Is DeepSeek V4-Flash Really 100x Cheaper Than Claude?

0731 post-training upgrade · Harness framework · competitor comparison · benchmark caveats · V4-Pro timeline

DeepSeek V4 Flash official release benchmark and API pricing comparison

On July 31, 2026, DeepSeek promoted V4-Flash-0731 to an official public API build: same 284B / 13B-active architecture 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 pricing. If you are weighing benchmark credibility, comparing against Kimi K3 or Qwen3.8-Max, or migrating off retired deepseek-chat aliases, this article delivers a release timeline, core pricing table, Harness breakdown, competitor comparison, six-step API runbook, and benchmark caveats you can act on. Data as of 2026-08-05.

01

What actually shipped on July 31 — and what didn't

It is easy to read "DeepSeek V4 official version" and assume a new model dropped. It didn't. This is the same 284B-parameter architecture with a re-run post-training pass. The flagship V4-Pro official release and DeepSeek's first in-house agent framework, Harness, remain unreleased with no confirmed date.

  1. 01

    April 24, 2026: DeepSeek-V4 preview launches with two open-weight (MIT) MoE models — V4-Pro (1.6T / 49B active) and V4-Flash (284B / 13B active) — both with a 1M-token context window. See our V4 GA breakdown.

  2. 02

    July 24, 2026: Legacy aliases deepseek-chat and deepseek-reasoner are retired; all traffic routes to the V4 family.

  3. 03

    July 27, 2026: Moonshot AI ships full open weights for Kimi K3 (2.8T total parameters), raising competitive pressure days before DeepSeek's own update. See our Kimi K3 open-weight guide.

  4. 04

    July 31, 2026: V4-Flash-0731 enters official public API beta; open weights land on Hugging Face the same day. Changelog names DeepSeek Harness for the first time, calling it "to be released soon." Crucially, this update is API-only — the consumer app and web chat are untouched.

  5. 05

    August 2, 2026: Alibaba Qwen3.8-Max API goes GA; weights still pending. See our Qwen3.8-Max open-source analysis.

  6. 06

    As of August 5, 2026: The official V4-Pro release remains unconfirmed. Some Chinese outlets, citing unnamed sources, report internal testing began the week of July 28 with a possible GA window of August 10–20 — this window is not confirmed by DeepSeek and should be treated as a rumor.

Teams tracking this release often hit five blind spots:

  1. 01

    Misreading "official version": Architecture and parameter count are identical to April's preview — performance gains came entirely from post-training.

  2. 02

    Benchmark framework dependency: Agent scores were measured using Harness's unreleased "minimal mode," not third-party agent tools.

  3. 03

    API-only limitation: Consumer app and web chat have not been updated to the 0731 build.

  4. 04

    Pro version gap: The flagship V4-Pro official release is still unavailable; complex reasoning tasks may require the preview build.

  5. 05

    Funding rumors: Reports of a ~$7.4B funding round and ~$48.7B valuation trace back to unnamed financial-media sources — not confirmed by DeepSeek or regulatory filings.

02

The numbers at a glance: pricing, parameters, and license

ModelStatusTotal / active paramsContextInput (miss / hit, $/M)Output ($/M)License
DeepSeek-V4-Flash-0731Official (Jul 31, 2026)284B / 13B1M$0.14 / $0.0028$0.28MIT
DeepSeek-V4-ProPreview only (Apr 24, 2026)1.6T / 49B1M$0.435 / $0.003625$0.87MIT
Kimi K3Open weights (Jul 27, 2026)2.8T / ~104B (community estimate)~1.05M$3.00 / $0.30$15.00Modified MIT
GLM-5.2Open (June 2026)~744B / ~40B1MNot verifiedNot verifiedMIT
Qwen3.8-MaxAPI GA (Aug 2, 2026); weights pending2.4T / 95B1M$2.00 / ~$0.17–0.25$6.00Open weights promised

All pricing figures above are vendor-published rates. DeepSeek has announced a future 2x peak-hour surcharge (9am–12pm and 2pm–6pm Beijing time) with no confirmed effective date yet. Compared to Claude Opus 4.8: cache-miss input runs roughly 36x cheaper, cache-hit input about 179x cheaper, and output about 89x cheaper per million tokens (per 21st Century Business Herald citing official pricing).

