100 RMB per Million Tokens: With Kimi Adopting "Affordable Luxury" Pricing, Will You Still Pay for K3?

Note: These are Jeffrey Ding’s informal and unofficial translations -- all credit for the original goes to the authors and the original text linked below. These are informal translations and all credit for the original work goes to the authors. Others are welcome to share excerpts from these translations as long as my original translation is cited. Commenters should be aware that the Google Doc is also publicly shareable by link. These translations are part of the ChinAI newsletter - weekly-updated library of translations from Chinese thinkers on AI-related issues: https://chinai.substack.com/ 

________________________________________________________________________________

Author: Hui Lei

Source: 第一新声 (Diyi xinsheng) — Beijing-based information platform focused on the AI value chain, with an affiliated think tank

Date: July 23, 2026

Original Mandarin: https://mp.weixin.qq.com/s/VK3Vyro2b5Q4wdt2HPzxOg 

"Attack the greatest challenges and attempt the most far-reaching goals." After opening with this classical Chinese language, Moonshot AI dropped a bombshell in the early hours of July 17: Kimi K3.

This was no ordinary version update; it was the debut of a "beast" boasting 2.8 trillion parameters and a context window of 1 million tokens—with a promise to go open-source in just ten days. The news sent Chinese social media into a frenzy; "Kimi K3" shot up the trending lists, and the comment sections were a mix of awe and skepticism. Some hailed it as an "honor for Chinese technology," while others gasped at the API pricing, which rivaled that of the hypothetical GPT-5.6 Terra[a].

Yet, the real drama of this launch unfolded on the other side of the ocean.

When Code Arena frontend programming leaderboard updated to show Kimi K3 claiming the global top spot with 1,679 points topping Claude Fable 5, Silicon Valley was instantly jolted wide awake.

Elon Musk[b], who always speaks his mind bluntly, commented "Impressive" on X, OpenAI executives questioned the open-source strategy, and the PHLX Semiconductor Sector Index[c] took a tumble.

Why did the model launch of a Chinese startup manage to rattle both Silicon Valley and Wall Street? What exactly has changed thanks to this open-source beast unleashed by Moonshot AI?

01. Parameters Are Merely an Entry Ticket; Hardcore Capability Is True Ability

To understand the source of this storm, we must first dissect the "core" of Kimi K3.

What immediately catches the eye is a dizzying array of figures: a total of 2.8 trillion parameters, making it the world's first open-source model in the 3-trillion-parameter class. Yet, scale is merely the surface; the real key lies in how this beast was tamed.

Source: https://www.kimi.com/blog/kimi-k3 

Officially disclosed technical architecture reveals that K3 is built upon the Kimi Delta Attention (KDA) hybrid linear attention mechanism and Attention Residuals technology.

In layman's terms, the former ensures the model isn't overwhelmed by computational load when processing ultra-long sequences (such as entire novels or codebases), while the latter prevents information loss as data propagates through networks spanning hundreds of layers. Together, these technologies underpin the efficient operation of a model with trillions of parameters.

Meanwhile, K3 employs a Mixture of Experts (MoE) sparse architecture, dynamically activating only 16 out of 896 experts. Coupled with the Stable LatentMoE framework and optimized training data recipes, its overall scaling efficiency has improved by approximately 2.5 times compared to its predecessor, K2. This means K3 can "squeeze out" more intelligence from the same amount of computing power. In the authoritative Artificial Analysis intelligence index, K3 scored 57 points, firmly securing third place globally, trailing only the formidable Claude Fable 5 and GPT-5.6 Sol.

However, what has truly electrified the developer community is the combination of its "long-range programming" and "native multimodal" capabilities. The use cases showcased by the developers are striking: in a GPU kernel optimization arena, K3 autonomously completed kernel rewriting and score verification within 24 hours, performing on par with Fable 5; it even built a compact compiler named MiniTriton from scratch, achieving performance comparable to the industry-standard Triton.

