Timeline · Product milestone chronicle
106 milestones (prehistory plus five eras), statically rendered so the full text reads without JavaScript. Every entry carries an official source and a verification time.
Data version = T-CLAW-14 generation base, 106 entries (T-GENE-10 merge, 2026-11-05)Prehistory2012—2019 · 9 entries
2012-09 · Prehistory
AlexNetModel🌐 Global
University of Toronto
60M parameters
Deep learning plus GPU becomes the norm, as ImageNet top-5 error drops from 26% to 15.3%
Source ↗2017-06 · Prehistory
Transformer (Attention Is All You Need)Model🌐 Global
Google
65M/213M parameters
A pure self-attention architecture that became the shared base of both GPT and BERT
Source ↗2018-02 · Prehistory
ELMoModel🌐 Global
Allen Institute for AI / UW
94M parameters (unverified)
Contextual word representations: an early sign of the pre-training paradigm
Source ↗2018-06 · Prehistory
GPT-1Model🌐 Global
OpenAI
117M parameters
Generative pre-training plus fine-tuning is established, with SOTA on 9 of 12 tasks
Source ↗2018-10 · Prehistory
BERTModel🌐 Global
Google
340M parameters (per paper)
Bidirectional MLM pre-training makes pre-train-then-fine-tune the default NLP pipeline
Source ↗2019 · Prehistory
Cohere foundedCompany event🌐 Global
Cohere
N/A (company event); Command is its enterprise generation/retrieval model line
A Toronto-born enterprise LLM company built around retrieval augmentation and private deployment
Source ↗2019 · Prehistory
Stability AI foundedCompany event🌐 Global
Stability AI
N/A (company event); released Stable Diffusion weights in 2022-08
Its open-weights strategy set off the image-generation ecosystem; cumulative model downloads passed 350 million
Source ↗2019-02 · Prehistory
GPT-2Model🌐 Global
OpenAI
1.5B parameters (only 124M released first; full weights open-sourced 2019-11-05)
Scaling-law shock plus a governance precedent for staged releases
Source ↗2019-03 · Prehistory
ERNIE 1.0 (first-generation ERNIE)Model🇨🇳 China
Baidu
About 110M parameters (BERT-base class)
The founding work of knowledge-enhanced pre-training, which set Baidu's ERNIE technical route
Source ↗API Era2020—2022 · 11 entries
2020-06-11 · API Era
GPT-3 APIModel🌐 Global
OpenAI
175 billion parameters; few-shot in-context learning; delivered through a weight-free API
In-context learning delivered through an API for the first time, opening the models-as-a-service era
Source ↗2020-11-30 · API Era
AlphaFold 2Model🌐 Global
Google DeepMind
CASP14 median GDT 92.4
Cracks a 50-year protein-folding problem: the first time AI solves a major scientific challenge
Source ↗2021-06-29 · API Era
GitHub CopilotProduct🌐 Global
GitHub (Microsoft)
Technical preview → GA (2022-06, $10 per month for individuals)
The first scaled AI pair programmer, bringing coding into the autocomplete era
Source ↗2021-11-08 · API Era
M6-10T multimodal large modelModel🇨🇳 China
Alibaba DAMO Academy
10 trillion parameters in a multimodal MoE, using about 1% of GPT-3's energy
The world's first 10-trillion-parameter-class multimodal pre-trained model, the largest of its kind at the time
Source ↗2021-12 · API Era
ERNIE 3.0 TitanModel🇨🇳 China
Baidu
260 billion parameters
At the time the largest Chinese knowledge-enhanced large model in the world
Source ↗2021-12 · API Era
MiniMax foundedCompany event🇨🇳 China
MiniMax
Yan Junjie leaves SenseTime to found the company, among the earliest general-AI entrants
Source ↗2022-04-06 · API Era
DALL·E 2Model🌐 Global
OpenAI
1024×1024 text-to-image with inpainting