03

Post-training upgrade and Harness: where the performance came from

The architecture didn't change — the training data did

V4-Flash-0731 is identical in size and structure to April's preview. DeepSeek says the entire performance jump on agent benchmarks came from re-running post-training, not scaling up. That cuts against the industry's default assumption that "bigger equals better" — a 284B/13B model is now beating a 1.6T/49B model from the same family on multiple agentic tasks, underscoring how post-training quality is rivaling raw parameter count in late-2026 competition.

Hybrid attention: CSA + HCA, mHC, and the Muon optimizer

DeepSeek's technical report ("DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence") describes three architectural changes carried over from the April preview:

  1. 01

    Hybrid attention (DSA): Combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to cut compute and memory costs at long context lengths.

  2. 02

    Manifold-Constrained Hyper-Connections (mHC): An enhancement over standard residual connections.

  3. 03

    The Muon optimizer: Used for faster convergence and training stability.

DeepSeek claims that at a 1M-token context length, V4-Pro needs only 27% of the per-token inference FLOPs and 10% of the KV cache footprint compared to V3.2. These are vendor-reported efficiency numbers; independent third-party reproduction has not been published yet.

Harness: DeepSeek's first in-house agent framework

July 31 also marked the first official mention of DeepSeek Harness — an in-house agent execution framework for file I/O, tool calls, and multi-step engineering tasks, positioned as DeepSeek's answer to Claude Code. Every agent benchmark DeepSeek published for V4-Flash-0731 (Terminal Bench 2.0, Toolathlon, etc.) was measured using Harness's "minimal mode," which is not yet publicly released, at max reasoning effort, top_p 0.95, temperature 1.0. DeepSeek's own changelog adds a caveat that agent scores are "extremely sensitive to harness choice" — worth taking at face value.

04

Competitor comparison and benchmark caveats: China's open-weight "hexagon war"

ModelLabIntelligence Index (Artificial Analysis)Avg. cost per taskDeepSeek official agent scores
V4-Flash-0731DeepSeek50$0.03Terminal Bench 2.0: 82.7 (Harness minimal mode)
Kimi K3Moonshot AI57$0.86
GLM-5.2Zhipu / Z.ai~1 point above V4-FlashNot verified
GPT-5.6 SolOpenAI9+ points above V4-Flash$1.86Closed source
Claude Fable 5Anthropic9+ points above V4-Flash$3.15Closed source

DeepSeek is not competing for the top of the leaderboard — it is optimizing for "good enough intelligence at a price nobody else can match." Per-task cost is roughly 1/29th of Kimi K3 and 1/105th of Claude Fable 5.

Benchmark caveats worth flagging: ① Terminal Bench 2.0's 82.7 score (vs V4-Pro preview's 67.9) was measured using unreleased Harness minimal mode at max settings; ② overseas developers report low input cache-hit rates and occasional safety-classifier timeouts; ③ V4-Pro / Harness release dates remain unconfirmed media rumors; ④ funding and IPO reports trace to unnamed sources, not official filings.

05

Six-step API selection runbook: from migration to production routing

V4-Flash-0731 supports OpenAI- and Anthropic-compatible APIs and drops into Claude Code, OpenCode, Cursor, and other agent toolchains. Here is a six-step path from evaluation to production:

  1. 01

    Audit legacy calls: Scan code for retired deepseek-chat / deepseek-reasoner aliases and establish monthly call volume and cache-hit-rate baselines.

  2. 02

    Switch model naming: Map deepseek-chat to deepseek-v4-flash (non-thinking); map deepseek-reasoner to Flash thinking mode or deepseek-v4-pro preview. Base URL stays the same.

  3. 03

    Evaluate cache strategy: Monitor input cache-hit rates — overseas community feedback suggests the official build may hit cache less often than expected, directly affecting bills. Test Prompt Caching on long-context workloads first.

  4. 04

    A/B competitor comparison: Run your own samples against Kimi K3 ($3/$15), Qwen3.8-Max ($2/$6), and V4-Flash ($0.14/$0.28) for latency, success rate, and per-task cost.

  5. 05

    Configure agent host environment: Harness is not public yet — use Claude Code or OpenCode as the execution framework in production; re-test Terminal Bench–class tasks once Harness GA lands.