Even more impressive is K3's ability to seamlessly integrate 3D reasoning, programming, and visual capabilities. Whether it is recreating the visual effects of the black hole "Gargantua,"[d] designing a 4mm² chip while meeting timing closure tests, or replicating the computational workflow of an astrophysics paper in just two hours, K3 demonstrates "Agent-level" productivity that far transcends the limitations of a mere "chatbot."

In short, K3 is no longer an "intern" capable only of generating code snippets; it has become a "full-stack engineer" capable of independently executing complex engineering projects.

02. A Turnaround in Perception and "Dimensional-Drop" Pricing: When the "Distillation Argument" Expires and the Price Structure Collapses

Following the release of Kimi K3, an intriguing phenomenon emerged in China’s domestic public discourse: skepticism regarding "distillation" virtually vanished.[e]

In the past, whenever domestic large models achieved a breakthrough, some critics would simply attribute it to "stealing techniques via distillation." This time, however, K3’s self-developed architecture—boasting 2.8 trillion parameters—and its top ranking in Code Arena rendered that catch-all explanation completely invalid.

If distillation alone were enough to easily surpass models like Claude Fable 5 and GPT-5.6 Sol, why would the global AI landscape remain a duopoly between China and the US? This shift in public sentiment signals that domestic models are being re-evaluated—moving beyond the label of mere "chasers."

Meanwhile, across the ocean, capital markets cast an even more dramatic vote with real money.

After K3’s launch, a sell-off of global semiconductor stocks intensified, with the Philadelphia Semiconductor Index (SOX) recording its worst weekly performance in nearly a year. Goldman Sachs characterized the decline as a "deleveraging event," explicitly noting that the issue was not an invalidation of the Scaling Law, but rather that simply increasing pre-training compute power was no longer the sole path to achieving a leading model.

As K3 demonstrated through architectural innovation that "brute-forcing with sheer power" and "finesse" can coexist, the market began to reassess profit growth expectations for storage, chip equipment, and even entire data centers.

The root of this panic lies in how Kimi K3’s pricing strategy precisely punctured the premium price bubble surrounding closed-source models.

For comparison, K3’s API pricing for input (cache miss[f]) and output is 20 RMB (approx. $3) ​​and 100 RMB (approx. $15) per million tokens, respectively—representing just 20% of the cost of Claude Fable 5 and 40% of Opus 4.8. Thanks to the "Mooncake" decoupled inference architecture, the cache hit rate in programming scenarios exceeds 90%, effectively cutting actual input costs by another 75%.

[g]

When open-source models offer capabilities that are close to or even partially surpassing closed-source flagships at one-tenth of the cost, the "tech moat plus high-priced subscription" model that Anthropic and OpenAI rely on has begun to show its first cracks.

03. Industry Watershed: From "Competing on Compute" to "Competing on Efficiency"— The Pricing Power of the Closed-Source Camp Loosens

The launch of Kimi K3 is far more than a mere product iteration; it marks a pivotal watershed moment for the global AI industry.

First, it signals that the "pre-training race based purely on piling up compute power" has entered its second half.

In recent years, the industry was locked in a "bigger is better" arms race regarding parameter counts, characterized by a frenzy to purchase Nvidia GPUs. The arrival of K3 demonstrates that qualitative leaps can also be achieved through innovations in model architecture (e.g., KDA/AttnRes), optimizations in training methods (e.g., quantization-aware training, expert parallelism), and improvements in data quality.

For Chinese AI companies constrained by limited compute resources and a tightening financing environment, this points toward a more sustainable path: shifting the competition from capital intensity to "intelligence" (algorithmic efficiency).

Second, it fundamentally undermines the investment thesis that "closed-source equals a competitive moat."

For the past two years, the narrative underpinning the high valuations of companies like OpenAI and Anthropic was simple: "Only we can build the smartest models, and you must pay a premium to use them." With the open-sourcing of K3, however, SMEs and developers worldwide can access a top-tier foundational model for self-deployment without paying exorbitant subscription fees to closed-source vendors.