Text-to-image climbs to photo realism: the start of generative vision
Source ↗2022-07 · API Era
Midjourney opens its betaProduct🌐 Global
Midjourney
Open beta, operated entirely through Discord
Text-to-image reaches mainstream consumers, yielding huge annual revenue with zero funding
Source ↗2022-08 · API Era
GLM-130B open-sourcedModel🇨🇳 China
Tsinghua University & Zhipu AI
130 billion parameters, a bilingual Chinese-English dense model
The first Chinese 100-billion-class bilingual dense model released openly, placing it in the global open-model front tier
Source ↗2022-08 · API Era
ModelBest foundedCompany event🇨🇳 China
ModelBest
Spun out of Tsinghua's NLP lab, it established the high-efficiency on-device “small but mighty” route
Source ↗2022-08-22 · API Era
Stable Diffusion open-sourcedModel🌐 Global
Stability AI
Open weights (SD1.4, about 890 million parameters)
Open text-to-image ignites the global AIGC creation wave
Source ↗ChatGPT EraFrom 2022.11 · 5 entries
2022-11-30 · ChatGPT Era
ChatGPTProduct🌐 Global
OpenAI
GPT-3.5; 100 million monthly actives two months after launch (UBS estimate)
Conversational AI reaches the general public: the starting point of the generative AI era
Source ↗2022-12-07 · ChatGPT Era
Perplexity AskProduct🌐 Global
Perplexity
Retrieval-augmented answers with live citations
The founding AI answer engine, whose live citations challenge traditional search
Source ↗2023-02-20 · ChatGPT Era
MOSS opens to the publicModel🇨🇳 China
Fudan University (Qiu Xipeng's team)
16 billion parameters
China's first ChatGPT-like dialogue model; its servers were overwhelmed on launch night
Source ↗2023-03-14 · ChatGPT Era
ChatGLM-6B open-sourcedModel🇨🇳 China
Tsinghua KEG & Zhipu AI
6 billion parameters; deployable in about 6 GB of VRAM after INT4 quantization
A Chinese open dialogue model that runs on consumer GPUs, setting off a local-deployment wave
Source ↗2023-03-16 · ChatGPT Era
ERNIE Bot launchedProduct🇨🇳 China
Baidu
The first launch event by a global major to take on ChatGPT, starting the Chinese LLM race
Source ↗Multimodal Era2023 · 25 entries
2023-01-23 · Multimodal Era
Microsoft invests further in OpenAICompany event🌐 Global
Microsoft
Multi-year, multi-billion-dollar commitment (reported $10B); Azure as exclusive cloud
A hyperscaler places its bet as the ChatGPT shockwave is fully capitalized
Source ↗2023-02-07 · Multimodal Era
New Bing (Bing Chat)Product🌐 Global
Microsoft
Driven by a custom GPT-4 and deeply integrated with Edge
The search war reopens, forcing every giant to respond to generative AI
Source ↗2023-03 · Multimodal Era
Runway Gen video generationProduct🌐 Global
Runway
Gen-2 generates 3-4 second clips from text or images behind a waitlist; later iterated to Gen-3/Gen-4
The first text-to-video system open to the public; video generation becomes usable
Source ↗2023-03-14 · Multimodal Era
GPT-4Model🌐 Global
OpenAI
8K/32K context; image input; parameters undisclosed (a rumoured 1.8T MoE remains unconfirmed)
Bar-exam-level tests become its benchmark, making the multimodal flagship the yardstick
Source ↗2023-03-24 · Multimodal Era
Baichuan AI foundedCompany event🇨🇳 China
Baichuan AI
N/A (company event); shipped Baichuan-7B/13B as open, commercially usable models within 100 days
Founded by Wang Xiaochuan with open, commercially usable Chinese base models, joining the domestic front tier
Source ↗2023-04 · Multimodal Era
Moonshot AI foundedCompany event🇨🇳 China
Moonshot AI
Founded by Yang Zhilin, it raised over $1 billion within its first year
