Honest plain-English takes on AI releases that matter
What changed
Filter
↓ Newest first · 59 changes
APRIL 2026
3 changes-
OpenAI launched GPT-5.5
OpenAI's spring update to GPT-5. 1-million-token context for everyone (not just Pro), better video understanding, and stronger agent reliability. The kind of release that doesn't make headlines but quiet…
-
AI tutor data came in — Khanmigo's first real-world results · era-defining
After two years of pilots, Khan Academy's AI tutor Khanmigo reached 1.5 million students in 130 countries. The 2025-2026 randomised trial data showed a measurable but modest jump in maths scores — 12% im…
-
Meta released Llama 5 · era-defining
Meta's open-weights flagship got a generational upgrade. 600B+ parameter mixture-of-experts, 5-million-token context window, native video and audio, and a community licence that finally got friendlier fo…
MARCH 2026
1 changeFEBRUARY 2026
3 changes-
Google launched Gemini 3.1 Pro (Deep Think mode)
Google's first big update to Gemini 3. New 'Deep Think' mode for hard problems — minutes-long reasoning runs with visible thought traces. 2-million-token context window for Pro users, double the Gemini 3…
-
Anthropic launched Claude Sonnet 4.6
Anthropic's mid-tier flagship got long-context and faster mode. 1-million-token context window finally on Sonnet, plus a 'fast mode' for everyday work that drops the thinking step when not needed. Same p…
-
OpenClaw — open-source AI agent framework went mainstream · era-defining
OpenClaw — an open-source framework for building autonomous AI agents — went from a side project (Nov 2025) to over 250,000 GitHub stars by early 2026. In February 2026, OpenAI hired the founder and star…
DECEMBER 2025
2 changes-
OpenAI launched GPT-5.2 (omnimodal)
OpenAI's December model bump made GPT-5 properly omnimodal — text, images, audio, video all handled by the same model with native generation in each. Expanded memory window to 1 million tokens for ChatGP…
-
AI-designed drugs entered human trials · era-defining
DeepMind's protein-prediction breakthroughs reached real patients. Isomorphic Labs — the DeepMind drug-discovery spinoff — moved its first AI-designed cancer and immunology drugs into human trials. The f…
NOVEMBER 2025
4 changes-
Anthropic launched Claude Opus 4.5
Anthropic's top tier got better and cheaper. Opus 4.5 took the leadership spot back from Gemini 3 within a week, with a meaningful price cut versus Opus 4. Aimed at the most demanding multi-step agent wo…
-
Google launched Gemini 3 · era-defining
Google's third Gemini generation. Reasoning across text, code, images, and video as a single model. Strong developer support out of the gate, and the first time Google's flagship genuinely felt like the …
-
Music labels stopped suing AI music makers — started licensing them · era-defining
After 18 months of suing AI music startups Suno and Udio for copyright infringement, Universal and Warner Music settled and signed licensing deals in October-November 2025. The 'sue them out of business'…
-
Self-driving taxis went mainstream in 10 US cities · era-defining
Waymo's fully self-driving robotaxi service expanded to 10 US cities with about 3,000 vehicles and 500,000 paid rides per week by early 2026 — up from ~50,000 in May 2024. Phoenix, San Francisco, LA, Aus…
OCTOBER 2025
4 changes-
Cursor 2.0 launched with Composer model and multi-agent UI
The popular AI code editor shipped a major rebuild. Cursor's own AI model (Composer), an interface where up to 8 agents work in parallel, and smarter codebase search. Optimised for high-speed code genera…
-
OpenAI completed its for-profit recapitalization · era-defining
After years of legal and political wrangling, OpenAI shifted from a capped-profit subsidiary of a nonprofit to a hybrid: a for-profit Public Benefit Corporation owned about 26% by the OpenAI Foundation. …
-
OpenAI launched ChatGPT Atlas · era-defining
OpenAI shipped Atlas — a web browser with ChatGPT built into the browsing experience. Sidebar AI assistant, optional browser memory, and an 'agent mode' where ChatGPT clicks around the web for you. macOS…
-
Anthropic launched Claude Skills · era-defining
Anthropic shipped Skills — packages of instructions, scripts, and templates that teach Claude how to do specific work. Like apps for an AI. Quickly became an open standard with Notion, Canva, Atlassian a…
SEPTEMBER 2025
4 changes-
OpenAI launched Sora 2 · era-defining
OpenAI's video AI grew up. Sora 2 generates 10-25 second clips with synchronised dialogue, sound effects, and physics that mostly behave. Came with a TikTok-style social app where you can drop your face …
-
Anthropic launched Claude Sonnet 4.5
Anthropic's mid-tier Claude got better at coding — meaningful jumps on practical engineering tests, computer-use tasks, and long agent runs. Same price as Sonnet 4, much stronger output. Took the coding …
-
AI took control of a fusion reactor · era-defining
DeepMind and Switzerland's EPFL handed plasma control of a real fusion reactor — the TCV tokamak — to a deep-learning AI. It held shapes that earlier control systems struggled with, including configurati…
-
Humanoid robots started doing real factory work · era-defining
Tesla put 1,000+ Optimus robots to work at Gigafactory Texas. Figure AI's robots ran shifts at BMW's Spartanburg plant. By late 2025, around 18,000 humanoid units had shipped globally — …
AUGUST 2025
2 changes-
OpenAI launched GPT-5 · era-defining
