2026

Vision-language-action models that directly control robots

Enhanced AI model update straining computing resources

AI predicts function of DNA sequences up to 1M base pairs

OpenAI expands access and verticals

432GB memory capacity for large model inference

336 billion transistors, 50 petaflops NVFP4 per GPU

Production-ready humanoid with all 2026 units pre-sold

Open model and data releases advancing physical AI

Exceeded $15B target with NVIDIA and Cisco as investors

Multiyear deal replaces on-device approach for cloud AI

Enhanced video generation with better consistency

Humanoid robot learns autonomously from experience

Multimodal model with 200M monthly active users on Ernie Bot

First custom inference chip with 140B transistors

1T MoE model with 100-agent swarm coordination

Ends Model S/X to build 1M humanoid robots per year

Life Biosciences ER-100 targets aging reversal in eye diseases

Combined entity valued at ~$1.2 trillion

Humanoid acquires skills from human demonstrations

Final specialist Codex before merger into mainline GPT-5.4

1M context window with parallel agent teams

744B-parameter open model trained entirely on Huawei chips

GPT-5.2-class performance powering 155M weekly users

Unitree robots achieve autonomous martial arts on world stage

Opus-level performance at one-fifth the cost

Open-weight 397B MoE with visual agentic capabilities

Four specialized AI agents debate before answering

Doubles ARC-AGI-2 reasoning score to 77.1%

AI video generation with native audio-visual synchronization

6 gigawatts of AMD GPUs in largest Mag 7 hardware deal ever

Amazon $50B, SoftBank $30B, NVIDIA $30B at $730B valuation

Two robots complete 90% of assembly work in three hours

Embodied AI systems ready for retail, manufacturing, and logistics

Unifies Codex into mainline, native computer use, 1M context

Seven new chips anchoring the Rubin generation AI platform

Two-version jump from M2.1 puts MiniMax in top-tier Chinese open-weight reasoning

Federal blueprint preempting state laws with light-touch AI regulation

First humanoid robot company to reach 10K production scale

Largest private funding round in history, with $3B from retail investors

31B dense model ranks #3 on Arena at 1452 Elo, Apache 2.0

Frontier Chinese model with 1M context and agentic coding focus

Gated research model after discovering thousands of zero-day vulnerabilities

First proprietary Meta model, non-open-weight multimodal reasoning

Stronger agentic coding and vision at unchanged pricing

SuperGrok Heavy beta with native video understanding and 1M context

300+ robots race alongside humans in world-first co-running event

Successor to K2.5 with stronger reasoning and Arena Elo

Open-weight 1T MoE matches frontier benchmarks at half the cost

First frontier model to cross 90% on ARC-AGI-2

Open-weight successor to V3, continues cost-disruption pattern

New default ChatGPT model, "strongest agentic coding model to date"

Flash-tier model with near-frontier reasoning: GPQA 92.2, ARC-AGI-1 96%

Frontier-tier Qwen flagship closes the gap to Western frontier

Successor to Opus 4.7 with stronger independent coding and agentic results

Frontier-tier open-weight successor to M2.7

First generally available Mythos-class model, positioned above Opus

Open-weight model tuned for long-horizon software engineering

Open-weight successor to GLM 5.1 with reasoning gains

First Sonnet of the Claude 5 generation; near-Opus capability at Sonnet cost

Full-sized humanoid combines local memory, emotional perception, and embodied AI

New frontier model targets coding, agentic tasks, and professional knowledge work

Sol, Terra, and Luna reach general availability with new agentic and coding frontiers

Multimodal reasoning model advances coding, computer use, and agent orchestration

Open-weight native multimodal model launches with a one-million-token context window

2.8T-parameter native-vision model brings open intelligence near the frontier

New Flash models improve agent quality and high-throughput efficiency

New Opus flagship approaches Fable-class capability at half the price

Open-weight multimodal MoE brings capable agents to a smaller active footprint

Post-training refresh lifts reasoning, coding, and agentic performance

High-risk AI system requirements become enforceable across EU

Four Qwen3.8 tiers span flagship, very large MoE, efficiency, and small open weights

Superintelligence Labs iteration lifts the Intelligence Index to 56.8

Frontier model taking the highest GPQA of the August wave

Flash-tier model ranking third overall on the Singularity Index

Open-weight pair with a large agentic gain over GLM 5.2

Efficiency-tier updates add a vision-capable variant

Mythos-level model advances long-running coding, research, and knowledge work

Pre-release model crosses OpenAI's highest cybersecurity preparedness threshold

First public riders join in Denver, San Diego, and Tampa

Coding and reasoning gains arrive alongside a restricted defensive cybersecurity variant

Intelligence Index climbs from 50.6 to 62.1 in under two months

New generation targets computer use, coding, science, and professional work

Genome-wide molecular predictions become a searchable research resource

Diffusion language model adds tunable reasoning and parallel tool calls

Asymmetric encoder-decoder architecture cuts active compute and cache requirements

Multi-model orchestration targets cost efficiency and complex agentic work

Six initial reports document unexpected model behavior

Human-behavior pretraining transfers to whole-body household tasks

Audio and video understanding model supports tool calling and text output

Verified research teams receive tailored biology safeguards

Prototype internal index reports 26% of AI R&D led by Claude under human supervision

Reasoning model enters independent evaluations with partial coverage

Evaluators will work inside the lab to assess models and safety practices

Multimodal open weights bring one-million-token context to the Flash tier

Text and vision reasoning model adds a new xhigh evaluation snapshot

Open-weight multimodal reasoning model enters independent benchmark coverage

GPT-6 expands into lower-cost coding and professional work models

New Claude generation targets agentic coding and knowledge work at lower cost

AI-led genome searches surface array-associated reverse transcriptases

Flash TTS and Flash-Lite TTS add expressive audio generation

First orbital Starship flight deploys its first operational payload

Faster, cheaper Sonnet joins the Claude 5.5 family

Near-Astra intelligence at one-fifth of the price, one week after GPT-6 Sol

First Gemini 4 model targets sustained reasoning and cyber defense

An unreleased OpenAI model posts hundreds of math results, including a claimed proof of the Unique Games Conjecture

Trillion-parameter open-weight contender from Europe enters preview

DOE, NIH, Google DeepMind, Isomorphic Labs, and Meta join the open data effort for a virtual cell

GPT-6 reaches ChatGPT users globally alongside Intelligent UI

Small, fast model completes the Claude 5.5 family at a quarter of the cost

Superintelligence Labs ships Muse Wonder and Muse Wonder Full for agents

Meta says it is the most capable US open-weight model to date