Source ↗2023-04 · Multimodal Era
Mistral AI foundedCompany event🌐 Global
Mistral AI
N/A (company event, no model spec); the company later entered the market with the open Mistral 7B
A European frontier lab founded to break the US closed-model monopoly with open weights and strong price-performance
Source ↗2023-04-06 · Multimodal Era
StepFun foundedCompany event🇨🇳 China
StepFun
Founded by former Microsoft global VP Jiang Daxin, betting on multimodality and scaling laws
Source ↗2023-04-07 · Multimodal Era
Qwen begins enterprise invite-only testingProduct🇨🇳 China
Alibaba
Alibaba's first consumer-facing LLM, which grew into the largest open-model family in the world
Source ↗2023-04-10 · Multimodal Era
SenseTime SenseNova launchedModel🇨🇳 China
SenseTime
A multi-model suite spanning language, text-to-image and 2D/3D digital humans; SenseCore with 27,000 GPUs and 5000 PFLOPS
The “large models plus large compute” strategy lands, with one of China's first published multimodal model suites
Source ↗2023-05 · Multimodal Era
01.AI founded and the Yi series open-sourcedCompany event🇨🇳 China
01.AI
N/A (company event); Yi series 6B/34B open weights
Founded by Kai-Fu Lee, a unicorn within six months, with the open Yi series taking on Llama 2
Source ↗2023-05-06 · Multimodal Era
iFlytek Spark cognitive model launchedModel🇨🇳 China
iFlytek
Spark V1.0 with seven core capabilities including text generation, language understanding and knowledge Q&A
The first fully in-house Chinese cognitive model, matching ChatGPT on three capabilities including Chinese text generation
Source ↗2023-07 · Multimodal Era
DeepSeek foundedCompany event🇨🇳 China
DeepSeek (incubated by High-Flyer)
Incubated by quant giant High-Flyer, with Liang Wenfeng leading its entry into large models
Source ↗2023-07-11 · Multimodal Era
Claude 2Model🌐 Global
Anthropic
100K context; free public beta on the web
A 100K-token context window opens to the public for the first time
Source ↗2023-07-13 · Multimodal Era
Interim Measures for the Management of Generative AI Services publishedCompany event🇨🇳 China
CAC and six other ministries
Effective 2023-08-15
Among the world's earliest dedicated generative AI rules, establishing the filing regime
Source ↗2023-07-18 · Multimodal Era
Llama 2 opens for commercial useModel🌐 Global
Meta
7B/13B/70B, free for research and commercial use under the licence terms
Commercially usable open weights: the watershed moment for open LLMs
Source ↗2023-08-31 · Multimodal Era
The first batch of large models approved for public accessCompany event🇨🇳 China
Baidu / Zhipu / SenseTime / Baichuan / ByteDance and 5 more
Products from the first eight companies opened to the public
Chinese LLMs are cleared for consumers, opening the first year of universal access
Source ↗2023-09-07 · Multimodal Era
Tencent Hunyuan debutsModel🇨🇳 China
Tencent
Over 100 billion parameters; over 2 trillion pre-training tokens
Tencent's in-house model debuts, deployed through its internal business scenarios
Source ↗2023-10-09 · Multimodal Era
Kimi assistant launchedProduct🇨🇳 China
Moonshot AI
A lossless 200,000-character context
It enters the consumer market through long-context differentiation and breaks out within six months
Source ↗2023-10-17 · Multimodal Era
ERNIE 4.0 releasedModel🇨🇳 China
Baidu
Officially described as on par with GPT-4 — the first Chinese flagship to match it
Source ↗2023-11 · Multimodal Era
Amazon Q launchedProduct🌐 Global
AWS
Q Business/Q Developer; 2024-04-30 GA
A hyperscaler enters the generative AI main stage with its own enterprise assistant
