China's open LLMs vs US cloud LLMs in 2026: a narrowing gap
Moonshot's Kimi K3 pulled the best open (Chinese) model to within about three points of the US closed frontier in July 2026. But the tidy 'China is cheap and open, the US is pricey and closed' split is breaking down.
목차
As of July 2026, China’s open-weight LLMs still trail the US cloud frontier — but the distance is months, not generations, and it just shrank again. Moonshot AI’s Kimi K3, released July 16, took the top open slot from GLM-5.2 and pulled the composite-intelligence gap from about nine points to about three (Tom’s Hardware, Simon Willison). By time, Epoch AI’s composite index puts the best open model about four months behind the best closed one (Epoch AI), and the UK’s AI Safety Institute has the cyber-capability gap narrowing from six-to-ten months to four-to-seven (UK AISI). Capability is converging — but price no longer splits as cleanly by side.
- ~3 pts
- top open vs top closed60 vs 57, Artificial Analysis index (K3 included)
- 2.8T
- Kimi K3 total parameterslargest open-weight model claimed (~50B active)
- $15 vs $6
- open K3 vs US Grok output pricecamp no longer equals price
- 1M tokens
- now the standard flagship contexton both sides
Numbers shift with who’s measuring and when. The figures below are from public sources as of July 23, 2026, with self-reported scores marked apart from independent ones.
Who’s at the frontier right now
The US side is all closed — no published weights, API only.
| Model | Maker | Released | In/Out ($/1M) | Context | Headline score (measured by) |
|---|---|---|---|---|---|
| Claude Fable 5 | Anthropic | 2026-06 | 10 / 50 | 1M | AA index 60 (independent, #1) |
| Claude Opus 4.8 | Anthropic | 2026-05 | 5 / 25 | 1M | AA index ~57 (independent) |
| GPT-5.6 Sol | OpenAI | 2026-07 | 5 / 30 | undisclosed | Terminal-Bench 2.1 88.8% (self) |
| Gemini 3.1 Pro | 2026-02 | 2 / 12* | 1M | GPQA 94%, SWE-bench 80.6% (indep.) | |
| xAI Grok 4.5 | xAI | 2026-07 | 2 / 6 | 500K | AA index 54 (independent) |
One thing stands out: transparency slipped. For GPT-5.6 Sol, OpenAI skipped the usual academic benchmarks (SWE-bench Verified, GPQA, AIME) and published only its own agentic metrics (Artificial Analysis). That makes head-to-head comparison harder than it was.
China’s side mostly does the opposite — publish the weights, and use a mixture-of-experts (MoE) design where total parameters are huge but active parameters are small, which keeps inference cheap. Kimi K3 bends that “cheap” rule, though.
| Model | Maker | License | Design (total/active) | In/Out ($/1M) | Headline score |
|---|---|---|---|---|---|
| DeepSeek V4-Pro | DeepSeek | MIT | MoE 1.6T / 49B | 0.435 / 0.87 | SWE-bench Verified 80.6% (Pro-Max) |
| Kimi K3 | Moonshot | TBA (K2 was Modified MIT) | MoE 2.8T / ~50B | 3 / 15 | AA index 57 (independent, top open) |
| GLM-5.2 | Zhipu (Z.ai) | MIT | MoE 744B | 1.4 / 4.4 | AA index 51, AIME 2026 99.2 |
| MiniMax M3 | MiniMax | custom | MoE 428B / 23B | 0.30 / 1.20 | AA index 44, multimodal |
| Qwen3.6-27B | Alibaba | Apache-2.0 | dense 27B | self-host | SWE-bench 77.2, AIME 94.1 |
| DeepSeek V4-Flash | DeepSeek | MIT | MoE 284B / 13B | 0.14 / 0.28 | SWE-bench Verified 79.0 |
Specs come from each lab’s model card and pricing page plus independent leaderboards (DeepSeek pricing, open-weight roundup, Kimi K3). What started this was DeepSeek-R1, released under MIT in January 2025 — the first reinforcement-learning reasoning model shipped fully open, and the first crack in the “the frontier stays private” assumption.
How big is the gap, really?
On overall intelligence, closed still wins — but barely. Artificial Analysis’ index tops out at Claude Fable 5 (60) and GPT-5.6 Sol (59), and sitting right below them is, surprisingly, an open model: Kimi K3 (57), released July 16. As the top open slot moved from GLM-5.2 (51) to K3, the gap fell from about nine points to about three, and K3 outscores even US-closed GPT-5.5 (55).
Source: Artificial Analysis (K3 included)
| Claude Fable 5 | 60 pts |
|---|---|
| GPT-5.6 Sol | 59 pts |
| Kimi K3 | 57 pts |
| GPT-5.5 | 55 pts |
| GLM-5.2 | 51 pts |
| DeepSeek V4-Pro | 44 pts |
Read K3 with care, though. Its flashier coding claims — topping a Frontend Code Arena, beating Claude Fable 5 — mostly trace to Moonshot’s own posts and aggregator blogs. What independent leaderboards confirm is the Artificial Analysis score of 57 (top open, 4th globally); advantages that appear only in vendor tables are best read as claims. And the index itself moves with version and timing — earlier Kimi K2.6 ranged from 43 to 54 depending on the snapshot — so “who measured which version” matters more than the raw number.
