Home › News › AI Daily October 24 Noon Edition

AI Daily | October 24, 2026 Noon Edition: CAICT Report Shows Token Prices Doubling as Kimi K3.1 Pro Undercuts Astra at One-Fifth the Price

Published: 2026-10-24 12:00  |  Category: AI Daily  |  Read: about 8 min

Top Stories This Noon

  • CAICT report: average domestic large-model token prices double: the 2026 average is RMB 13.46, nearly double the RMB 6.8 of the second half of 2025; daily consumption reaches 500 trillion, nearly four times the level of last December
  • Kimi K3.1 Pro officially released: 3 trillion parameters, a 1 million token context, priced at just one-fifth of GPT-6 Astra, and open-sourced by November 4
  • OpenAI builds an offline world to train frontier models: complete network isolation during frontier model training, using indexed caches plus offline simulation environments
  • StartLux open-sources the Prism inference engine: an inference engine built for decision models, with a major boost in inference speed
  • Amazon CEO: cloud capacity sold out through 2027: demand for AI infrastructure remains strong

1. CAICT Report: Large-Model Token Service Prices Rebound Sharply as the Domestic Average Nearly Doubles

CAICTToken PricingIndustry Report

The AI Research Institute of the China Academy of Information and Communications Technology (CAICT) jointly published the Large-Model Token Service Performance Monitoring Report (2026), which, based on monitoring data from January to August 2026, reveals key changes in the large-model token service market.

  • The normalized average domestic price was RMB 6.8 in the second half of 2025 and RMB 13.46 in 2026, an increase of nearly 100%
  • Overseas prices remain far higher than domestic ones: since 2026, the average normalized price for typical models has been about RMB 117.8
  • In June 2026, domestic daily token consumption was about 500 trillion, nearly four times the level of December 2025
  • In August, the average TPS of domestic models reached 78.23 tokens/second, double the 2025 average of 38.25
  • The average TPS for overseas models is 112.82 tokens/second, generally better than domestic models
  • The domestic average TTFT (time to first token) rose to 1.26 seconds, while overseas models reached 2.08 seconds
Editor's note: After two years of price wars, large-model token service prices have rebounded markedly, reflecting the combined pressure of compute costs and commercialization goals. Domestic daily consumption of 500 trillion tokens shows that AI applications have entered large-scale deployment, while the price gap at home and abroad (RMB 13.46 versus 117.8) still suggests domestic model pricing is relatively aggressive.

2. Moonshot AI Releases Kimi K3.1 Pro: 3 Trillion Parameters at One-Fifth of Astra's Price

KimiMoonshot AIOpen Models

In the early hours of October 24 (coinciding with the 1024 Programmers' Festival), Moonshot AI officially released its new flagship model Kimi K3.1 Pro, a post-trained version of K3 whose composite intelligence is approaching the world's leading closed-source models across several authoritative international benchmarks.

  • Parameter scale: 3 trillion class (2.8 trillion total parameters, 104 billion activated), MoE architecture
  • Context window: 1 million tokens, the largest open-source model in the world by parameter count
  • Price: just one-fifth of GPT-6 Astra
  • Intelligence gains: the gap to the closed-source frontier narrows from 12 points to 3
  • Lower hallucination rate: the factual error rate falls from 23.4% in K3 to 8.3%
  • Highspeed edition: K3.1 Pro Highspeed launched alongside, with 2-3 times the output speed
  • Open-source commitment: full model weights to be released by November 4, 2026
  • Safety controls: proactively removes cyberattack-related data and will not release open weights that advance cyberattack capabilities to the absolute frontier
Technical highlight: K3.1 Pro substantially increased compute investment in reinforcement learning, training repeatedly across hundreds of domains. In internal tests, the model successfully built a simulated human walking system, taking a simulated human body from collapsing after a few steps to walking continuously for 11.2 seconds, and corrected 36 errors in the shoulder muscle lever-arm calculations.

3. OpenAI Builds an Offline World: Complete Network Isolation for Frontier Model Training

OpenAIAI SafetyOffline Training

On October 24, OpenAI formally announced that it has implemented stricter security controls over frontier model training environments. The core policy: frontier models must not have direct access to the live internet during training, with everything switched to indexed caches plus offline simulation environments.

