OpenAI Releases GPT-5.6 Model Family and Moonshot Launches Kimi K3 Coding Models
모델 출시 | Thu Jul 30 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 3 sources
OpenAI unveiled the GPT-5.6 Sol, Terra, and Luna three-tier lineup while Moonshot expanded its Kimi K3 coding model lineup.
Analysis
[OpenAI] released the GPT-5.6 Sol, Terra, and Luna three-tier model family [2]
- Sol is the flagship model that surpasses Claude Fable 5 at less than half the cost with max reasoning
- Terra delivers GPT-5.5 level performance at half the price
- Luna is the fastest and cheapest
- priced 80% lower than Sol
- Designed to balance capability and cost across the spectrum of tasks
[OpenAI Inference Stack] optimized infrastructure efficiency using GPT-5.6 Sol [2]
- Significantly reduced serving costs through improved load balancing
- GPT-5.6 Sol autonomously rewrote Triton and Gluon kernels
- cutting end-to-end serving costs by 20%
- Improved speculative decoding boosted token generation efficiency by over 15%
- Verified kernel correctness with the FpSan open-source tool
[OpenAI Agentic Harness] optimized context management and prompt caching [2]
- Exposes tools and plugins only when needed via deferred discovery
- Tool outputs are limited to 10
- 000 tokens by default
- Maximizes prompt cache hit rate with an append-only history structure
- Rust orchestration layer connects models
- tools
- and user environments
[OpenAI GPT-5.6 Sol] tripled ARC-AGI-3 benchmark score through harness configuration [1]
- Enabling retained reasoning and compaction tripled the score and reduced output tokens by 6x
- Score rose from 13.3% in the official harness to 38.3%
- Rolling truncation and reasoning disposal in the harness caused performance degradation
- Reasoning is automatically preserved by using the Responses API and passing previous response IDs
[Moonshot Kimi Code] unveiled the Kimi K3 and K3-256k coding model lineup [3]
- Kimi K3 is the flagship coding model with 2.8T parameters and a 1M context window
- K3-256k is a 256k context version that effectively reduces consumption
- K2.7 Code HighSpeed delivers output speeds 5-6x faster
- K3 supports both images and video
- while K3-256k supports images only