Moonshot AI released the Kimi K2.6 model on April 21, 2026. This release triggered a traffic surge that exceeded company expectations and caused service unavailability for some users. A discrepancy in the backend system for tracking Agent usage quotas also led to erroneous deductions for certain user entitlements. The surge in traffic led to member queues and brief service unavailability. Moonshot AI announced a compensation plan on the evening of April 22, 2026, via its official social media account @Yue Zhi An Mian Kimi to reset all users’ monthly quotas to 100% and directly clear their used amounts after the K2.6 launch caused massive traffic surges. This decision aimed to restore the user experience after the system error. The K2.6 model achieves wide-ranging improvements in general Agent capabilities, code generation, and visual understanding. Its model-driven Agent cluster architecture supports 300 sub-agents working in parallel to complete 4,000 collaborative steps. The K2.6 model enables uninterrupted coding for 13 hours and processes over 4,000 lines of code. For high-load workflows, the K2.6 model integrates automated task processing for proactive Agent frameworks like OpenClaw and Hermes Agent. These frameworks support continuous autonomous operation for up to five days. The K2.6 performance matches or exceeds international models like GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro. I find the decision to reset quotas for everyone a blunt way to manage user frustration.
API pricing and context windows
Developers use the Moonshot AI API platform to access various models. The Kimi K3 model provides a 1,048,576-token context window for long-horizon coding and knowledge work. Kimi K3 costs $0.30 per million tokens for cache-hit input, $3 for cache-miss input, and $15 for output. The K2.6 model provides a 256K context window for general-purpose multimodal tasks. K2.6 costs $0.16 per million tokens for input, $0.95 for cache-hit input, and $4 for output. K2.7 Code also provides a 256K context window and accepts text, image, and video input. K2.7 Code costs $0.19 per million tokens for input, $0.95 for cache-hit input, and $4 for output. The HighSpeed variant of K2.7 Code uses the same model at higher output speeds. HighSpeed costs $0.38 for input, $1.90 for cache-hit input, and $8 for output. K3.1 Pro Highspeed provides higher speed for intensive workloads and released on October 23, 2026. For membership users, Kimi provides tiers like Adagio, Moderato, Allegretto, Allegro, and Vivace. Adagio costs $0 per month. Moderato costs $15 per month and includes more agent credits and Kimi Code. Allegretto costs $31 per month and provides five times the Kimi Code credits. Allegro costs $79 per month and provides up to a 1M-token K3 chat context. Vivace costs $159 per month and increases multipliers to 10 times for agent credits and 30 times for Kimi Code credits. You should distinguish Kimi membership benefits from pay-as-you-go platform API charges when planning spend.
| Model | Context | Input ($/1M) | Output ($/1M) |
|---|---|---|---|
| Kimi K3 | 1,048,576 | $3.00 | $15.00 |
| Kimi K2.6 | 256K | $0.95 | $4.00 |
| K2.7 Code | 256K | $0.95 | $4.00 |
| K2.7 HighSpeed | 256K | $1.90 | $8.00 |
Competition against ByteDance and US adoption
ByteDance manages the Doubao model, which reached 163 million monthly active users. Moonshot AI recently secured $500 million in Series C funding to expand its compute infrastructure. This capital infusion raised the company’s valuation to $4.3 billion. While Chinese firms compete locally, Moonshot AI models gain popularity in the United States. Raffi Krikorian, the chief technology officer at Mozilla, switched to the Kimi K3 model for daily tasks because it felt faster than the Claude chatbot. Other US firms, including cryptocurrency exchange Coinbase, have explored Chinese models to reduce costs. Kimi K2.5 also scored 50.2 on the Humanity’s Last Exam benchmark and beat the 45.5 score of GPT-5.2. DeepSeek also attracted global attention earlier this year when it launched a model that matched some US systems at a much lower cost. Z.ai’s GLM-5.2 also saw use for managing documents and emails. The rise of Chinese AI systems has raised concerns among some US technology companies regarding intellectual property and access to advanced technology. Washington has imposed restrictions on China’s access to some advanced AI chips. I consider Moonshot AI’s recent expansion a high-stakes gamble to secure dominance through sheer compute power. Analysts note that the demand for cost-effective AI models grows as companies use AI agents for multi-step tasks. Experts say this competition helps Chinese AI companies expand their influence globally. Can Moonshot AI sustain its growth while it manages the high costs of research and development?
