Skip to main content
Four Chinese AI model orbs in a cascade formation Four luminous blue orbs representing DeepSeek V4, Kimi K2.6, GLM-5.1, and MiniMax M2.7 arranged in a staggered vertical formation. A cascading flow of light connects them, flowing downward like a flood. Space Mono annotations show model names, prices, and the 12-day release window. DEEPSEEK V4 $0.28/M TOKENS KIMI K2.6 MOONSHOT AI GLM-5.1 Z.AI · MIT LICENSE MINIMAX M2.7 OPEN WEIGHTS 12 DAYS   ·   4 MODELS   ·   APR 7-24, 2026

AI & Discovery  ·  Geopolitics

The Flood


In twelve days in April 2026, four Chinese AI laboratories released open-weights frontier models at prices that make the Western equivalent look, in at least one analyst's phrase, like a different economic universe.

Between April 7 and April 24, 2026, four Chinese AI laboratories released frontier-class open-weights models. Twelve days. Four models. MIT licences. Price tags that make the equivalent Western models look obscene. This was not a coordinated campaign - but it had the effect of one.

Section 01

Twelve Days


The sequence began on April 7th, the same day Anthropic formally announced Mythos. Z.ai released GLM-5.1. A week later, MiniMax published M2.7. On April 18th, Moonshot dropped Kimi K2.6. DeepSeek, which had effectively restarted the Chinese open-source AI conversation in early 2025, completed the sequence with V4 Pro on April 21st and V4 Flash on April 24th.

No single lab coordinated the window. The timing reflects the compressed development cycles now operating across the Chinese AI ecosystem, not a strategic communications plan. But the effect was the same: a cascade that arrived faster than Western engineers could absorb it, each model landing before the previous one had been properly benchmarked and integrated into workflows.

The models share three properties that make the release window significant. First, they are open-weights: the model weights are publicly available, freely downloadable, and - in the case of DeepSeek, Kimi K2.6, and GLM-5.1 - MIT-licensed, meaning any organization can deploy, modify, and build on them without restrictions. Second, they are frontier-competitive: not at the Mythos tier, but within striking range of GPT-5.5 on most standard benchmarks. Third, they are priced for mass adoption at a scale that fundamentally changes the unit economics of AI deployment.

Open-weights vs. open-source

Open-weights means the trained model parameters are public. Open-source, strictly speaking, means the training code, data, and process are also disclosed. Most "open" Chinese models are open-weights only - you can run the model, but you cannot reproduce the training run. MIT-licensed open-weights is nevertheless a substantial commitment: it permits commercial use, modification, and redistribution without restriction, including building closed products on top of the base model.

Section 02

What They Build


12 Days from first to fourth release — April 7-24, 2026
87 BenchLM score — DeepSeek V4 Pro (Chinese leaderboard #1)
61% Share of global OpenRouter token consumption — Chinese models, Feb 2026

On BenchLM, the most widely used aggregated capability benchmark in 2026, DeepSeek V4 Pro scores 87. Kimi K2.6 scores 84. GLM-5.1 scores 83. For reference, GPT-4o - the model that defined the frontier in mid-2024 - scores in the low 60s. These are not catch-up models. They are models that would have been considered frontier-class a year ago, now released as open weights.

Model comparison: Chinese open-weights frontier vs. Western closed models, May 2026
Model Lab BenchLM License Price / 1M out tokens Weights
Chinese Open-Weights — April 2026 Release Window
DeepSeek V4 Pro DeepSeek 87 MIT $1.10 Public
DeepSeek V4 Flash DeepSeek 81 MIT $0.28 Public
Kimi K2.6 Moonshot AI 84 MIT $0.70 Public
GLM-5.1 Z.ai 83 MIT $0.50 Public
MiniMax M2.7 MiniMax 80 Proprietary $0.40 API only
Western Closed Models — for comparison
GPT-5.5 Instant OpenAI 82 Proprietary ~$15.00 API only
Claude Opus 4.7 Anthropic 88 Proprietary ~$30.00 API only
Mythos Preview Anthropic ~100+ Restricted Not available Glasswing only
Figure 1 — Model comparison. Prices are approximate API costs per million output tokens as of May 2026. BenchLM scores sourced from benchlm.ai.

The capability comparison requires a clear-eyed reading. These models are not at Mythos's level. The gap between DeepSeek V4 Pro (87 BenchLM) and Mythos (estimated well above 100 on equivalent measures) is substantial. On long-horizon autonomous tasks, Mythos's METR time horizon of 16 hours is not something any of these models approach.

