跪拜 Guibai
← Back to the summary

Kimi K3 Tops Code Benchmarks, Then Locks Its Best Features Behind a $96/Month Paywall

1. Code Track Tops the Leaderboard — Development Capability Fully Loaded

After Arena.AI's latest front-end code leaderboard update, Kimi-K3 took first place directly, with a score of 1679 leaving a group of overseas models behind by a margin. Second place is Claude Fable 5 at 1631; third place GPT-5.6 Sol only scored 1618. The remaining Claude models on the list — Opus, Sonnet — along with domestic models like Zhipu, ByteDance, and Tongyi Qianwen, all have noticeably lower scores. For everyday development needs like writing business logic, fixing bugs, and refactoring engineering code, the scores show that its overall output quality is indeed better than the vast majority of current mainstream large models.

img_v3_0213m_4f385867-b2d6-45e8-ba3e-33a8d95f496g.jpg

2. Real Workplace Tasks and Long-Chain Multi-Step Tasks — All in the First Tier

Leaving aside the code-specific track, two evaluation suites biased toward real office scenarios better reveal its foundation for professional work.

img_v3_0213m_11a039aa-bdf6-4b71-a1ec-7501e3dd519g.webp

  1. GDPval-AA v2 Professional Comprehensive Evaluation This leaderboard covers 44 professions and 9 major industries, specifically measuring an AI's ability to handle real-world work tasks. K3 scored 1687, ranking only behind two top overseas models, Claude Fable 5 Max and GPT-5.6 Sol Max, and significantly higher than the commonly used Claude Opus 4.8 Max (1600 points). Whether writing industry proposals, organizing research materials, or doing data analysis — the kind of daily office work — its overall completion quality is in the first tier.
  2. AA-Briefcase Long-Chain Agent Task Test This benchmark specifically measures the ability to handle multi-step, long-duration coherent work, suitable for doing a complete project: from finding materials, building a framework, writing in sections, to proofreading and optimization. K3 scored 1527, second globally, only behind Claude Fable 5 Max, while surpassing GPT-5.6 Sol Max. If you frequently use AI for coherent long-chain work, its advantages in contextual memory and sustained output stability are obvious.
  3. Million-Token Context Brings a Unique Tier of Long-Text Retrieval Capability K3 comes standard with a 1-million-token super-large context window, achieving the industry-best score of 91.2 on the BrowseComp long-text retrieval test. Without manually splitting documents or compressing text, you can directly feed in hundreds of thousands of words of manuals, multiple reports, or entire source code repositories, and the model can precisely locate key information and cross-reference multiple materials. Massive long-text processing and deep document retrieval are its most prominent strengths.

3. Complaints Fully Loaded: Four Membership Tiers Split Core Features, Full Version Has a High Barrier

Help me boost and win big prizes together! Complete Kimi registration, and both you and I will 100% win a prize, up to 1 year of membership equivalent benefits: https://kimi-bot.com/activities/zh-cn/viral-referral/share?scenario=invite&from=share_poster&invitation_code=9QLPRA

image.png

Although the model's performance improvement is eye-catching, the membership pricing rules this time have drawn complaints from many long-time users. It is divided into four monthly subscription tiers: 49 RMB Andante, 99 RMB Moderato, 199 RMB Allegretto, and 699 RMB Allegro. The permissions for two core scenarios are completely separated, with very granular restrictions:

  1. Coding scenario (Kimi Code) The 49 RMB tier cannot use K3 at all; the 99 RMB tier only unlocks a 256K context window. To use K3's full 1M window, you need to subscribe to at least the 199 RMB/month tier or above.
  2. Mobile/Client General Chat Even if you subscribe to the 199 RMB mid-tier membership, you cannot use the full 1M context on mobile. Only the most expensive 699 RMB top-tier membership unlocks the full version of K3 on the client.

Simply put: to fully experience its strongest million-token context and top-tier coding ability, you need to spend at least 199 RMB per month; to use the full-spec long-text processing on your phone anytime, anywhere, you have to go straight to the top-tier 699 RMB. For the average office worker or light developer, the cost is on the high side.