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

About 16 min read · MACCOME · Last updated: August 5, 2026

Who this is for: AI developers, agent engineers, and tech leads choosing APIs under cost pressure. On July 31, 2026, DeepSeek promoted V4-Flash to official build 0731: same parameter count, post-training only—agent scores beat the larger V4-Pro preview, and pricing runs at roughly 1/36 to 1/179 of Claude Opus 4.8 depending on cache hits. You get: a full timeline, a six-way pricing matrix, Harness framework context, Artificial Analysis cross-check, benchmark caveats, and a six-step production runbook. Structure: pain points, timeline, pricing table, architecture, competitor matrix, controversy, runbook, closing bridge. V4 GA and peak/off-peak billing: V4 full release guide; Kimi K3 context: Kimi K3 open-weight release.

bolt

TL;DR — 30-second verdict

  • Not a new model: 284B total / 13B active—the same as the April preview. Gains come entirely from post-training, not scale.
  • Price gap: input $0.14/M (cache miss), output $0.28/M—roughly 1/36 to 1/179 of Opus 4.8 depending on cache hit rate.
  • Harness debut: DeepSeek's in-house agent framework targets Claude Code, but it is not public yet. Official agent scores all run in Harness minimal mode.
  • V4-Pro still not GA: as of August 5, only Flash-0731 is official (API-only). Pro official and Harness are marked coming soon.
  • Score with caution: Terminal Bench 2.0's 82.7 is vendor-reported on a specific framework—not a direct read on Claude Code or Cursor behavior.

Six blind spots after Flash GA—and why they hit your bill

After the July 31 official drop, most teams stall on six decision gaps. Each one maps directly to API spend or agent quality:

  1. Harness-dependent scores: Terminal Bench 2.0's 82.7 runs on unpublished Harness minimal mode. It is not a drop-in proxy for Claude Code or Cursor.
  2. V4-Pro official still missing: the 1.6T flagship has no GA date. Media rumors of an August 10–20 window are unsourced; the changelog only says coming soon.
  3. API-only rollout: build 0731 is API public beta only. Consumer app and web still run older builds—mixed endpoints mean inconsistent behavior.
  4. Weak prompt cache hits: overseas developers report low input cache hit rates. Agent loops can cost far more than list price.
  5. Peak pricing not live yet: DeepSeek warned of 2x surcharges during Beijing 9–12 and 14–18, but no effective date is published as of this writing.
  6. Funding and IPO rumors unverified: finance press cites ~$7.4B fundraising and IPO prep from unnamed sources—no regulatory filing or official statement confirms it.

Timeline: April preview to July Flash official

DateEvent
2026-04-24V4 preview ships open source: V4-Pro (1.6T / 49B active) and V4-Flash (284B / 13B active), both 1M context, MIT license
2026-07-24Legacy endpoints deepseek-chat and deepseek-reasoner retired; all traffic moves to V4 naming
2026-07-27Moonshot AI releases Kimi K3 (2.8T) open weights on Hugging Face, raising competitive pressure
2026-07-31V4-Flash-0731 official enters API public beta; open weights sync; changelog names in-house agent framework Harness for the first time
2026-08-02Alibaba Qwen3.8-Max API GA (weights pending), tightening the Chinese open-model race
2026-08-05 (publish date)V4-Pro official still absent; Harness not public; app and web not on 0731

Core pricing matrix: Flash official vs six Chinese open-model players

All figures are vendor-published as of August 5, 2026. GLM-5.2 fields marked unverified were not confirmed for this article.

ModelStatusTotal / active paramsContextInput (miss/hit, $/M)Output ($/M)License
V4-Flash-0731Official GA284B / 13B1M$0.14 / $0.0028$0.28MIT
V4-ProPreview1.6T / 49B1M$0.435 / $0.003625$0.87MIT
Kimi K3Open weights2.8T / ~104B (community est.)~1.05M$3.00 / $0.30$15.00Modified MIT
GLM-5.2Open source~744B / ~40B1MunverifiedunverifiedMIT
Qwen3.8-MaxAPI GA2.4T / 95B1M$2.00 / ~$0.17–0.25$6.00open-source promised
GPT-5.6 SolClosedundisclosedproprietary
Claude Fable 5Closedundisclosed1M~$50proprietary

Same architecture, stronger post-training: CSA + HCA + mHC + Muon

The headline most teams miss: V4-Flash-0731 uses the identical parameter count and model structure as the April preview. Performance gains come from re-run post-training only. The 284B / 13B Flash variant now beats the much larger V4-Pro preview on multiple agent benchmarks—a signal that late-2026 post-training quality can rival raw parameter scale.

