The heartbeat of AI
Every major model, release, and breakthrough from the labs shaping AI — as a living heatmap you can actually feel. Watch the pace. Compare the players. See what shipped, and why it mattered.
2026
52 events2025
65 eventsThe timeline
Google DeepMind
Model★ LandmarkGemini 4 Argon
Google's new frontier model, state of the art on DeepSWE (77.9%); rolling out to cyber defenders first.
OpenAI
ModelGPT-6.1 Sol (DevDay 2026)
Near-Astra coding and computer use at a fifth of Astra's API price, launched at DevDay.
Anthropic
ModelClaude Sonnet 5.5
A clear upgrade over Sonnet 5: 30%+ faster and up to 30% cheaper for most work.
- Model
GPT-6 Sol & Luna
Two cheaper GPT-6 tiers that balance capability and cost for everyday work.
- Model
Claude Opus 5.5
Performs at Fable 5.1's level on most tasks at about 40% lower cost than Opus 5.
xAI
ModelGrok 4.7
SpaceXAI's new frontier coding and agent model; its Fast variant is exclusive to Cursor and Grok Build.
DeepSeek
ModelDeepSeek-V4.1-Flash
A 552B MoE with a new asymmetric encoder–decoder design (8B / 16B active) and native vision.
- Product
Agents API
A managed service for building cloud agents on the Codex harness, with long-running sessions.
Mistral AI
Funding★ LandmarkMistral raises €3B Series D
A €3B round at a valuation above €21B to push sovereign, open-weight AI to the frontier.
- Research
AlphaGenome Atlas
Predicts the molecular effects of all 9 billion possible single-letter DNA changes in the genome.
- Research★ Landmark
Claimed Navier–Stokes proof
OpenAI published an AI-generated claimed Millennium Prize solution with a Lean proof; still under review.
- Model★ Landmark
GPT-6 Astra
OpenAI's most intelligent and aligned model yet: state of the art in computer use, coding and science.
- Model
Gemini 3.8 Flash & 3.8 Flash Cyber
Next-generation Flash models for agentic workflows, plus a cybersecurity-tuned variant.
- Model★ Landmark
Claude Fable 5.1 & Mythos 5.1
Anthropic's most advanced models for coding and knowledge work, aimed at speeding up science.
- Safety
Claude text watermarking
Future Claude models will watermark generated text to comply with the EU AI Act.
- Model
DeepSeek-V4-Pro GA
V4-Pro went GA with agent upgrades, adjustable reasoning effort and Responses API support.
- Product
Ads in ChatGPT (test)
OpenAI began testing clearly labeled ads to fund free access; expanded to Europe a week later.
Qwen · Alibaba
Open weights★ LandmarkQwen3.8-Max
Alibaba's largest model (2.4T params, 95B active); its weights followed — a first for Max-class Qwen.
- Research
Gemini Robotics 2
Whole-body control for humanoid robots — walking, crouching and manipulating at once.
- Model★ Landmark
Claude Opus 5
A step-change for the Opus tier, built for long-running agents, coding and professional work.
- Product
Health in ChatGPT
US users can securely connect medical records and Apple Health for personalized insights.
Microsoft AI
ModelMAI-Image-2.5-Pro & MAI-Voice-2-Flash
Microsoft's highest-fidelity image model, plus a voice model 2x faster and 32% cheaper.
- Model
Gemini 3.6 Flash & 3.5 Flash-Lite
New Flash-tier models, plus Gemini 3.5 Flash Cyber specialized for cybersecurity work.
Meta AI
ModelMuse Spark 1.1 & Meta Model API
An updated multimodal reasoning model for agents, plus a public preview of the Meta Model API.
- Model★ Landmark
GPT-5.6
More intelligence per token and per dollar; became Microsoft 365 Copilot's preferred model at launch.
- Research
Robostral Navigate
An 8B robot-navigation model scoring 76.6% on R2R-CE from a single RGB camera, no LiDAR.
- Model
Grok 4.5
SpaceXAI's model for coding, agents and knowledge work at $2 / $6 per million tokens.
- Product
GPT-Live voice models
New natural-conversation voice models powering ChatGPT Voice; GPT-Live-1 hit the API in September.
- Model
Muse Image & Muse Video
Meta Superintelligence Labs' image and video models, with precise multi-reference editing.
- Product
Fable 5 redeployed
Fable 5 returned from July 1 with updated cyber safeguards and a new jailbreak framework.
- Model
Claude Sonnet 5
Anthropic's most agentic Sonnet yet, with top-tier coding and professional-work intelligence.
- Product
Jalapeño inference chip (with Broadcom)
A custom chip co-designed with Broadcom specifically for LLM inference at scale.
- Safety★ Landmark
US directive suspends Fable 5 access
A US export-control directive cut off Fable 5 and Mythos 5 access for all foreign nationals.
- Model★ Landmark
Claude Fable 5 & Mythos 5
First publicly available Mythos-class frontier model, state-of-the-art with new safety routing.
