Launches on 2026-09-28

VIDRAFT’s open 180B reasoning model: Alibaba’s Qwen3.8-Flash-Next retrained on its own checked answers, reading text and images, with a built-in confidence score that is returned before it answers.

Anthropic’s mid-tier model at $2 in and $10 out per million tokens, half Opus 5.5’s price, with a 1M-token context, 128,000 output tokens and a retirement floor of 28 September 2027. It replaces Sonnet 5.

IQuestLab’s open agentic-coding model: a 320B mixture of experts with 15B active, a 524,288-token context and English and Chinese text, for self-hosting with SGLang or vLLM. Modified MIT licence with a display requirement.

A managed Postgres platform for AI applications from pgEdge: built-in MCP, RAG and PostgREST servers, pgvector search and copy-on-write branches, movable to your own cloud or on-premises. Managed plans from $25 a database a month.

ElevenLabs’ text-to-speech model released 28 September 2026: a new architecture for expressive speech in 90+ languages, with multiple speakers, inline emotion and sound-effect tags, and long-form stitching. Hosted only.

The low-latency member of ElevenLabs’ v4 text-to-speech family, released 28 September 2026 for voice agents: ElevenLabs states a median inference latency of about 100 ms and about 150 ms to first speech. Hosted only.

Nace AI's decision model, about 9B parameters: takes a state as text or JSON plus typed questions and returns a probability for each option, not free text. Open weights under a modified Open RAIL-M licence.