AI Model Releases Timeline

Z.ai releases GLM-5.2, which Artificial Analysis rates the leading open-weights model on its Intelligence Index v4.1 at 51.

744B total, 40B active open-weights

leading open-weights model on the Artificial Analysis Intelligence Index v4.1, scoring 51 against MiniMax-M3 at 44, DeepSeek V4 Pro at 44 and Kimi K2.6 at 43; same size as GLM-5.1 but 11 points higher (Artificial Analysis)

GPT-5.6 Sol, Terra and Luna go public.

proprietary

Kimi K3 takes the number one spot on the arena leaderboard from Claude and GPT.

Google's Gemini 3.5 is reported delayed while headlines claim the opposite.

DeepSeek releases V4-Flash-0731, pushing Terminal Bench 2.1 from 61.8 to 82.7 at unchanged prices.

Terminal Bench 2.1 at 82.7, up from 61.8, at unchanged prices (DeepSeek)

Alibaba announces Qwen3.8-Max, 2.4 trillion parameters, with open weights promised within a week.

2.4T total, 95B active per token

first Qwen-Max-class model to get open weights (Alibaba)

Meta launches Muse Code, with a cheap tier that trains on user prompts and code.

Meta releases Muse Glimmer, a 30B open-weights model that runs on a single consumer GPU.

30B Apache-2.0

DeepMind opens WeatherNext, which beats existing cyclone forecasts by a full day.

cyclone track and intensity a full day further out than prior systems (Nature paper, DeepMind)

OpenAI ships GPT-5.6-Cyber, a security-specialized model that answers 95 percent of the questions its general models refuse.

proprietary

answers 95% of advanced security queries against about 2% for Daybreak Blue (OpenAI)

DeepMind releases SL2T, a sign language translation model shipping in the Pixel 11 keyboard.

DeepSeek ships V4 Pro and raises API prices up to 12x with peak-hour billing.

peak output tokens from $0.87 to $3.96 per million, cache-hit input from $0.003625 to $0.044 (DeepSeek)

OpenAI previews Ultrafast, running GPT-5.6 Sol at up to 750 tokens per second.

up to 750 output tokens per second, up to 14x Standard, on Cerebras hardware (OpenAI)

Google releases Gemini 3.7 Flash three weeks after 3.6, while Gemini 3.5 Pro stays delayed.

1,000,000-token context proprietary

FrontierCode 43.6% (from 34.4%), DeepSWE 65.3% (from 49.0%); $0.75/$3.75 per million until 2027-01-01 (Google)

Nvidia's MotionBricks animates game characters and a humanoid robot from one model trained on 350,000 clips.

one controller trained on 350,000 mocap clips, 15,000 FPS at 2 ms latency (Nvidia Research)

Anthropic's risk report discloses an unreleased Model 2 more capable than Mythos 5.

"somewhat more capable than Mythos 5", no plans to release externally (Anthropic risk report)

OpenAI pauses frontier RL training for two weeks and keeps its largest run on hold.

RL training on newest deployment-bound models paused two weeks; largest frontier run on hold (OpenAI)

Claude, working with mathematicians, finds a rank 30 elliptic curve, the first improvement on the 2024 record; a second improvement follows within the week.

rank 30 then rank 31 elliptic curve, with Levent Alpoge and Ava Howell; Epoch AI marked the problem "Solved (AI)"

A model listed only as Ox Alpha appears on OpenRouter with no lab attached, a 1,048,576-token context window and near-unlimited free usage.

1,048,576-token context

near 80% on a 10-task DeepSWE subset in one independent test; no lab attached

Z.ai delays GLM-5.3 open weights by two weeks after the model finds 2,436 flaws across 269 open-source projects, 1,097 of them medium to high severity.

2,436 flaws across 269 open-source projects, 1,097 medium to high severity (Z.ai testing)

OpenAI cuts GPT-5.6 Sol API and credit pricing by over 20 percent for three months.