The index blends dozens of benchmarks, so the picture changes by axis. Pure reasoning (AIME 2025, GPQA Diamond) has saturated: AIME sits at 99-100% for GPT-5.5 and GLM-5.2, and 96% for Kimi. Almost nothing separates them there. The real gap is in coding and agents (SWE-bench Verified), where Claude Fable 5 leads at a self-reported ~95% — but open has closed in. DeepSeek V4-Pro-Max now posts 80.6%, tied with US model Gemini 3.1 Pro. What you feel in production isn’t “can it solve the problem” but “can it finish a long task reliably.”
There’s one more thing to watch: self-reported and independent scores diverge more on newer models. GPT-5.5’s SWE-bench is 88.7% by OpenAI’s count but 80.6% when lmcouncil measured it — an eight-point spread. Read “who’s first” only after “who measured it.”
About price — now it’s per model, not per camp
If capability is a few points apart, cost is what’s left. But the summer-2026 price sheet isn’t as tidy as it used to be.
Source: each lab's pricing + OpenRouter (July 2026)
| Claude Fable 5 | 50 USD |
|---|---|
| GPT-5.6 Sol | 30 USD |
| Kimi K3 | 15 USD |
| Gemini 3.1 Pro | 12 USD |
| xAI Grok 4.5 | 6 USD |
| GLM-5.2 | 4.4 USD |
| MiniMax M3 | 1.2 USD |
| DeepSeek V4-Pro | 0.87 USD |
| DeepSeek V4-Flash | 0.28 USD |
The floor is still Chinese open: DeepSeek V4-Pro’s $0.87 output is about one fifty-seventh of Claude Fable 5’s $50. But two exceptions just appeared. Kimi K3, for all the buzz, is open yet costs $15 to output — the priciest open model, roughly 17x DeepSeek — Moonshot raised the price along with the performance. And xAI’s Grok 4.5 is US-closed but outputs at $6, cheaper than K3. “China cheap, US expensive” is now a per-model call, not a per-camp one.
The bigger picture still holds, though. API price is only the start: because the weights are open, you can pull the model onto your own GPUs, and past a certain volume self-hosting undercuts metered calls — here’s what that takes. It also keeps your data in-house, which for finance, healthcare or government matters more than price. And “open” isn’t uniform: DeepSeek and GLM are MIT, Qwen is Apache-2.0, but MiniMax adds commercial limits, and Kimi K3’s license wasn’t public at writing (weights were due July 27; K2 was a Modified MIT with attribution at scale). Read the terms before you build on one.
When should you use which?
The choice comes down to constraints, not a leaderboard.
| Situation | Reasonable pick | Why |
|---|---|---|
| Data can’t leave your network | Self-hosted open-weight | Nothing goes out over an API |
| Very high call volume | Cheap open (DeepSeek) or self-host | Unit cost drops by tens of times |
| Hardest reasoning / long agent runs | US closed API | Still ahead on reliability and completion |
| Cost-sensitive but need quality | Cheap closed too — Grok 4.5, GPT Luna | US closed now has low-cost tiers |
| Serving in a specific language | Test both | Public non-English benchmarks are thin — measure it |
Why China opens and the US closes — and the exceptions
This part is structural, not about scores. Since 2022 the US has restricted exports of high-end AI accelerators (Nvidia’s H100 class and up) to China. Short on top chips, Chinese labs took two paths: efficiency — MoE designs that keep active parameters at 13-50B — and openness, releasing weights so the world runs the models on its own hardware. Opening isn’t philanthropy; it’s strategy. When open weights become the default, tooling and fine-tuning stack on top, and that stack is influence. It routes around the chip block through software.
But draw that picture too cleanly and you’re wrong. China’s big labs have started closing their best models too — Alibaba moved its flagship to closed with Qwen 3.7, and Baidu’s ERNIE 5.1 keeps its weights private. And the US has open weights of its own: Nvidia’s Nemotron 3 Ultra (550B) beats MiniMax M3 and DeepSeek V4-Pro on the composite index (Artificial Analysis). “China open, US closed” is a trend, not a law. The shock DeepSeek-R1 delivered in January 2025 was about exposing that split; the summer’s Kimi K3 shows it’s now tangled enough to hold exceptions.
The bottom line
The performance gap shrank to months and a few points, and price has started to split by model rather than by camp. The hardest problems and the most demanding agent work still sit with US closed models — but for the many tasks that don’t need that ceiling, open models are becoming the “good enough, far cheaper” answer. With expensive open models like Kimi K3 and cheap closed ones like Grok, the rational 2026 move isn’t “China or America” — it’s weighing models one at a time, by use.
Figures are as of July 23, 2026 and move fast; models, prices and benchmarks change month to month. Self-reported (self) and independent (indep.) numbers are labeled separately, and anything unverified is left as such. Reconfirm the latest before relying on it.
FAQ
- Are Chinese open models better than US cloud models now?
- At the very top, US closed still leads narrowly — Claude Fable 5 sits at 60 on the composite index. But Kimi K3 (57) has closed the gap to about three points and outscores US-closed GPT-5.5, and for most workloads open models are now the 'good enough and far cheaper' choice.
- What does 'open-weight' actually mean?
- The model's weight files are published, so anyone can download and run it on their own servers. It differs from 'open source' (full code) — the training data usually stays private.
- Are open models free and unrestricted?
- No. Self-hosting removes API fees but adds GPU cost, and licenses vary. DeepSeek and Qwen are commercially permissive; some models require attribution or restrict commercial use. Kimi K3 is also the priciest open model to call by API.
- Which should a company pick?
- If data can't leave your network or cost dominates, self-host an open-weight model. If you need top-end reasoning or agent reliability, a US closed API is safer. Increasingly the answer is to weigh individual models, not whole camps.