- Model★ Landmark
MAI-Thinking-1 & MAI-Code-1-Flash
Microsoft's first in-house reasoning and coding models, building a frontier stack independent of OpenAI.
- Product
Vibe agent platform
Le Chat relaunched as an autonomous work/code agent — Mistral's pivot to an enterprise agent OS.
- Model
Claude Opus 4.8
Flagship emphasizing honesty and reliability, adding Dynamic Workflows in Claude Code and a faster fast mode.
- Product
Gemini Spark personal agent
A 24/7 background agent wired into Gmail and Calendar — Google's most direct agentic consumer product.
- Model★ Landmark
Gemini 3.5 Flash & Gemini Omni (I/O 2026)
3.5 Flash beat 3.1 Pro at 4x speed/half cost; Omni generated video from any input modality.
- Open weights
Mistral Medium 3.5
A 128B open-weight model with 256K context and cloud agents that submit GitHub PRs autonomously.
- Open weights★ Landmark
DeepSeek-V4
First Chinese frontier model built to run natively on Huawei Ascend chips, MIT license, 1M context.
- Model★ Landmark
GPT-5.5
OpenAI's smartest and most intuitive model yet, framed as a step toward an AI super app.
- Product
8th-gen TPUs (8t & 8i)
Split-chip strategy with dedicated training and inference chips for the agentic era.
- Model
Claude Opus 4.7
Most capable public model at launch, with self-checking and stronger agentic coding.
- Safety★ Landmark
Mythos Preview & Project Glasswing
Withheld a frontier model with extreme cyber capability, giving vetted partners restricted access.
- Product
MAI-Transcribe-1, Voice-1, Image-2
Three production proprietary models for speech, voice, and image — all without OpenAI technology.
- Model
Gemini 3.1 Pro
Major agentic-reasoning leap (77.1% ARC-AGI-2, 80.6% SWE-bench Verified).
- Open weights★ Landmark
Qwen3.5
A 397B natively multimodal open-weight agent model competitive with GPT-5.2 on multimodal benchmarks.
- Model
GPT-5.3-Codex
Most capable agentic coding model to date, 25% faster while advancing reasoning.
- Model★ Landmark
Claude Opus 4.6
First Opus with a 1M-token context window, plus agent teams for enterprise work.
- Funding★ Landmark
xAI raises $20B Series E
The largest private AI round at the time, at a $230B valuation, backed by Nvidia.
- Model
Gemini 3 Flash
High-efficiency model beating Gemini 2.5 Pro at a quarter the cost, made default in the app.
- Model
GPT-5.2
Most capable model for professional knowledge work, topping GDPval across 44 occupations.
- Open weights★ Landmark
Mistral Large 3
A 675B MoE open-weight frontier model claiming parity with closed rivals, plus small models.
- Open weights★ Landmark
DeepSeek-V3.2
Flagship with integrated reasoning and sparse attention, claimed GPT-5-level at a fraction of cost.
- Model★ Landmark
Claude Opus 4.5
Flagship positioned as best in the world for coding and agents, adding an effort parameter.
- Model★ Landmark
Gemini 3 Pro
New flagship shattering benchmarks (37.4 on Humanity's Last Exam), launched across all surfaces day one.
- Model
GPT-5.1
A smarter, more conversational GPT-5 update with adaptive reasoning that adjusts thinking time.
- Research
MAI Superintelligence Team
A dedicated unit under Mustafa Suleyman targeting superhuman AI in domains like medical diagnosis.
- Model
Claude Haiku 4.5
First Haiku with extended thinking and computer use, bringing frontier features to a fast cheap tier.
- Product
DevDay 2025: AgentKit & Apps SDK
Shipped AgentKit, an Apps SDK on MCP, and GPT-5 Pro to make ChatGPT an agent-building platform.
- Model★ Landmark
Sora 2
A more physically accurate video model with synchronized audio, launched with a social app.
- Model★ Landmark
Claude Sonnet 4.5
Billed as the best coding model in the world, running autonomously for around 30 hours.
- Open weights
Qwen3-VL
Open vision-language series whose 235B flagship matched Gemini 2.5 Pro on visual benchmarks.
- Funding★ Landmark
Series C — €1.7B led by ASML
Europe's largest AI round at the time, valuing Mistral at €11.7B.
- Funding
Series F at $183B valuation
Raised $13B, roughly tripling the company valuation in six months.
- Model
MAI-Voice-1 & MAI-1-preview
Microsoft's first in-house foundation models, a step toward reducing OpenAI dependence.
- Model★ Landmark
GPT-5
Unified flagship with a real-time router between fast and deep-reasoning models for all users.
- Open weights★ Landmark
gpt-oss-120b and gpt-oss-20b
OpenAI's first open-weight models since GPT-2, Apache 2.0 and near o4-mini reasoning quality.
- Open weights
Wan 2.2
First open-source video model using a mixture-of-experts architecture, cinematic 720p on consumer GPUs.
- Product
ChatGPT agent
Unified agent combining Operator, deep research, and ChatGPT on its own virtual computer.
- Open weights
Voxtral
Mistral's first open-source speech models, beating Whisper large-v3 on multilingual audio.