What they are is highly competitive with every Western model that is actually publicly available. DeepSeek V4 Pro trails Claude Opus 4.7 by one BenchLM point and exceeds GPT-5.5 Instant by five. At 1/27th of the price of Opus 4.7.

Section 03

The Price Weapon


The numbers need to be held next to each other. DeepSeek V4-Flash costs $0.28 per million output tokens. Claude Opus 4.7 costs approximately $30.00 per million output tokens. That is not a price difference. It is a 107-fold gap. For high-volume production workloads generating tens of millions of tokens per day, the choice between Western and Chinese models is, economically, not a choice at all.

Price per 1M output tokens — selected models, May 2026
Claude Opus 4.7
$30.00
GPT-5.5 Instant
$15.00
DeepSeek V4 Pro
$1.10
Kimi K2.6
$0.70
DeepSeek V4 Flash
$0.28

* Cache-hit pricing (DeepSeek, Kimi) reduces effective input cost to $0.03-0.07/M for agent workloads with stable system prompts.

DeepSeek pioneered cache-hit pricing: if your application reuses the same system prompt across calls - as virtually all production AI applications do - the input cost for cache-hit tokens drops to $0.07 per million. Kimi K2.6 followed with similar pricing. For agentic workloads where the system prompt is long and stable, effective input cost can reach $0.03 per million tokens. Not cheaper than Western models. Categorically different.

By February 2026, the market had already shifted to reflect this. On OpenRouter - the world's largest AI model API aggregation platform, used by developers who have made a direct comparison across providers - Chinese models accounted for 61% of total token consumption among the top ten models globally. Four of the five most-used models on the platform were Chinese. China's daily AI token usage surpassed 140 trillion in March 2026, up from 100 billion at the start of 2024. That is a more than 1,000-fold increase in two years.

"We're seeing a rapid commoditization of frontier-tier AI capabilities - what required GPT-4 class compute eighteen months ago is now available at inference prices that make enterprise deployment economics completely different."
State of AI: May 2026 — Air Street Press

Section 04

The Geopolitical Layer


Stanford HAI's analysis of China's open-source AI strategy reaches an uncomfortable conclusion: the commitment to open weights is strategic, not ideological. China's open-source AI stance is not analogous to the open-source software movement, which emerged from a philosophy of shared knowledge and user freedom. It is a market capture strategy. If the global developer ecosystem builds on Chinese model infrastructure, Chinese labs control the interfaces through which the world's AI applications are shaped - and the dependencies that come with that control compound over time.

The parallel to the open-source software platform wars is deliberate and exact. Linux won the server. Android won mobile. Not because proprietary alternatives were weaker, but because open platforms captured developer adoption early and made themselves the default substrate. DeepSeek, Kimi, and GLM are attempting the same playbook for AI inference.

The adoption data suggests it is working. Southeast Asia has moved to what analysts call a "dual-stack" approach: Western hyperscaler infrastructure for compute (AWS, Google Cloud) combined with Chinese open-weight models for application-layer AI, where cost efficiency dominates the decision. The choice in regional boardrooms is not geopolitical - it is economic. Chinese models are where the unit economics work.

The neutrality problem

DeepSeek and other Chinese frontier models decline entirely to answer questions flagged as sensitive under Chinese state standards: Taiwan's political status, the 1989 Tiananmen Square events, the Xinjiang detention system, criticism of the Chinese Communist Party. This is not a minor content policy divergence. When the applications through which billions of people access information are powered by models that systematically decline these questions, the question of what those models will and will not say becomes a governance issue, not a technical one. The Centre for International Governance Innovation notes that "AI neutrality" - the idea that a model can be deployed as a globally neutral information tool - is incompatible with models trained under national content restrictions.

The picture that emerges from the April 2026 release window is one of deliberate, accelerating divergence. At the frontier, the West has Mythos - the most capable AI ever built, locked away and inaccessible. Below the frontier, Chinese labs have flooded the deployable tier with open-weights models at prices that make the Western equivalents economically indefensible for most applications. The two strategies are mirror images: Anthropic and OpenAI competing on raw capability at the top, Chinese labs commoditizing everything below.

For developers building production systems today, the practical implication is immediate: Chinese models have won the cost argument for the middle tier of AI workloads. For governments thinking about AI infrastructure dependency, the April 2026 release window is the clearest illustration yet of what strategic open-source looks like at scale. And for anyone watching the long-term trajectory - where does capability lead when the price approaches zero? - the answer is somewhere none of our existing governance frameworks were designed to handle.

Ko-fi Buy me a coffee
Scroll to Top