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

If you are evaluating agent API options in the summer 2026 frontier-model wave and need to decide whether DeepSeek V4-Flash-0731 belongs in production pipelines, this article walks through the timeline, core specs, head-to-head comparison, and caveats. On July 31, DeepSeek promoted V4-Flash to an official build with the same 284B-parameter architecture as April's preview — only post-training changed — beating its own larger V4-Pro preview on agent benchmarks at roughly 1/36 to 1/179 of Claude Opus 4.8's price. The flagship V4-Pro and in-house Harness framework remain unreleased.

What actually shipped on July 31 — and what did not

• April 24, 2026: DeepSeek-V4 preview launches with two open-weight MIT MoE models — V4-Pro (1.6T total / 49B active) and V4-Flash (284B total / 13B active) — both with a 1M-token context window.
• July 24, 2026: Legacy aliases deepseek-chat and deepseek-reasoner are retired; all traffic routes to the V4 family.
• 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.
• July 31, 2026: deepseek-v4-flash is promoted to an official public API beta (build tag "0731"). Same architecture, same parameter count — gains came entirely from a fresh post-training pass. Open weights land on Hugging Face under MIT the same day. The changelog also names "DeepSeek Harness," an in-house agent framework, for the first time. Crucially, this update is API-only — the consumer app and web chat are untouched.
• 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.

The numbers at a glance

All pricing figures above are vendor-published rates. DeepSeek has also announced a future 2x peak-hour surcharge (9am–12pm and 2pm–6pm Beijing time) with no confirmed effective date yet. Prices and specs may change — verify on official pages before committing.

How DeepSeek squeezed more out of the same model

The architecture did not 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 — cutting against the industry's default assumption that bigger equals better.

Hybrid attention, hyper-connections, and a new optimizer. DeepSeek's technical report ("DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence") describes three architectural changes:

1. Hybrid attention combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), marketed as "DSA" (DeepSeek Sparse Attention), aimed at cutting compute and memory at long context lengths;
2. Manifold-Constrained Hyper-Connections (mHC), an enhancement over standard residual connections;
3. The Muon optimizer 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 — vendor-reported efficiency numbers without independent third-party reproduction yet.

Harness: DeepSeek's first in-house agent framework. July 31 marked the first official mention of DeepSeek Harness, positioned as an in-house alternative to Claude Code. Every agent benchmark DeepSeek published (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."

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

Intelligence Index and per-task cost figures come from Artificial Analysis, an independent benchmarking firm. DeepSeek's own agent benchmarks use a different methodology and are listed separately. V4-Flash's per-task cost is roughly 1/29th of Kimi K3, 1/62nd of GPT-5.6 Sol, and 1/105th of Claude Fable 5 — DeepSeek is not competing for the top of the leaderboard; it is optimizing for "good enough intelligence at a price nobody else can match." For more pricing context, see our GPT-5.6 price cuts and API pricing comparison and Qwen3.8-Max release breakdown.

Hard data and sources you can cite

• Post-training drove the jump: V4-Flash-0731 scored 82.7 on Terminal Bench 2.0 versus V4-Pro preview's 67.9 — same 284B/13B architecture (DeepSeek changelog, Harness minimal mode).
• Long-context efficiency: At 1M tokens, V4-Pro needs 27% of V3.2's per-token FLOPs and 10% of KV cache (DeepSeek technical report).
• Independent cost comparison: Artificial Analysis average per-task cost — V4-Flash $0.03 vs Kimi K3 $0.86 vs Claude Fable 5 $3.15.
• OpenRouter usage: The V4-Flash preview reportedly topped the most-used model ranking for seven consecutive weeks — a signal that large-scale agent workloads prioritize unit economics over peak intelligence.

Verifiable official and third-party sources — confirm latest figures before publishing:

DeepSeek official API documentation and changelog

Hugging Face DeepSeek model cards (DeepSeek-V4-Pro, DeepSeek-V4-Flash)

Artificial Analysis independent model benchmarks

The catch: why you should not trust the benchmarks blindly

1. Headline agent scores are harness-dependent and self-reported. Terminal Bench 2.0's 82.7 was measured using DeepSeek's own unreleased Harness in minimal mode. Until third parties reproduce with other agent harnesses (Claude Code, Cursor, etc.), treat these as "vendor plus specific framework" results.
2. Real-world usability complaints exist. Chinese financial outlet 21st Century Business Herald, citing overseas developer community feedback, reported low input cache-hit rates and occasional safety-classifier timeouts on the official V4-Flash build.
3. V4-Pro and Harness release dates are unconfirmed. An August 10–20 GA window from unnamed sources is a rumor — DeepSeek's changelog only says the official V4-Pro release "will follow as soon as possible."
4. Funding and IPO reports are not independently confirmed. Outlets report a roughly $7.4 billion funding round valuing DeepSeek at approximately $48.7 billion — these trace back to unnamed sources in financial media, not regulatory filings or DeepSeek statements.

Why Chinese developers call this the "kill line"

Before V4-Flash-0731 shipped, Chinese AI forums had taken to mocking DeepSeek founder Liang Wenfeng with the nickname "Liang Baikai" — a pun roughly meaning "Liang Empty Promise," a jab at V4-Pro's slipped mid-July target. Once the official Flash build outperformed expectations, the same communities flipped back to calling him "Liang Sheng" ("Liang the Sage").

More substantively, there is a term circulating in Chinese developer circles: "斩杀线" (zhǎn shā xiàn), literally "kill line." The idea is that DeepSeek's combination of good-enough performance plus rock-bottom price sets an effective bar: competitors whose models do not clearly beat DeepSeek on capability, and cannot undercut it on price either, risk losing relevance. That framing helps explain moves like OpenAI reportedly cutting prices by 80% on its low-cost GPT-5.6 Luna tier around the same period. On July 31, when V4-Flash went official, chipmakers 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.

FAQ

DeepSeek V4-Flash-0731 pushes the "good-enough intelligence at rock-bottom price" agent API playbook to a new level — but Harness is not public yet and V4-Pro has no confirmed GA date. Until independent reproduction lands, if you need to validate local agent workflows on real Apple Silicon hardware and compare harness behavior side by side, a dedicated physical host you can reset on demand is often more predictable than a shared GPU pool. VMSPIN's day-rental cloud Mac mini ships with MDM auto-provisioning and SSH + VNC dual access — suited for short benchmark runs and multi-toolchain comparisons. Pair it with our Mac mini M4 rental guide and pricing page if you are weighing API costs against local deployment.