[h]

When the capabilities of open-source models catch up to 90% of their closed-source counterparts—while costing only 20% as much—the pricing power of closed-source vendors begins to erode irreversibly. It is precisely this shift in pricing power that is causing anxiety on Wall Street.

As for Kimi K3’s pricing strategy, it can be summarized in one sentence: by matching the performance of flagship overseas models while anchoring prices to the ceiling for domestic models, thereby driving China’s large models to move beyond “cutthroat price competition” toward “value monetization.”

Let’s look at some hard data. According to official figures from the Moonshot AI open platform, K3’s API pricing is set at: 20 RMB per million tokens for input (cache miss) and 2 RMB per million tokens (cache hit), with output priced at 100 RMB per million tokens. Converted to USD, this equates to $3 for input and $15 for output. A cross-comparison with overseas flagship models reveals the following: According to Anthropic’s official website, Claude Fable 5 is priced at $10 for input and $50 for output, while OpenAI’s GPT-5.6 Sol is priced at $5 for input and $30 for output. K3’s list price is approximately one-third that of Fable 5 and 60% of Sol’s price.

[i]

Domestically, the disparity is even more pronounced. A Goldman Sachs research report indicates that K3’s blended rate[j] is approximately $2.30 per million tokens—significantly higher than Alibaba’s Qwen3.7 Max ($1.40), Zhipu’s GLM-5.2 ($0.90), MiniMax’s M3 ($0.22), and DeepSeek’s V4 Pro ($0.18). For comparison, Zhipu’s GLM-5.2 has input and output prices of $1.40 and $4.40, respectively; K3’s output price is roughly 3.4 times that figure. Furthermore, K3’s API output price represents a more than 3.5-fold increase over its predecessor, the K2.6.[k]

So, what justifies Kimi K3’s pricing strategy?

First, superior performance. In the "Frontend Code Arena" blind test for frontend programming, K3 topped the global rankings with a score of 1,679, surpassing both Claude Fable 5 (1,631 points) and GPT-5.6 Sol (1,618 points). In the Artificial Analysis intelligence index, K3 ranked third globally with a score of 57.

Morgan Stanley noted that K3 demonstrates how Chinese large models have fully caught up with leading US vendors across three dimensions: scale, performance, and pricing.

Second, cost efficiency. A high list price does not equate to high actual costs. Moonshot AI disclosed that, thanks to its proprietary "Mooncake" decoupled inference architecture, K3 achieves a cache hit rate exceeding 90% in programming scenarios, meaning the actual input cost for enterprise users is only one-quarter of the list price. Artificial Analysis estimates K3’s average cost per task at approximately $0.94—roughly on par with GPT-5.6 Sol ($1.04) and half that of Claude Opus 4.8.

[l]

Third, strategic positioning. This is not merely a price hike; it is a carefully orchestrated battle for "pricing power." Goldman Sachs characterized the K3 pricing as setting a "new high for domestic models," while Morgan Stanley viewed it as a landmark signal of a "pricing correction" within the domestic large model industry—marking the end of the cutthroat "low-price-for-traffic" model that dominated the past two years and shifting the sector toward a virtuous cycle where performance drives monetization, which in turn fuels further R&D. The market cast its vote with real capital: just 48 hours after launch, Kimi had to suspend new consumer subscriptions due to a shortage of computing power, while Moonshot AI’s market share surged from 1.6% to 7.9%.

Ultimately, it reshaped the psychological landscape of global AI competition. Elon Musk’s endorsement and OpenAI’s skepticism formed an intriguing contrast: the former identified it as a rival, while the latter perceived it as a threat.

Moving from being "left in the dust" to "breathing down the leader's neck," Chinese AI is no longer merely a "cost-effective alternative"; it is beginning to define the technological frontier.

04. The "K3 Moment" in the Open-Source World: Making Intelligence Accessible to All is no longer Just a Slogan

Although Kimi K3 still lags noticeably behind Claude Fable 5 and GPT-5.6 Sol in terms of complex reasoning and user experience, its greatest contribution may lie in reigniting the open-source community's imagination regarding "top-tier models."

