 27% of V3.2 per-token FLOPs and 10% of KV cache—pending independent reproduction.

Harness: first in-house agent framework

July 31 marked the first official mention of DeepSeek Harness—file I/O, tool calls, engineering workflows—DeepSeek's answer to Claude Code. Every published agent score (Terminal Bench 2.0, Toolathlon, etc.) used Harness minimal mode (not yet public). The changelog warns scores are "extremely sensitive to harness choice."

05

Side-by-side: China's open-weight scrum

ModelLabTotal paramsAA Intelligence IndexCost / task (AA)
V4-Flash-0731DeepSeek284B50$0.03
Kimi K3Moonshot2.8T57$0.86
GLM-5.2Zhipu~744B~1 pt above FlashNot verified
GPT-5.6 SolOpenAIUndisclosed9+ pts above$1.86
Claude Fable 5AnthropicUndisclosed9+ pts above$3.15

On Artificial Analysis's independent index, Flash isn't the smartest—but per-task cost is ~1/29 of Kimi K3 and ~1/105 of Claude Fable 5. DeepSeek optimizes for "good-enough intelligence at unmatched price," which also explains seven weeks atop OpenRouter usage for the preview.

Decision matrix

ScenarioLean towardWhy
High-volume agents / API batchesV4-Flash-0731Lowest cost; 0731 agent scores beat Pro preview
Max independent indexKimi K3 or closed flagshipHigher AA score, much higher unit cost
Self-host + 1M contextV4-Flash MIT weightsOpen license on Hugging Face
Waiting on HarnessHold for Pro officialFramework unreleased; third-party agents still required
06

Caveats: don't trust headlines blindly

  1. 01

    Harness lock-in: Terminal Bench 2.0 at 82.7 (vs Pro preview 67.9) used unreleased Harness minimal mode—treat as vendor + framework, not portable truth

  2. 02

    Usability reports: Overseas devs cite low input cache-hit rates and occasional safety-classifier timeouts

  3. 03

    Pro/Harness dates: August 10–20 GA rumors are unnamed-source media—not DeepSeek confirmation

  4. 04

    Funding rumors: ~$7.4B round / ~$48.7B valuation reports lack regulatory filings—background only

07

"Kill line" pricing and the chip-stock shrug

Chinese forums nicknamed founder Liang Wenfeng "Liang Empty Promise" when Pro slipped past mid-July; after 0731 over-delivered, "Liang the Sage" returned—a fast sentiment barometer.

Developers also use 斩杀线 (zhǎn shā xiàn)—"kill line": good-enough capability plus rock-bottom price sets a bar competitors must beat on both axes. OpenAI's reported 80% Luna price cut fits that framing. On July 31, Nvidia, Broadcom, and AMD barely moved—markets now treat "DeepSeek efficiency" as routine engineering, unlike the 2025 R1 chip selloff.

Five-step developer checklist

  1. 01

    Point production at deepseek-v4-flash; retire legacy aliases

  2. 02

    Separate vendor agent scores from AA independent index when routing

  3. 03

    Schedule bulk jobs outside announced 2× peak windows (date TBD)

  4. 04

    Validate agents in OpenClaw / Claude Code—don't rely on TB2.0 alone

  5. 05

    Watch changelog for Harness and Pro—API and weights may diverge again

Quotable figures

  • Flash 0731: 284B / 13B active; input as low as $0.0028/M cache-hit
  • vs Opus 4.8: ~1/36 cache-miss input, ~1/179 cache-hit, ~1/89 output (vendor list, per 21st Century Business Herald)
  • Terminal Bench 2.0: 82.7 (Harness minimal mode, vendor-reported)
08

FAQ

Yes—V4-Pro and V4-Flash, including 0731, are MIT open weights on Hugging Face.

List pricing roughly 36×–179× cheaper on input and ~89× on output vs Opus 4.8 per million tokens (vendor rates, not audited).

No confirmed date—changelog says "as soon as possible." August rumors unconfirmed.

SWE-bench Verified is stronger signal. Agent scores need third-party Harness reproduction.

DeepSeek's in-house agent framework (files, tools, engineering tasks)—named July 31, not yet public.

Closing

V4-Flash-0731 proves post-training and pricing can matter more than another parameter bump—but validating OpenClaw multi-model routing, comparing Harness vs Claude Code, or batch-migrating APIs still benefits from a real macOS graphical session. If your daily driver is Windows or Linux, buying a Mac just to click through agent permissions is heavy; SSH alone won't cover system dialogs. Rent a VNCMac remote Mac by the hour, wire up DeepSeek V4 in VNC, run agent workloads in isolation, and stop when the project ends. See Mac plans and our SSH-VNC guide; also read our V4 GA overview and Kimi K3 open weights pieces.

Sources: DeepSeek API docs/changelog, technical report, Artificial Analysis (via financial media), 21st Century Business Herald. Verify pricing and release status before publishing—figures as of August 5, 2026.