Source ↗2023-11-01 · Multimodal Era
Microsoft 365 Copilot GAProduct🌐 Global
Microsoft
$30 per user per month
Generative AI joins the office suite with explicit per-seat pricing
Source ↗2023-11-24 · Multimodal Era
Qwen-72B open-sourcedModel🇨🇳 China
Alibaba
72 billion parameters
One of the largest Chinese open models of its time, it quickly topped several open leaderboards
Source ↗2023-12 · Multimodal Era
Suno AI music generationProduct🌐 Global
Suno
The web app generates complete songs with vocals (about 2 minutes); v3 fully opened in 2024-03
One-click songs from text, integrated into Microsoft Copilot, bringing music generation to the masses
Source ↗2023-12-06 · Multimodal Era
Gemini 1Model🌐 Global
Google
Ultra scores 90.0% on MMLU (CoT@32); natively multimodal
A natively multimodal flagship that let Google strike back head-on for the first time
Source ↗Agent Reasoning Era2024—2025 · 33 entries
2024-02 · Agent Reasoning Era
MiniCPM-2B open-sourcedModel🇨🇳 China
ModelBest
2.4B parameters, deployable on-device
The “small but mighty” line validates the high-density small-model route, and on-device AI takes off
Source ↗2024-02-01 · Agent Reasoning Era
Coze launches in ChinaProduct🇨🇳 China
ByteDance
A low-code / no-code bot development platform
It brings “agents for everyone” to China and accelerates ByteDance's app ecosystem
Source ↗2024-02-15 · Agent Reasoning Era
Sora announcedModel🌐 Global
OpenAI
Text-to-video up to 60 seconds at 1080p
A leap in text-to-video that ignites the world-simulator contest
Source ↗2024-03 · Agent Reasoning Era
Groq LPU inference engine and GroqCloudInfrastructure🌐 Global
Groq
LPU Inference Engine and the GroqCloud self-serve inference platform; officially measured at about 300 tokens/s for Llama-2 70B
Its deterministic LPU architecture pushes open-model inference to hundreds of tokens per second, reshaping the inference cost curve
Source ↗2024-03-18 · Agent Reasoning Era
Kimi supports a lossless 2-million-character contextProduct🇨🇳 China
Moonshot AI
A lossless 2-million-character context
It sets off a long-context arms race, and demand overloads Kimi's servers
Source ↗2024-03-23 · Agent Reasoning Era
Step series models releasedModel🇨🇳 China
StepFun
Step-1/1V at 100 billion parameters, with Step-2 previewed as a trillion-parameter MoE
China's first publicly stated trillion-parameter MoE route, going all in on multimodality
Source ↗2024-05-06 · Agent Reasoning Era
DeepSeek-V2 releasedModel🇨🇳 China
DeepSeek
MoE with 236B total and 21B active parameters; API input at 1 yuan per million tokens
Extreme price-performance undercuts industry pricing and ignites the LLM price war
Source ↗2024-05-13 · Agent Reasoning Era
GPT-4oModel🌐 Global
OpenAI
Average audio response of 320 ms; free tier opened
Real-time omni-modal interaction reaches the public for free for the first time
Source ↗2024-05-15 · Agent Reasoning Era
Doubao models releasedModel🇨🇳 China
ByteDance (Volcano Engine)
Doubao general pro-32k input at 0.0008 yuan per 1K tokens (about 1/150 of comparable specs)
ByteDance enters with a price war, bringing “cent-era” pricing to large models
Source ↗2024-05-30 · Agent Reasoning Era
Tencent Yuanbao launchedProduct🇨🇳 China
Tencent
A consumer AI assistant built on Hunyuan, drawing on WeChat ecosystem sources
Tencent's consumer AI flagship lands as big-tech assistant competition heats up
Source ↗2024-09-12 · Agent Reasoning Era
o1Model🌐 Global
OpenAI
AIME 74% (GPT-4o: 13%)