API and credit pricing cut over 20% for three months; $4/$20 per million, down from $5/$30, through 2026-11-21 (OpenAI)

Anthropic ships a browser inside Claude Cowork: the desktop app opens a side panel and navigates, clicks and fills forms for Pro, Max and Team plans on macOS and Windows.

Alibaba open-weights Qwen3.8-Flash-Next: 176 billion parameters (125B model plus a 51B N-gram lookup) with only about 6B active per token.

176B total (125B model plus 51B N-gram lookup), about 6B active per token 262,144-token context

Instinct, an AI assistant reached by text or phone call that acts through connected apps, is valued at $2.5 billion while still not publicly available.

$250M Series B, $350M total at a $2.5B valuation, still private beta (TechCrunch)

Anthropic opens the research preview of the Model Hardware Standard, letting agents discover and operate lab equipment; QuEra laser stabilization went from 58% to 99.3%.

QuEra laser stabilization from 58% to 99.3% (Anthropic)

fal ships MiniMax H3 Max, rendering 5-second audio-video clips in under 3 seconds, faster than playback.

5-second 768p clip with audio in under 3 seconds, about 35x the official MiniMax endpoint (fal)

OpenAI retires the original DALL-E GPT from ChatGPT; ChatGPT Images takes over.

GPT-5.6 breaks the 2018 record on large gaps between primes set by Ford, Green, Konyagin, Maynard and Tao, per Oxford number theorist Jared Lichtman.

improved the 2018 Ford-Green-Konyagin-Maynard-Tao bound on large prime gaps, reported by Jared Lichtman; no proof check yet

Dyson launches CameraJet, its first toothbrush: a $499 device with a 100,000-pixel camera at 28fps and AI that fires mouth rinse at gaps between teeth.

$499 device, 100,000-pixel camera at 28 images per second, ML trained on about 470,000 dental images (Dyson)

Anthropic launches Claude Fable 5.1 and Mythos 5.1: 55.8% on Terminal-Bench 4.0, cache reads 75% cheaper; Fable models run on usage credits for Pro subscribers.

proprietary

Fable 5.1 at 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5; cache reads 75% cheaper (Anthropic)

OpenAI designates Astra the first model at its Critical cybersecurity threshold and says it will release it with limited cyber access via Daybreak Blue.

first model designated at the Critical cybersecurity threshold of OpenAI's Preparedness Framework (OpenAI)

Google releases Gemini 3.8 Flash, third Flash in six weeks, ahead of Opus 5 on Google's tables.

Meta releases Muse Spark 1.3, reporting 75.4% on DeepSWE v1.1.

75.4% on DeepSWE v1.1 (Meta)

OpenAI launches GPT-6 Astra, scoring 99.9% on ARC-AGI-3 and 100% on ExploitBench; rollout to all paid ChatGPT tiers, API, Azure and AWS Bedrock within days.

proprietary

99.9% on ARC-AGI-3 and 100% on ExploitBench (OpenAI)

Google DeepMind launches WeatherNext 3, the first global weather model to produce a new forecast every hour from live satellite data, at 5 km resolution; rolling into Search, Maps and Gemini.

first global weather model with hourly forecasts, 5km resolution against a 25km grid (Google DeepMind)

Meta launches Muse, a personal AI agent that reads email, books travel, negotiates and pays with the user's saved card; US on iOS, Android and web, free tier plus $20 and $100 plans.

OpenAI fixes three problems behind GPT-6 Astra's quality drop (old skills blocking self-checks, an opt-in context experiment hitting about 4,000 to 5,000 users, misconfigured engines) and resets usage limits.

Anthropic releases Claude Opus 5.5 at $4/$20 per million tokens, 20% below Opus 5, raises five-hour usage limits on paid plans and gives subscribers a limit reset.

40% less than Opus 5 on typical workloads (Anthropic)

OpenAI releases GPT-6 Sol ($2/$10 per million tokens) and GPT-6 Luna ($0.10/$0.50), half the GPT-5.6 prices, and tells VentureBeat the prices are permanent.