September model releases push agentic and multimodal systems forward

The latest LLM release cycle through Sep 25, 2026 was led by a dense run of agentic and multimodal launches. Perceptron released Mk1.5 for embodied reasoning and robotics perception, while Upstage introduced Solar Mini 4 as a 35B-total / 3B-active MoE with a 524,288-token context. Fireworks debuted Ember-1 as a Research Preview focused on shorter reasoning traces for coding, knowledge work, and agentic workflows.

Major frontier labs also refreshed their lineups. OpenAI launched GPT-6 Luna and GPT-6 Sol on 2026-09-22 after releasing GPT-6 Astra earlier in the month, positioning Astra for computer/browser use, professional artifacts, long-session coding, and defensive cybersecurity. Anthropic released Claude Opus 5.5 for long-running agentic coding and knowledge work, reporting roughly Claude Fable 5.1-level quality on most tasks at about 40% lower running cost than Opus 5.

Open-weight activity remained strong across China and Asia. Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash under MIT, Alibaba’s Qwen team shipped Qwen3.8-Omni-Flash, DeepSeek released V4.1-Flash with a 1M-token context and reduced KV-cache memory, and Shanghai AI Laboratory published Atria Dawn Preview. The month also brought specialized releases for CRM reasoning, form filling, cybersecurity, coding, extraction, and on-device deployment.