Trading inference-time compute for intelligence: the start of the reasoning-model paradigm
Source ↗2024-10-22 · Agent Reasoning Era
Computer UseProduct🌐 Global
Anthropic
A screenshot-click-type loop, in public beta with 3.5 Sonnet
A frontier model is officially given the ability to operate a computer for the first time
Source ↗2024-10-25 · Agent Reasoning Era
AutoGLM releasedProduct🇨🇳 China
Zhipu AI
Phone-use capability with autonomous control of a real device
The first real-device phone-use agent, opening the on-device agent chapter
Source ↗2024-11-25 · Agent Reasoning Era
Model Context Protocol(MCP)Infrastructure🌐 Global
Anthropic
An open protocol unifying model-to-data and model-to-tool connections
An open standard for connecting AI to the outside world (the USB-C of AI)
Source ↗2024-12-03 · Agent Reasoning Era
Amazon NovaModel🌐 Global
Amazon
Micro/Lite/Pro GA the same day; Premier in 2025Q1
AWS rounds out its in-house model family with a price-performance focus
Source ↗2024-12-26 · Agent Reasoning Era
DeepSeek-V3 releasedModel🇨🇳 China
DeepSeek
MoE with 671B total and 37B active parameters; 14.8T tokens; about $5.576 million in GPU training cost
It matches GPT-4o-class performance at roughly one-tenth the training cost, rewriting training economics
Source ↗2025-01-21 · Agent Reasoning Era
Stargate programInfrastructure🌐 Global
OpenAI/SoftBank/Oracle/MGX
$500 billion over four years; first site in Abilene, Texas
The largest AI compute infrastructure program in history is launched
Source ↗2025-02-17 · Agent Reasoning Era
Grok-3Model🌐 Global
xAI
Trained on the Colossus supercomputer at roughly 10 times the previous generation's compute; Think and DeepSearch modes
Its build-your-own-supercomputer route reaches the front tier for the first time
Source ↗2025-04-09 · Agent Reasoning Era
Google ADK + A2AInfrastructure🌐 Global
Google
An open-source Python framework, released alongside the A2A protocol
Google standardizes agent development and interoperability
Source ↗2025-04-16 · Agent Reasoning Era
o3Model🌐 Global
OpenAI
ARC-AGI-1 87.5% (high-compute tier, 2024-12 preview score)
A reasoning model clears ARC-AGI, a human-like abstraction benchmark
Source ↗2025-07 · Agent Reasoning Era
KiroProduct🌐 Global
AWS
Spec-driven workflows plus agent hooks on a VS Code base
Spec-driven development enters the agentic IDE race
Source ↗2025-07-09 · Agent Reasoning Era
Grok-4Model🌐 Global
xAI
Grok 4 Heavy multi-agent; HLE 44.4% (with tools, xAI figure)
Multi-agent concurrent reasoning tops the Humanity's Last Exam benchmark
Source ↗2025-07-17 · Agent Reasoning Era
Lovable becomes a unicornCompany event🌐 Global
Lovable
A $200 million Series A at a $1.8 billion valuation, eight months after launch
Vibe coding reaches the masses and it becomes one of the fastest unicorns ever
Source ↗2025-08-06 · Agent Reasoning Era
Jules general availabilityProduct🌐 Global
Google
An isolated VM per task; previewed 2024-12, public beta 2025-05
Google turns the coding agent into an asynchronous back-office teammate
Source ↗2025-08-07 · Agent Reasoning Era
GPT-5Model🌐 Global
OpenAI
Unified routing that switches between chat and reasoning automatically; open to all
Reasoning folds seamlessly into the main flagship, available to everyone
Source ↗2025-09 · Agent Reasoning Era
Replit Agent 3Product🌐 Global
Replit
Up to 200 minutes of autonomous operation; writes its own tests and fixes itself
A long-running autonomous full-stack development agent reaches consumer pricing
Source ↗2025-10-01 · Agent Reasoning Era