- Product★ Landmark
Meta Superintelligence Labs
Consolidated all AI work under a new lab and aggressively recruited frontier talent.
- Research
Agentic Misalignment research
Showed leading models would blackmail or leak data as insider threats when cornered.
- Funding★ Landmark
Meta invests $14.3B in Scale AI
Took a 49% stake and brought on Alexandr Wang as Chief AI Officer to lead a superintelligence push.
- Model
Magistral
Mistral's first reasoning family with transparent, multilingual chain-of-thought.
- Safety
Activating ASL-3 protections
First deployment of AI Safety Level 3 safeguards under the Responsible Scaling Policy.
- Model★ Landmark
Claude Opus 4 & Sonnet 4
Launched the Claude 4 generation with frontier coding and agents, and Claude Code GA.
- Open weights
Devstral
First Mistral model built for software-engineering agents, leading open-source SWE-bench.
- Product
Project Mariner web agent
Moved from prototype to limited release, enabling autonomous multi-tab web browsing.
- Model★ Landmark
Veo 3 & Imagen 4 (I/O 2025)
Veo 3 generated synchronized dialogue and sound natively; Imagen 4 added 2K output and typography.
- Research★ Landmark
AlphaEvolve
An evolutionary coding agent that discovered faster algorithms and recovered global Google compute.
- Model
Phi-4 reasoning
Small reasoning models beating DeepSeek-R1 and o1-mini on AIME math despite far fewer parameters.
- Research
DeepSeek-Prover-V2
A 671B open theorem-prover for Lean 4 setting a new formal-math state of the art.
- Product
Meta AI app & Llama API (LlamaCon)
Meta's first standalone AI app and a developer API, unveiled at its inaugural LlamaCon.
- Open weights★ Landmark
Qwen3
Eight open-weight hybrid-reasoning models claiming parity with Google and OpenAI flagships.
- Model★ Landmark
o3 and o4-mini
Frontier reasoning models that agentically use every ChatGPT tool, setting new math/coding records.
- Model
Veo 2 (GA for developers)
Cinematic video generation via the Gemini API — physics-accurate HD clips from text or image.
- Model
GPT-4.1 family
API models with 1M-token context and major coding and instruction-following gains over GPT-4o.
- Product
Ironwood TPU (7th gen)
First TPU designed purely for inference at scale, signaling a dedicated inference era.
- Open weights★ Landmark
Llama 4 (Scout & Maverick)
First open-weight natively multimodal MoE models, with Scout's 10M-token context window.
- Funding★ Landmark
$40B round at $300B valuation
The largest private tech fundraise on record, led by SoftBank, doubling OpenAI's valuation.
- Product
xAI acquires X
An all-stock merger combining the social network's data with xAI's model development.
- Research
Tracing the Thoughts of an LLM
Landmark interpretability work mapping how Claude reasons internally via circuit tracing.
- Model
Qwen2.5-Omni
Any-to-any model handling text, images, audio, and video with real-time speech, built for edge.
- Model★ Landmark
Gemini 2.5 Pro
First thinking flagship with integrated reasoning, topping LMArena and math/science/coding.
- Open weights
DeepSeek-V3-0324
First open-weights model to outscore all proprietary non-reasoning models on Artificial Analysis.
- Open weights
Gemma 3
Multimodal, multilingual open-weight models (1B-27B) runnable on a single GPU.
- Open weights
QwQ-32B
A 32B open reasoning model rivaling DeepSeek-R1 at a fraction of the parameter count.
- Model
Phi-4 multimodal & mini
Tiny models processing speech, vision, and text, outperforming Gemini 2.0 Flash at their size.
- Open weights
Wan 2.1
Open-source text- and image-to-video model (Apache 2.0) that runs on consumer hardware.
- Model★ Landmark
Claude 3.7 Sonnet & Claude Code
First hybrid reasoning model with extended thinking, launched with the Claude Code agentic tool.
- Model★ Landmark
Grok 3
Frontier flagship on the 200k-GPU Colossus cluster, with DeepSearch and Think Mode.
- Product
Le Chat relaunch
Overhauled assistant with apps, web search, and a Pro tier — a first direct ChatGPT challenge.
- Model
Gemini 2.0 Flash (GA)
First natively multimodal-output model went GA with a 2M-token context and native tool use.
- Product
Deep research
Agentic ChatGPT feature that synthesizes hundreds of web sources into analyst-grade reports.
- Model
Qwen2.5-Max
Alibaba's largest MoE flagship, claiming to outperform DeepSeek-V3 and GPT-4o on key benchmarks.
- Open weights
Qwen2.5-VL
Open vision-language models (3B-72B) for long-video understanding and GUI agent control.
- Product
Operator
OpenAI's first computer-using agent that browses the web to complete tasks autonomously.
- Open weights★ Landmark
DeepSeek-R1
Open-source reasoning model matching o1 at a fraction of the cost, triggering a global cost-shock.
- Open weights
Phi-4 open-sourced
Released the 14B small reasoning model's weights under an MIT license for free commercial use.