  6. 06

    Build primary/backup routing and monitoring: Route high-volume batch tasks through V4-Flash; keep V4-Pro preview or Claude as fallback for complex reasoning; set cost caps and peak-hour 2x surcharge alerts.

python
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": "user", "content": "Analyze this code..."}]
)
06

Industry context, hard numbers, and where VpsMesh fits

  • The "kill line" effect: Chinese developer circles use zhǎn shā xiàn — literally "kill line" — to describe how DeepSeek's "good-enough performance plus rock-bottom price" sets a bar competitors must beat on capability or undercut on cost. That framing helps explain moves like OpenAI's reported 80% GPT-5.6 Luna price cut around the same period.
  • Community sentiment flip: Before V4-Flash-0731 shipped, forums mocked founder Liang Wenfeng as "Liang Empty Promise" over V4-Pro delays; after the Flash build outperformed expectations, the same communities flipped back to "Liang the Sage" — a telling barometer of how fast sentiment swings in China's AI developer community.
  • Muted chip-stock reaction: Nvidia, Broadcom, and AMD saw no significant movement on July 31 when V4-Flash went official — a contrast to early 2025, when DeepSeek-R1's efficiency claims triggered a global AI-chip selloff. Markets now treat "DeepSeek does more with less compute" as routine engineering, not an automatic bearish signal.
  • OpenRouter call volume: The V4-Flash preview reportedly topped OpenRouter's most-used model ranking for seven consecutive weeks — evidence that large-scale, low-cost agent calls are its core user base, not leaderboard chasers.
  • Terminal Bench 2.0: Official score of 82.7 vs V4-Pro preview's 67.9 — but measured with unreleased Harness; treat as vendor-reported until independently reproduced.

Pure API calls cover V4-Flash agent inference, but iOS signing chains, Xcode local builds, Metal workloads, and 24/7 agent long-runs still need real macOS nodes — laptops cannot reliably host persistent OpenClaw / Claude Code agents, and VMs add performance tax plus EULA risk. For production iOS CI/CD and AI agent automation, VpsMesh Mac Mini cloud rental is usually the better fit: bare-metal Apple Silicon, root access, and predictable monthly cost pair cleanly with DeepSeek API in a cloud-inference-plus-local-build architecture. Local inference options: see our ds4 + V4 Flash local inference guide.

Sources: DeepSeek official API docs and changelog · technical report · Artificial Analysis · 21st Century Business Herald · Kuai Technology · V2EX community discussion. Verify official pricing, benchmarks, and V4-Pro/Harness release status before production decisions.

FAQ

Frequently asked questions

Yes. Both V4-Pro and V4-Flash, including the July 31 official V4-Flash-0731 build, ship as open weights under the MIT license on Hugging Face, and can be used, fine-tuned, and redistributed commercially without additional permission.

Based on figures reported by 21st Century Business Herald, official V4-Flash pricing runs roughly 36x cheaper than Claude Opus 4.8 on cache-miss input, about 179x cheaper on cache-hit input, and about 89x cheaper on output, per million tokens. These are vendor list prices, not an independent audit.

There is no confirmed date. DeepSeek's changelog says only that the official V4-Pro release "will follow as soon as possible." Reports of an August 10–20 general-availability window come from unnamed sources in Chinese media and have not been confirmed by DeepSeek. For agent host deployment, see Mac Mini M4 rental pricing.

Partially. Widely-adopted third-party benchmarks like SWE-bench Verified carry more weight. But agent-specific scores (Terminal Bench 2.0, Toolathlon, etc.) were measured with DeepSeek's own unreleased Harness framework, and the company itself warns these numbers are highly sensitive to harness choice — wait for independent reproduction with Claude Code, Cursor, and other agent tools before treating them as general capability claims.

DeepSeek's first self-developed agent execution framework, positioned as an in-house alternative to Claude Code, for file editing, tool calls, and multi-step engineering work. It was named for the first time in the July 31, 2026 changelog and is not yet publicly available.

If you use OpenAI- or Anthropic-format APIs to access DeepSeek, no code changes are needed — deepseek-v4-flash automatically points to the new official build. If you still call the retired deepseek-chat / deepseek-reasoner aliases, switch immediately. More deployment details at our help center.