China's AI company Moonshot unveiled its new flagship model "Kimi K3" —and in independent tests, it’s nearly on par with the best models from Anthropic and OpenAIin independent tests, which may have contributed to the sharp decline in AI stock prices.

What happened?

Moonshot AI, a Chinese AI company with a valuation of approximately $31.5 billion, has unveiled its new model, Kimi K3. By comparison, Anthropic and OpenAI are each projected to be valued at over $1 trillion —more than 30 times that amount.

Key figures:

Size: 2.8 trillion parameters (= the “adjustment screws” that store the model’s knowledge—roughly speaking: the more there are, the more powerful the model, but also the more expensive it is to run). This makes K3 the largest model ever to come out of China.

Open-Weight: Moonshot plans to make the model available for free download by July 27—companies could then run it on their own hardware without paying licensing fees to Moonshot. As of now, however, it’s only available through Moonshot’s paid interface.

Context: Capable of processing 1 million tokens at a time (tokens = text units; 1 million tokens correspond roughly to 750,000 words).

Unknown: The actual training costs. Anthropic also published evidence in February 2026 that Chinese labs had copied U.S. models via “distillation” (= a model learns by querying and imitating vast amounts of responses from another model). Whether and to what extent this plays a role in K3 remains unclear.

The comparison: How much do the models cost—and how smart are they?

Pricing is per 1 million tokens, broken down by input (what you send to the model) and output (what it responds with):

Kimi K3 (Moonshot): $3 input / $15 output – Intelligence Index: 57

Claude Fable 5 (Anthropic): $10 / $50 – Intelligence Index: 60 (1st place)

GPT-5.6 Sol (OpenAI): $5 / $30 – Intelligence Index: 59

Claude Opus 4.8 (Anthropic): $5 / $25 – Intelligence Index: 56

Grok 4.5 (xAI): $2 / $6 – the most affordable model among the top performers

DeepSeek V4 Pro (China): approx. $0.44 / $0.87 – low-cost segment, but index only 44

The Intelligence Index comes from Artificial Analysis, an independent testing provider that compares models across nine task areas (programming, logic, knowledge work) on a scale of up to 100.

Strengths – K3 trails the leader, Fable 5, by only 3 points, but costs only one-third as much per output token ($15 instead of $50) and half as much as GPT-5.6 Sol. In Artificial Analysis’s practical test, completing a task with K3 cost an average of $0.94—compared to $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8. When it comes to building web interfaces, K3 even took first place in blind tests with real developers.

Weakness – K3 isn’t quite as cheap as the headline suggests: It costs three to four times as much as its predecessor, is the most expensive model among its Chinese competitors, and, according to tests, “burns” up to 1.9 times as many tokens as GPT-5.6 Sol for the same task—so the cost advantage per token effectively disappears in practice. And the promised free model weights have not yet been released.

What could this mean for the market?

The AI stock boom hangs by a thread: U.S. labs like Anthropic, OpenAI, and xAI have committed to spending hundreds of billions of dollars—on chips, data centers, and electricity. Chip manufacturers, data center operators, energy providers, and construction companies have seen their stock valuations rise because this money is expected to flow to them. The plan is to refinance this through high prices for cutting-edge AI.

If a model now delivers nearly the same performance for a fraction of the price—and is soon expected to be available for free download—it could create the impression that the U.S. labs’ pricing power is crumbling (“commoditization of intelligence”: cutting-edge AI is becoming an interchangeable mass-market commodity). This wouldn’t immediately cost revenue, but it sows doubt about the business case behind the massive investments—and in overheated market phases, even a single doubt can trigger a chain reaction.

Opportunities – Cheaper AI could accelerate AI adoption overall: More adoption means greater demand for computing power for ongoing operations (inference)—and K3 is just as hardware-intensive in operation as U.S. models. Chip and infrastructure providers would benefit from this, regardless of which model is running.

The latest quarterly figures support this interpretation: On Tuesday, ASML raised its annual forecast for the second time (to €43–45 billion in revenue) and reported that its capacity is nearly fully booked through 2027. On Wednesday, TSMC increased its capital expenditure budget to a record-high $60–64 billion and stated that investments over the next three years would once again be significantly higher than those of the past three years—and that it is monitoring the construction progress of data centers to ensure its own chips do not end up in inventory. This can be interpreted as a sign that the demand side remains strong.

Risks – If price compression were to actually hit U.S. labs, training investments could be stretched out or cut back—which would eventually affect the recipients of this spending (chips, data centers, energy). Furthermore, those who train using distillation save precisely the billions in GPU costs that make the U.S. approach so expensive. The cost bases of the providers are therefore not directly comparable, but the price risk for U.S. labs still exists. A countervailing factor is that Chinese models are unlikely to be considered for security-critical U.S. applications (government agencies, robotics) due to potential backdoors.

Context

The bottom line, in my view, is that Kimi K3 is less a “DeepSeek moment 2.0” than a reality check for the valuation logic:

In 2025, the question was which chips China would use for training—now the question is whether cutting-edge intelligence permanently justifies premium prices.

The supply chain’s response (ASML, TSMC) has been clear so far: demand exists—the response at the model level is less certain.

Limitations

Initial benchmarks are partly based on manufacturer specifications = independent long-term tests are still pending.

Training costs and the potential distillation ratio for K3 cannot be verified.

Moonshot’s valuation ($31.5 billion) is from a private funding round; the trillion-dollar valuations of Anthropic and OpenAI are prospective—neither is directly comparable to market capitalizations.

Sources:

Artificial Analysis – Kimi K3 Benchmarks: https://x.com/ArtificialAnlys/status/2077832874183860404

OpenRouter – Kimi K3 Prices & Specs: https://openrouter.ai/moonshotai/kimi-k3

Anthropic – Official API Pricing Documentation: https://platform.claude.com/docs/en/about-claude/pricing

OpenRouter – Grok 4.5 Pricing: https://openrouter.ai/x-ai/grok-4.5

Price Comparison: Kimi K3 / Fable 5 / GPT-5.6 Sol: https://drawpie.com/blog/kimi-k3-vs-fable-5-vs-gpt-5-6-sol-price-benchmarks/

Decrypt – Cost per Task in Benchmark Comparison: https://decrypt.co/373716/china-kimi-k3-largest-open-source-ai-model-ever-beats-claude-fable-gpt-5-6-sol

The Decoder – K3 Classification & Price Level: https://the-decoder.com/kimis-open-model-k3-nears-gpt-5-6-sol-and-fable-5-while-signaling-the-end-of-super-cheap-chinese-ai/

ASML – Q2 2026 Earnings Call Transcript: https://www.investing.com/news/transcripts/earnings-call-transcript-asml-q2-2026-beats-guidance-as-ai-demand-lifts-outlook-93CH-4792156

TSMC – Q2 2026 Earnings Call Transcript: https://www.investing.com/news/transcripts/earnings-call-transcript-tsmc-lifts-2026-outlook-as-ai-demand-stays-hot-in-q2-2026-93CH-4794777

Anthropic – Report on Distillation Attacks (February 2026): https://anthropic.com/news/detecting-and-preventing-distillation-attacks

$NVDA (-0,72 %)
$2330
$ASML (-1,16 %)
$MSFT (+0,19 %)
$AMZN (+0,08 %)
$GOOGL (+0,29 %)