Is DeepSeek's new model really 100x cheaper than Claude? Inside the V4-Flash benchmarks

Why you shouldn't trust the benchmarks blindly

The July 31 numbers are real vendor claims — but several friction points matter before you rewrite your agent stack:

• Headline agent scores are harness-dependent: Terminal Bench 2.0 at 82.7 (versus V4-Pro-preview's 67.9) was measured using DeepSeek's own unreleased Harness in minimal mode at max settings. DeepSeek's changelog explicitly warns agent scores are "extremely sensitive to harness choice."
• API-only update: The July 31 build is an API public beta. Consumer app and web chat were untouched — developer API experience and end-user product lag are not the same thing.
• Usability complaints exist: Chinese financial outlet 21st Century Business Herald, citing overseas developer feedback, reported low input cache-hit rates and occasional safety-classifier timeouts on the official build.
• Flagship and framework still pending: Official V4-Pro and Harness have no confirmed date. An August 10-20 GA window circulating in Chinese media traces to unnamed sources — DeepSeek's own text is only "as soon as possible."
• Long-horizon agent eval needs isolation: Multi-step tool chains and overnight coding runs need a wipeable dev machine — shared CI sandboxes with stale SDKs cannot reproduce vendor showcase cases.

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

• April 24, 2026: DeepSeek-V4 preview launches — V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), both 1M-token context, MIT open weights.
• July 24, 2026: Legacy aliases deepseek-chat and deepseek-reasoner retired; all traffic routes to the V4 family.
• July 27, 2026: Moonshot AI ships Kimi K3 open weights (2.8T), raising competitive pressure days before DeepSeek's update.
• July 31, 2026: deepseek-v4-flash promoted to official public API beta (build tag "0731"). Same architecture — gains from fresh post-training only. Open weights on Hugging Face same day. Changelog names "DeepSeek Harness" for the first time.
• As of August 5, 2026: Official V4-Pro release unconfirmed. Some Chinese outlets report internal testing began the week of July 28 — not confirmed by DeepSeek.

The numbers at a glance

All pricing is vendor-published. DeepSeek has announced a future 2x peak-hour surcharge (9am–12pm and 2pm–6pm Beijing time) with no confirmed effective date yet.

How DeepSeek squeezed more out of the same model

V4-Flash-0731 is identical in size and structure to April's preview. The entire agent benchmark jump came from re-running post-training, not scaling up — a 284B/13B model now beats a 1.6T/49B sibling on multiple agentic tasks.

• Hybrid attention (DSA): Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA), cutting compute and memory at long context.
• Manifold-Constrained Hyper-Connections (mHC): Enhancement over standard residual connections.
• Muon optimizer: Faster convergence and training stability versus traditional optimizers.
• Efficiency claims (vendor-reported): At 1M-token context, V4-Pro needs only 27% of V3.2's per-token FLOPs and 10% of KV cache — no independent third-party reproduction seen yet.
• DeepSeek Harness: First official mention July 31 — in-house agent framework positioned as DeepSeek's answer to Claude Code. Every published agent benchmark used Harness's unreleased "minimal mode."

DeepSeek V4-Flash vs Kimi K3 vs GLM-5.2 vs Qwen3.8-Max

Intelligence Index and per-task cost come from Artificial Analysis (independent). DeepSeek's own agent benchmarks use a different methodology and are listed separately. The tension: on AA's index, V4-Flash trails Kimi K3 and GLM-5.2 — but per-task cost is roughly 1/29th of Kimi K3, 1/62nd of GPT-5.6 Sol, and 1/105th of Claude Fable 5. That explains why the preview reportedly topped OpenRouter's most-used ranking for seven weeks.

Six steps to migrate and evaluate V4-Flash-0731

Until Harness ships and V4-Pro goes official, the pragmatic path is API-first with your own workload samples.

1. Retire deprecated model names: deepseek-chat and deepseek-reasoner were retired July 24. Route to deepseek-v4-flash or deepseek-v4-pro (preview).
2. Get an API key and verify endpoints: Confirm OpenAI-compatible and Anthropic-compatible routes cover your agent framework.

from openai import OpenAI
client = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://api.deepseek.com",
)
resp = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Refactor this function without changing behavior: ..."}],
    temperature=1.0,
    top_p=0.95,
)
print(resp.choices[0].message.content)

3. A/B on real workload samples: Blind-score 20–50 representative tasks against Kimi K3 and Qwen3.8-Max APIs — not just DeepSeek's launch table.
4. Profile cache hit rates: Community reports cite low cache-hit rates on the official build — test cache strategy on long-context agent runs and watch the bill.
5. Separate Harness scores from general capability: Terminal Bench 82.7 was measured with unreleased Harness minimal mode. Wait for reproduction with Claude Code, Cursor, or your production harness.
6. Run long-horizon agents on an isolated Mac: Multi-step tool chains and overnight coding on a wipeable machine — shared CI cannot keep a clean control environment.

The "kill line" and what Chinese developers are saying

Before V4-Flash-0731 shipped, Chinese AI forums mocked founder Liang Wenfeng as "Liang Baikai" — a pun roughly meaning "Liang Empty Promise" — after V4-Pro's mid-July target slipped. Once the official Flash build outperformed expectations, communities flipped back to "Liang Sheng" ("Liang the Sage").

More substantively, Chinese developer circles use "斩杀线" (zhǎn shā xiàn) — literally "kill line." DeepSeek's "good-enough performance plus rock-bottom price" sets an effective bar: competitors that don't clearly beat DeepSeek on capability and can't undercut on price risk losing relevance. That framing helps explain moves like OpenAI reportedly cutting GPT-5.6 Luna prices by 80% around the same window.

On July 31, Nvidia, Broadcom, and AMD saw no significant stock movement — a contrast to early 2025 when DeepSeek-R1's efficiency claims triggered a global AI-chip selloff. Funding and IPO reports (~$7.4B round, ~$48.7B valuation) trace to unnamed financial media sources, not regulatory filings — presented here as background, not verified fact.

Hard numbers and sources

• Scale: V4-Flash 284B total / 13B active, 1M-token context, MIT open weights.
• Pricing: Input $0.14 (cache miss) / $0.0028 (hit), output $0.28 per million tokens.
• Terminal Bench 2.0 (vendor + Harness): V4-Flash-0731 82.7 vs V4-Pro preview 67.9.
• Artificial Analysis: Intelligence Index 50, $0.03 average cost per task.
• Efficiency (vendor-reported): At 1M context, V4-Pro FLOPs 27% of V3.2, KV cache 10% of V3.2.

Verify latest pricing, benchmarks, and V4-Pro/Harness status before publishing — figures current as of August 5, 2026.

DeepSeek official documentation and model cards:

DeepSeek API documentation and changelog

Hugging Face — DeepSeek-V4-Flash model card

Independent benchmarking:

Artificial Analysis — Intelligence Index and task costs

Chinese media coverage:

21st Century Business Herald — DeepSeek update coverage

Frequently asked questions

V4-Flash-0731 delivers "good-enough intelligence at a price nobody else can match" — but Harness is unreleased, flagship Pro is still pending, and cache and classifier feedback are mixed. API smoke tests, Kimi K3 blind comparisons, and long-horizon agent runs need a wipeable, root-capable Mac — shared sandboxes cannot hold a clean control. KVMFLUX cloud Mac mini M4 rentals start daily with real Apple Silicon and minute-level SSH access: wipe and rerun when model policy shifts overnight, wire V4 into your agent pipeline before Harness goes public.