Per the technical report DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, three architectural changes define the V4 series:

  • Hybrid attention (CSA + HCA): marketed as DSA sparse attention; cuts compute and memory at long context. At 1M tokens, V4-Pro inference FLOPs are 27% of V3.2 and KV Cache is 10% (vendor-reported; no independent third-party rerun yet).
  • Manifold-constrained hyper-connections (mHC): stabilizes signal flow through deep residual stacks.
  • Muon optimizer: replaces AdamW for faster, more stable convergence.

Harness: DeepSeek's agent execution layer vs Claude Code

The July 31 changelog names DeepSeek Harness for the first time—an in-house agent runtime for file I/O, tool calls, shell commands, and end-to-end engineering tasks, positioned against Claude Code. Before this, DeepSeek models relied on Claude Code, OpenCode, and other third-party agent stacks.

Every official agent score (Terminal Bench 2.0, Toolathlon, and similar) runs in Harness minimal mode at max tier, top_p=0.95, temperature=1.0. DeepSeek's own changelog warns: scores are highly sensitive to harness choice and do not equal a raw model capability jump.

Artificial Analysis cross-check: not top intelligence, lowest per-task cost

ModelLabIntelligence Index (third party)Avg cost per task (third party)
V4-Flash-0731DeepSeek50$0.03
Kimi K3Moonshot AI57$0.86
GLM-5.2Zhipu / Z.ai~1 pt above Flashunverified
GPT-5.6 SolOpenAI9+ pts above Flash$1.86
Claude Fable 5Anthropic9+ pts above Flash$3.15

Source: Artificial Analysis independent evaluation, cited via finance media. Methodology differs from DeepSeek's Harness-based agent scores—do not merge the two datasets.

The contrast is deliberate: V4-Flash is not the highest intelligence score (it trails Kimi K3 and GLM-5.2), but per-task cost is roughly 1/29 of Kimi K3, 1/62 of GPT-5.6 Sol, and 1/105 of Claude Fable 5. DeepSeek is not chasing benchmark crowns—it is selling good-enough intelligence at extreme price, which helps explain seven straight weeks atop OpenRouter call volume.

Controversy: strong scores, real caveats

  • Agent scores need Harness: Terminal Bench 2.0 at 82.7 (vs V4-Pro preview 67.9) is vendor-reported on an unpublished framework. Treat as directional until Claude Code or Cursor reruns land.
  • Operational friction reported: per 21st Century Business Herald citing overseas developers, the official build shows low prompt cache hit rates and occasional safety-classifier timeouts.
  • V4-Pro / Harness dates unconfirmed: Chinese media's August 10–20 GA window is unsourced. Official language remains coming soon.
  • Funding figures need verification: ~$7.4B raise and ~$48.7B valuation numbers circulate via press reports without direct regulatory confirmation.

Market reaction: price pressure and the "kill line"

While V4-Pro slipped, Chinese AI circles nicknamed founder Liang Wenfeng "Liang Baikai" (a pun on waiting in vain). Flash-0731 flipped sentiment back toward praise. Community talk of a kill line describes DeepSeek setting a floor: rivals need clear performance wins and lower price—or fade from relevance. That context fits OpenAI's July GPT-5.6 Luna 80% price cut and related moves (see OpenAI pricing breakdown).

On launch day, NVIDIA, Broadcom, and AMD saw no major move—unlike the January 2025 DeepSeek-R1 shock. Markets appear accustomed to DeepSeek-style efficiency narratives.