Microsoft Agent Framework(MAF)Infrastructure🌐 Global
Microsoft
Unifies Semantic Kernel and AutoGen in an open-source preview
Microsoft consolidates its two frameworks as the enterprise agent stack takes shape
Source ↗2025-10-06 · Agent Reasoning Era
AgentKitProduct🌐 Global
OpenAI
Agent Builder/ChatKit/Evals
Agent building moves from writing code to visual assembly
Source ↗2025-10-28 · Agent Reasoning Era
OpenAI restructuring (PBC)Company event🌐 Global
OpenAI
OpenAI Group PBC; Microsoft holds about 27% (about $135 billion)
Its capital structure is settled and the Microsoft relationship is repriced
Source ↗2025-10-29 · Agent Reasoning Era
Cursor 2.0 and ComposerProduct🌐 Global
Anysphere (Cursor)
The in-house Composer model, with eight agents in parallel
The AI-native IDE enters the multi-agent parallel era
Source ↗2025-11 · Agent Reasoning Era
Microsoft Agent 365Product🌐 Global
Microsoft
A control plane for registration, permissions and observability; GA 2026-05-01 at $15 per user per month
Managing agents becomes managing digital employees
Source ↗2025-11-18 · Agent Reasoning Era
Gemini 3Model🌐 Global
Google
Gemini 3 Pro preview with multi-step reasoning and agentic ability
Google's flagship returns to the front tier with native agent capabilities
Source ↗2025-11-24 · Agent Reasoning Era
Claude Opus 4.5Model🌐 Global
Anthropic
SWE-bench Verified 80.9%, the first in the industry past 80%
The first model past 80% on a real-world coding benchmark, with a price cut
Source ↗Open-Source Era2025—2026 · 23 entries
2025-01-20 · Open-Source Era
DeepSeek-R1 releasedModel🇨🇳 China
DeepSeek
Open-sourced under MIT; trained purely with reinforcement learning; API priced at about 1/27 of o1
An open reasoning model matches o1 for the first time, shaking the closed-model high-compute narrative
Source ↗2025-01-27 · Open-Source Era
The DeepSeek shockwaveCompany event🇨🇳 China
DeepSeek / global market
Nvidia falls 16.97% in a single day, shedding about $589 billion in market value
DeepSeek tops the US App Store as global capital reprices Chinese AI
Source ↗2025-02-10 · Open-Source Era
Qwen has the most derivative models in the worldCompany event🇨🇳 China
Alibaba
More than 90,000 Qwen-derived models, overtaking Llama
Qwen overtakes Llama to become the largest open-model family in the world
Source ↗2025-02-18 · Open-Source Era
Step-Video-T2V open-sourcedModel🇨🇳 China
StepFun & Geely Auto
Text-to-video, and the largest open-source model by parameters at release
An automaker and a model lab open-source together, filling the open-source gap in video generation
Source ↗2025-02-24 · Open-Source Era
Alibaba commits 380 billion yuan to AI over three yearsInfrastructure🇨🇳 China
Alibaba
Over 380 billion yuan for cloud and AI hardware infrastructure over the next three years
The largest AI compute investment by a Chinese private company, exceeding the previous decade combined
Source ↗2025-03-06 · Open-Source Era
Manus launchedProduct🇨🇳 China
Butterfly Effect (Monica team)
SOTA on the GAIA benchmark
The first general-purpose AI Agent to break out, when invite codes were almost impossible to get
Source ↗2025-04-05 · Open-Source Era
Llama 4(Scout/Maverick)Model🌐 Global
Meta
MoE; Scout offers a 10-million-token context
MoE plus a ten-million-token context in open weights: open source closes in on the frontier
Source ↗2025-04-29 · Open-Source Era
Qwen3 open-sourcedModel🇨🇳 China
Alibaba
0.6B–235B (A22B) MoE under Apache 2.0, with hybrid reasoning
A flagship-class hybrid reasoning model released open at every size, topping Hugging Face within a day
