AI Model Release Tracker

Still training as of Aug 2026. Q1/Q2 windows passed. Leading rumor: ~6T MoE on Colossus 2, multi-agent 4→16+. No official date.

RUMOR: 2.1T params, 'better than 4.6 in every way', slightly slower to serve. Expected a few weeks after 4.6 (late Aug / early Sep).

Embodied reasoning model for robotics: video understanding, multi-step task orchestration, multi-robot collaboration, stronger safety.

RUMOR: Musk targeted ~Aug 7, 1.5T params with significantly improved SFT & RL. Target missed as of Aug 11 (xAI docs still point to 4.5). 'Musk time' may slip to Aug 14.

GPT-5.6 family fully opened: Sol ($5/$30), Terra ($2/$12), Luna ($0.2/$1.2). 1.05M context, 128K output, knowledge cutoff 2026-02-16. Originally delayed in June by White House intervention.

Next-gen speech-to-speech voice model: reasons while speaking, 1.5-2x transcription accuracy vs Deepgram Nova 3 / ElevenLabs Scribe v2. $0.08/min.

Still in partner testing as of Aug 2026. Unclear if it ships before Gemini 4 or is superseded by it.

Latest V4-Flash update, accessed via deepseek-v4-flash alias. $0.14/$0.28 per 1M tokens. Official notice: significant price increase coming soon.

CONFIRMED: most ambitious pre-training run yet. Pichai (7/23 earnings call): larger base model than Gemini 3 Pro; coding & autonomous agents are priorities. Release window (Nov/Dec 2026) is analyst inference, not official.

Efficiency-first Flash refresh: 17% fewer output tokens than 3.5 Flash, DeepSWE v1.1 49% (vs 37% for 3.5 Flash, 12% for 3.1 Pro). Post also confirmed Gemini 4 pre-training has started.

V9-Medium, 1.5T parameters, trained with Cursor data. $2/$6 per 1M tokens, 500K context. Coding-focused, agentic.

Fable 5 returns globally July 1. $10/$50 per 1M tokens. Anthropic also proposed industry-wide jailbreak severity scoring framework with Glasswing partners.

720p clips up to 15s with native audio. #1 on Image-to-Video Arena.

Opus 5 ($5/$25), Sonnet 5 ($3/$15, promo $2/$10 until 8/31), Haiku 4.5 ($1/$5).