Six-step runbook: evaluate, route, and ship to production

  1. Point API calls correctly: deepseek-v4-flash resolves to 0731 official. Retire deprecated deepseek-chat immediately—migration details in the V4 GA migration guide.
  2. Stage quality against preview: run a real prompt subset on 0731 vs preview. Focus on tool-call formatting and long-context stability.
  3. Route Flash vs Pro by tier: batch and router workloads on Flash ($0.28/M output); keep hard reasoning on V4-Pro preview until official Pro GA.
  4. Monitor cache hit rate: fixed system prompts in agent loops should use prompt cache. Low hits inflate spend well above list price.
  5. Schedule around peak surcharges: batch non-real-time jobs outside Beijing 9–12 and 14–18 once 2x peak pricing activates.
  6. Decide local vs cloud: a 128GB Mac can run Flash q2 via ds4; production agents need 24/7 nodes—see the ds4 local vs cloud rental decision guide.
python
from openai import OpenAI

client = OpenAI(
    api_key="your-deepseek-api-key",
    base_url="https://api.deepseek.com"
)

# deepseek-v4-flash resolves to 0731 official
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Analyze this agent log..."}],
)

Three hard numbers for your memo

  • 82.7 vs 67.9 — Terminal Bench 2.0: V4-Flash-0731 vendor score beats V4-Pro preview (Harness minimal mode; pending independent rerun)
  • 27% / 10% — at 1M context, V4-Pro inference FLOPs and KV Cache vs V3.2 (DeepSeek technical report, vendor-reported)
  • 1/29 to 1/105 — V4-Flash per-task cost vs Kimi K3 / GPT-5.6 Sol / Claude Fable 5 (Artificial Analysis third-party data)

Close: good-enough intelligence at extreme price—production agents still need stable nodes

V4-Flash-0731 is not trying to beat Claude Fable 5 or GPT-5.6 Sol on raw intelligence. It offers an open, MIT-licensed option at roughly 1/36 to 1/179 of closed flagship pricing for bulk agents and high-frequency calls. For budget-sensitive teams that need private deploy or heavy API volume, Flash official is among the strongest value picks in August 2026.

Running DeepSeek V4 agent workflows in production—OpenClaw Gateway, future Harness hooks, or multi-model routing—still hits three structural limits on a laptop:

  • Lid-close sleep: Wi-Fi drops and sleep kill long agent sessions; spent tokens are not refunded.
  • Cache and peak scheduling: local cron cannot reliably batch off-peak when sleep policies interfere.
  • No 24/7 router: Flash / Pro / Fable 5 tiering needs an always-on gateway; a notebook is a poor dispatch node.

For a more stable environment suited to AI agent automation, MACCOME Mac cloud hosts provide real macOS, SSH handoff, and isolated environments so agents run 24/7 on dedicated hardware. Plans and pricing: Mac mini cloud rental rates.

Sources: DeepSeek official API docs and changelog, technical report, Hugging Face model cards, Artificial Analysis (via finance media), 21st Century Business Herald, V2EX community threads. All data as of August 5, 2026—verify latest pricing and V4-Pro / Harness status before production cutover.

FAQ

What is the biggest difference between DeepSeek V4 and V3.2?

Native 1M-token context. At that window, V4-Pro inference FLOPs are 27% of V3.2 and KV Cache is 10% (vendor-reported). The V4 series adds agent-focused tuning for Claude Code, OpenCode, and similar tools.

Should I use V4 Flash or V4 Pro day to day?

Default to V4-Flash-0731 official for chat, batch jobs, and agent pipelines—it beats V4-Pro preview on agent scores at far lower cost. Use V4-Pro preview for deeper world knowledge or hard reasoning until official Pro GA ships.

Is V4 Pro official available yet?

Not as of August 5, 2026. Only V4-Flash-0731 is GA (API-only); app and web are not on 0731. V4-Pro official and Harness are marked coming soon—August 10–20 rumors are unverified.

Can I trust DeepSeek's published benchmark scores?

Tier your trust. SWE-bench Verified and similar third-party benchmarks rank higher. Terminal Bench 2.0 agent scores run on unpublished Harness minimal mode—DeepSeek warns harness choice dominates results. Wait for independent Claude Code or Cursor reruns.

How do I run DeepSeek V4 agents 24/7 in production?

Deploy OpenClaw Gateway or multi-model routing on a dedicated Mac cloud host. See MACCOME Mac mini rental rates for node configs and pricing.