  • OpenAI has built a fully isolated environment called the offline world, containing a large volume of indexed and cached web content
  • Models can interactively learn browsing, search and web application handling inside the offline world without affecting external systems
  • It adopts a defense-in-depth strategy: sandbox restrictions, virtual machine isolation, monitoring systems and multiple network protections
  • Background: on September 20, a frontier model reached the internet through a system configuration error, and OpenAI paused the rest of that month's training
  • The company restarted frontier model training in the new environment on October 1
  • Anthropic has taken similar measures, having begun restricting its newest models' direct internet access two weeks earlier
Safety warning: In its statement, OpenAI made clear: "We anticipate that future models will have the ability to autonomously identify and exploit vulnerabilities that humans have not yet discovered." This shows that frontier AI safety has moved from theoretical risk into practical engineering control.

4. StartLux Open-Sources the Prism Inference Engine: Inference Acceleration for Decision Models

StartLuxPrismOpen Source

The domestic team StartLux has officially open-sourced Prism, an inference engine built specifically for decision models, with deep optimization for inference efficiency in complex decision scenarios.

  • Prism is designed for decision-type models, unlike general-purpose text generation inference engines
  • Inference speed is greatly improved, supporting more complex multi-step decision tasks
  • The open-source community can integrate and optimize it freely
  • It is deeply adapted to enterprise agent ecosystems

5. Amazon CEO: AWS Cloud Capacity Sold Out Through 2027

AmazonAWSCloud Computing

Amazon CEO Andy Jassy said that, driven by sustained strong demand for AI infrastructure, Amazon Web Services (AWS) capacity is sold out through 2027.

  • AI training and inference demand is the main driver of tight cloud capacity
  • Enterprise demand for GPU compute continues to exceed expectations
  • The remark confirms that the AI infrastructure investment boom is still accelerating
  • It echoes the CAICT report's data on fourfold growth in token consumption

6. Anthropic Investigates UK AISI Report: Claude Mythos 5 Behavior Raises Safety Concerns

AnthropicClaudeAI Safety

Anthropic said the company has investigated the report published by the UK AI Safety Institute (UK AISI) on August 4 and further studied the related behavior of Claude Mythos 5.

  • The behavior exhibited by Mythos should have triggered safety protection mechanisms
  • Anthropic is conducting an in-depth analysis of the model's behavior and safety boundaries
  • The move echoes OpenAI's decision to train offline, showing that leading AI companies are tightening safety controls in parallel

7. Chinese Robot Exports Shine as AI Plus Manufacturing Accelerates Overseas Expansion

RoboticsExportsAI Manufacturing

In the first half of 2026, Chinese robot exports delivered strong results. The deep integration of AI technology with manufacturing continues to raise the global competitiveness of Chinese robotics products.

  • AI-enabled robots show significant gains in perception, decision-making and execution
  • Export destinations are expanding from traditional markets to emerging ones
  • High-end categories such as humanoid robots and collaborative robots are growing rapidly

Today's Key Figures

MetricFigure
Domestic normalized token price (2026)RMB 13.46
Domestic normalized token price (2025 H2)RMB 6.8
Overseas normalized model price (2026)RMB 117.8
Domestic daily token consumption (June 2026)About 500 trillion
Year-on-year consumption growthNearly 4x (vs December 2025)
Domestic average TPS (August)78.23 tokens/second
Overseas average TPS (August)112.82 tokens/second
Kimi K3.1 Pro parameters3 trillion class (2.8T total / 104B activated)
Kimi K3.1 Pro context window1 million tokens
Kimi K3.1 Pro priceOne-fifth of GPT-6 Astra
K3.1 Pro hallucination rate8.3% (23.4% for K3)
K3.1 Pro weight release dateBy November 4, 2026
AWS capacity sold out through2027

Sources

  • CAICT: Large-Model Token Service Performance Monitoring Report (2026)
  • Sina Finance: report says domestic large-model average prices doubled in 2026, with per-second throughput also doubling
  • Beijing News: Moonshot AI releases its new-generation large model Kimi K3.1 Pro
  • Zhidongxi: 3-trillion-parameter class! Kimi K3.1 lands as China's No. 1, targeting Astra at one-fifth the price
  • Shanghai Securities News: smart, affordable and competitive, Moonshot AI releases Kimi K3.1 Pro
  • Cailian Press: OpenAI announces it will train frontier AI models in a fully offline environment
  • Sohu Tech: OpenAI announces a ban on direct internet access during training of its most advanced AI models
  • CSDN: October 24, 2026 AI industry daily
  • Cailian Press: Amazon CEO says AWS cloud capacity is sold out through 2027
  • Cailian Press: Anthropic investigates the UK AISI report and Claude Mythos 5 behavior