China's AI Knife Fight: DeepSeek's New Model Runs 100x Cheaper Than Anthropic's Flagship

The price of artificial intelligence just printed a new low. DeepSeek's V4-Flash, officially released Friday, costs roughly three cents to run through a standard battery of benchmark tests, according to San Francisco research firm Artificial Analysis. Moonshot AI's Kimi K3 costs 86 cents; OpenAI's GPT-5.6 Sol, $1.86; and Anthropic's Claude Fable 5, the industry's top-scoring model, $3.15. In realized terms, the Chinese model is more than 100 times cheaper to run than the American flagship, according to Reuters.

V4-Flash is the cheapest well-known model in the world, and by a wide margin. The figure is even more striking because of a detail buried in the firm's write-up: V4-Flash is unusually verbose. It consumes tokens heavily and still lands at three cents. The per-token price is doing all the work.

The same firm supplies the caveat. V4-Flash scores 50 out of 100 on the Artificial Analysis Intelligence Index, a composite of nine benchmarks spanning coding, reasoning, and workplace tasks. That ties Google's Gemini 3.6 Flash and puts it one point behind Meta's Muse Spark 1.1 and Zhipu's GLM-5.2. Moonshot's Kimi K3 scores 57, while Anthropic's Claude Opus 5 and Fable 5, along with OpenAI's GPT-5.6, score at least nine points higher.

That said - V4-Flash kicks massive ass at routine tasks, but the frontier models still own the heavy lifting. In difficult, multi-step agentic work, small reliability gaps at each step compound into enormous end-to-end differences. Essentially that's the current state of play for frontier pricing. But for the routine volume that makes up most production traffic - summarization, boilerplate code, and back-office automation - V4-Flash is where it's at.

Chinese Knife Fight

V4-Flash's 3-cent print is the fourth shot in an 18-day barrage: China's AI labs are cutting each other for domestic share, and the fallout is repricing the model market everywhere else. On July 16, Moonshot shipped Kimi K3, a 2.8-trillion-parameter model that promptly took the number-one slot on Arena's Frontend Code leaderboard from Fable 5 and GPT-5.6 Sol. On July 19, Alibaba rushed a preview of Qwen3.8-Max onto the stage at the World AI Conference in Shanghai. There was no pricing, no model card, and its claim of ranking "second only to Fable 5" rested on Alibaba's internal evaluations.

On July 27, Moonshot answered by open-sourcing K3's full weights, the largest open-weight release in history. DeepSeek shipped V4-Flash on July 31. Then on August 3 - Monday, the same morning the Reuters story ran - Alibaba took Qwen3.8-Max to general availability: 2.4 trillion total parameters, 95 billion active parameters, a one-million-token context window, and flat pricing of $2 per million input tokens and $6 per million output tokens, with no long-context surcharge.

Qwen's release promptly landed at No. 4 on the Frontend Code Arena with 1,668 points - one point behind Claude Opus 5 at high effort, eight behind Kimi K3, and ahead of both Fable 5 at 1,630 and GPT-5.6 Sol at 1,620. Of the five models Arena identifies on the cost-performance Pareto frontier, four are Chinese: Kimi K3, Qwen3.8-Max, GLM-5.2, and V4-Flash. The lone American entry, Opus 5, occupies the expensive tip, defended by 37 Elo points.

Qwen3.8-Max by @Alibaba_Qwen has reshaped the cost-performance Pareto frontier in Frontend Code Arena, with pricing of $2 per input MToken and $6 per output MToken.

Top models on the Pareto frontier:
- Claude-Opus-5
- Kimi-K3
- Qwen3.8-Max
- GLM-5.2
- DeepSeek-V4-Flash

Congrats… https://t.co/3S4tW1KmlI pic.twitter.com/CiWU7Hh4BD

— Arena.ai (@arena) August 3, 2026

Alibaba's own benchmark table is more candid. Qwen3.8-Max edges Fable 5 and Opus 4.8 on Terminal-Bench, 86.6 to 84.6, while trailing badly on hard repository engineering: it scores 67.7 on SWE-bench Pro against Fable 5's 80.0. The preference-judged coding moat is gone. The deep-engineering moat remains intact - for now.

All of this raises a more basic question: who is paying for three-cent inference?

Until this spring, DeepSeek had never taken outside money. Founder Liang Wenfeng bankrolled the company through his quant fund, High-Flyer. In late May, DeepSeek closed its first external round - more than 50 billion yuan, or roughly $7.4 billion, at a valuation above $50 billion - as first reported by The Information.

The round's structure is unusual. Commercial investors, reportedly led by Tencent and CATL, bought into a limited partnership controlled by Liang, with no voting rights and a five-year lockup. Exactly one party received direct equity and a vote: China's state-backed National AI Industry Investment Fund. Within weeks, DeepSeek was in talks for a follow-on round at roughly $71 billion, with proceeds earmarked for data centers and chips.

Add the 75% API discount the company made permanent earlier this year, and the arrangement begins to resemble industrial policy conducted through an API: state-privileged capital underwriting below-cost tokens to capture global share. It is working: in June, DeepSeek accounted for nearly 23% of the tens of trillions of tokens flowing through Vercel's enterprise AI gateway, compared with Anthropic's 32%.

Meanwhile...

As we (and now Wall Street) have been noting, the Token Expenditure Index - a usage-weighted average of what the market pays per million tokens, blended across frontier APIs and open-weight platforms - peaked above 2.0 in May after nearly doubling from its December launch, and is now slip sliding lower. 

Strategist Andreas Steno Larsen called it the one everyone should be watching, warning that sustained weakness in token pricing would end the memory, hardware, and data-center trades for this cycle. The index last printed 1.3394, roughly a third below its May high. Bloomberg flagged the rollover in early July as evidence that AI vendors were losing pricing power with increasingly cost-sensitive customers; Silicon Data's own commentary interpreted it as usage drifting back toward open-weight models.

Also relevant - the pushback to data centers amid a capex boom running north of $700 billion. As we reported last month, from the nationwide July 18 protests organized by Tea Party veteran Amy Kremer's Humans First to the widening fracture inside the Republican coalition over land, water, and power - domestic politics has entered the chat, something Beijing doesn't have to deal with - so now they've got a three-cent benchmark financed on terms no Western lab can match. Through open-weight releases, it is also portable onto American silicon, where US inference providers will happily serve Chinese models at commodity margins. Export controls cannot contain a set of weights on the torrent.

The model to watch is V4-Pro, the heavier system DeepSeek has confirmed without naming a release date. Flash at three cents pressures the budget tiers at OpenAI and Google. If Pro lands anywhere near frontier scores at DeepSeek prices, the last 37 Elo points - and the frontier premium that OpenAI and Anthropic both charge to underwrite the buildout - will be directly in play.