Historically, open-source models played the role of "chasers"—replicating whatever closed-source models released, but at a lower cost.

K3, however, chose a more arduous path: actively pushing parameter scales to new limits, challenging the boundaries of next-generation capabilities—such as million-token context windows and long-horizon agents—and committing to a full open-source release on July 27.

This combination of "pushing capabilities up, driving prices down, and making code public" allows developers worldwide—particularly those in resource-constrained research institutions and Global South nations—to stand on equal footing at the starting line of the intelligent era. This aligns perfectly with the Chair’s Statement issued at the 2026 World Artificial Intelligence Conference, which declared that "open source and openness are vital pathways for the inclusive development of artificial intelligence."

Of course, significant challenges remain. K3 is not without flaws; the company has candidly acknowledged limitations such as sensitivity to thinking history[m] and a tendency to be "excessively proactive[n]." Furthermore, computing resources are currently strained due to a surge in users, prompting the introduction of purchase limits for premium members.

Additionally, it remains to be seen whether the high price point will be accepted by ordinary consumers and whether enterprise clients are willing to pay for the long-term stability and security of an "open-source model."

Nevertheless, Kimi K3 has demonstrated its capabilities and made a significant breakthrough: it proves that on the journey toward AGI, Chinese players can do more than just "keep pace" or "run alongside" global leaders—they can even "take the lead" in specific arenas.

As technological barriers crumble line by line before code destined for open release, and as the spark of knowledge passes from hand to hand among developers worldwide, an era of intelligence defined by genuine co-creation and shared progress may finally be taking shape.

[a]From OpenAI's website description: Terra, a balanced model for everyday work, GPT‐5.6 Terra and GPT‐5.6 Luna outperform Fable 5 at around one-sixteenth the cost

[b]Chinese tech media is still obsessed with Elon Musk

[c]From Wikipedia: The PHLX Semiconductor Sector (SOX) is a capitalization-weighted index comprising the 30 largest U.S.-traded companies (not necessarily U.S. companies) primarily involved in the design, distribution, manufacture, and sale of semiconductors.

[d]Interstellar reference!

[e]This is an interesting note

[f]When it's a cache miss, the system has to process your full prompt rather than routing your request to a server that recently processed the same prompt.

"Model prompts often contain repetitive content, like system prompts and common instructions. OpenAI routes API requests to servers that recently processed the same prompt, making it faster and less expensive to reuse an exact prompt prefix than to process it from scratch." https://developers.openai.com/api/docs/guides/prompt-caching

[g]Cool to learn more about cache misses and hits

[h]But most K3 users probably won't be able to self-host it

[i]good details

[j]This is essentially an average cost of use metric; assuming that 70% of your tokens go to inputs that are cache hits; 20 % go to fresh inputs (cache misses), and 10% go to outputs

[k]This is a really interesting comparison

[l]Source: https://artificialanalysis.ai/models/kimi-k3.

Some other notes from Artificial Analysis: 

"Kimi K3 is amongst the leading models in intelligence, but somewhat expensive when comparing to other models of similar price. It's also notably slow and very verbose."

"When evaluating the Intelligence Index, it generated 130M tokens, which is very verbose in comparison to the median of 63M."

[m]"K3 was trained in the preserved thinking history mode. If the agent harness fails to pass back all the historical thinking content as required, or if an ongoing session with another model is switched over to K3, generation quality may become highly unstable. We recommend using a harness with verified compatibility, such as Kimi Code, and avoiding switching to K3 in the middle of a session." https://www.kimi.com/blog/kimi-k3

[n]This seems important to track: "Excessive proactiveness. K3's training places particular emphasis on long-horizon, challenging tasks. As a result, when it encounters minor issues or ambiguous user intent during task execution, it may make unexpected decisions on the user's behalf. If your application requires the agent to operate within well-defined boundaries and refrain from excessive improvisation, please impose more explicit behavioral constraints on K3 in the system prompt or in AGENTS.md." https://www.kimi.com/blog/kimi-k3