Source ↗2025-06-30 · Open-Source Era
ERNIE 4.5 series open-sourcedModel🇨🇳 China
Baidu
Ten models, headed by ERNIE-4.5-VL-424B-A47B, all under Apache 2.0
Baidu opens its core flagship for the first time as big tech pivots to open source
Source ↗2025-07-11 · Open-Source Era
Kimi K2 open-sourcedModel🇨🇳 China
Moonshot AI
MoE with 1 trillion total and 32 billion active parameters, open-sourced
China's first trillion-parameter open model, with agentic ability aligned to the frontier
Source ↗2025-07-28 · Open-Source Era
GLM-4.5 open-sourcedModel🇨🇳 China
Zhipu AI
MoE with 355B total and 32B active parameters under MIT
An agentic open flagship that tops the open leaderboards for coding and agentic tasks
Source ↗2025-08-05 · Open-Source Era
gpt-oss open weightsModel🌐 Global
OpenAI
120B/20B MoE; Apache 2.0
The first open weights since GPT-2, bringing flagship capability downmarket
Source ↗2025-10 · Open-Source Era
MiniMax M2 open-sourcedModel🇨🇳 China
MiniMax
MoE with 230B total and 10B active parameters, with interleaved thinking
10B-class active parameters matching Claude-level coding, a price-performance benchmark
Source ↗2025-11-06 · Open-Source Era
Kimi K2 Thinking open-sourcedModel🇨🇳 China
Moonshot AI
1 trillion total and 32 billion active parameters; about 3,000 consecutive tool calls
An open reasoning model benchmarked against GPT-5, praised by Hugging Face's founder
Source ↗2025-12-31 · Open-Source Era
Dola passes 10 million daily activesProduct🇨🇳 China
ByteDance (global AI assistant)
More than 10 million daily active users
ByteDance's AI assistant scales overseas with chat, writing, translation and image tools
Source ↗2026-01-09 · Open-Source Era
MiniMax lists in Hong KongCompany event🇨🇳 China
MiniMax
Priced at HK$165, up 109.09% on day one, for a market cap above HK$100 billion
The first IPO by a Chinese LLM company, breaking the ice for AI asset securitization
Source ↗2026-02-11 · Open-Source Era
GLM-5 releasedModel🇨🇳 China
Zhipu AI
744B total and 40B active parameters with DSA sparse attention, open-sourced
A new SOTA in open coding and agentic ability, taking on closed flagships head-on
Source ↗2026-02-16 · Open-Source Era
Qwen3.5 released and open-sourcedModel🇨🇳 China
Alibaba
Qwen3.5-Plus launched on Lunar New Year's Eve, with 397B-A17B weights open-sourced on 24 February
A native multimodal and agent foundation: Qwen moves toward native agents
Source ↗2026-04 · Open-Source Era
Claude Managed AgentsProduct🌐 Global
Anthropic
A managed agent runtime in public beta, with statefulness and tool calling
Enterprise agents move from framework orchestration to managed runtimes
Source ↗2026-04-24 · Open-Source Era
DeepSeek-V4 preview releasedModel🇨🇳 China
DeepSeek
Pro and Flash editions with a 1M context
Focused on code intelligence and production-grade reasoning, with V4-Pro following in August
Source ↗2026-07-24 · Open-Source Era
Claude Opus 5Model🌐 Global
Anthropic
Fifth-generation Opus, strengthened for agentic coding and long tasks
Anthropic's flagship enters its fifth generation, going deeper on long tasks
Source ↗2026-09-03 · Open-Source Era
GPT-6(Astra)Model🌐 Global
OpenAI
Astra at $10/$50 per million tokens; Sol and Luna added on 09-22
The GPT-6 generation opens, focused on lower prices and fewer errors
Source ↗2026-09-30 · Open-Source Era
Gemini 4(Argon)Model🌐 Global
Google
Argon is the launch model, with a 1M-token context and a staged rollout
The Gemini 4 generation begins as agentic engineering lands
Source ↗