OpenAI shipped GPT-5 — the successor everyone had been waiting for. Stronger reasoning, bigger memory window (272K tokens), much-reduced made-up-fact rate. Replaced GPT-4o as the default in ChatGPT for f…
-
EU AI Act phase 2 (general-purpose AI rules) took effect
Twelve months after the EU AI Act was published, the rules for general-purpose AI providers kicked in. OpenAI, Anthropic, Google, Meta, Mistral and others now have to publish documentation, manage risks,…
JULY 2025
2 changes-
Perplexity launched Comet — the AI browser · era-defining
Perplexity shipped Comet — a web browser built around an AI assistant from the ground up. Limited release for $200/month Max subscribers in July, then free for everyone in October. The first AI browser t…
-
xAI launched Grok 4
xAI's thinking model upgrade with multi-agent workflows and stronger real-time data integration through DeepSearch. Caught xAI up to where the rest of the leaders had moved during the first half of 2025.…
JUNE 2025
2 changes-
AI doubled some Indian farmers' incomes in real fields · era-defining
'Saagu Baagu' in Telangana, India — an AI-driven precision-farming programme — produced documented results: 21% higher chili yields, 9% less pesticide, doubled incomes. Now scaling to 50…
-
AI cancer detection became routine in real hospitals · era-defining
By mid-2025, over 950 AI medical devices had been cleared by the US FDA — most of them in radiology. Real hospitals were using AI as a 'second reader' on mammograms, colonoscopies, lung scans, and prosta…
MAY 2025
4 changes-
Anthropic open-sourced tools to look inside an AI's mind · era-defining
Anthropic open-sourced its 'circuit tracing' tools — a way to look inside a large language model and watch which features and pathways it actually uses to answer a question. Researchers found Claude plan…
-
Anthropic launched Claude 4 (Sonnet 4 + Opus 4) · era-defining
Anthropic's next major Claude family. Sonnet 4 (the daily driver) and Opus 4 (the heavy hitter, available again after a year off). Big jumps on long, agent-style coding tasks — Opus 4 ran a coding task a…
-
Klarna replaced 700 workers with AI, then started rehiring humans · era-defining
Swedish fintech Klarna laid off 700 customer-service staff in 2023-2024 to replace them with OpenAI-powered chatbots. By mid-2025, customer satisfaction had crashed and complaints had spiked — so the CEO…
-
Pulitzer Prizes required AI disclosure for the second year
The 2025 Pulitzer Prizes required entrants to declare any use of AI in their work for the second year running — and added a new rule for photojournalism: original, unedited image files only. At least one…
APRIL 2025
6 changes-
Alibaba released Qwen3 — China's open-source AI ace
Alibaba's Qwen3 series shipped — eight open-weights models from 0.6B to 235B parameters, fully open under Apache 2.0, with 'thinking on demand' toggles, 119-language support, and native MCP integration. …
-
AI's data centers reopened fossil fuel power plants · era-defining
Through 2024-2025, AI data centers drove a major surge in electricity demand — global data center power use hit ~415-460 TWh in 2024, US data centers alone 4% of national electricity (could hit 12% by 20…
-
OpenAI launched o3 + o4-mini
OpenAI's full o3 shipped to ChatGPT subscribers, plus a smaller, faster o4-mini. Took the leadership spot back from Gemini 2.5 within weeks of losing it. The December 2024 'o3' announcem…
-
Stanford's 2025 AI Index Report — the year in numbers
Stanford's annual AI Index — the most thorough numbers on the AI year — landed with hard data. AI is now in 78% of organisations (up from 55%), inference costs fell 280x in 18 months, FDA-approved AI med…
-
Meta released Llama 4 (Scout + Maverick)
Meta's first mixture-of-experts open-weights models — Scout and Maverick. The model has many sub-models inside; only the relevant ones light up for any question. Cheaper to run for what you get. Scout sh…
-
Cognition shipped Devin 2.0
Cognition's autonomous coding agent went generally available with a complete redesign and a price drop from $500 to $20 per month. New: an agent-native code editor, parallel agent runs, and 'Devin Search…
MARCH 2025
3 changes-
Court let The New York Times' AI copyright lawsuit proceed · era-defining
Judge Sidney Stein let the heart of The New York Times' copyright lawsuit against OpenAI and Microsoft proceed. The Times claims OpenAI illegally trained on its articles. OpenAI claims fair use. The case…
-
Google launched Gemini 2.5 Pro
Google's thinking model. Pauses to reason like OpenAI's o-series, with a million-token context window kept from earlier Geminis. Briefly held the top spot on most leaderboards before OpenAI's o3 took it.
-
AI voice clone scams hit billions in losses · era-defining
By early 2025, AI voice clone scams cost Americans more than $12.5 billion in 2024 alone — a 25% jump. AI-driven romance scams cost another $3 billion. Average loss per voice-clone victim: over $6,000. T…
FEBRUARY 2025
5 changes-
OpenAI launched GPT-4.5 (Orion)
OpenAI's biggest, most expensive non-thinking model — codename 'Orion'. Smoother writing, fewer made-up facts, and very pricey. OpenAI itself called it underwhelming relative to cost, then quietly retire…
-
Anthropic launched Claude 3.7 Sonnet + Claude Code
Anthropic upgraded its mid-tier model with an extended thinking mode you can turn on for harder questions, and shipped Claude Code — a command-line tool that lets developers run Claude a…
-
xAI launched Grok 3
Elon Musk's xAI released Grok 3 — bigger and faster than Grok 2, with a free DeepSearch mode that browses the live web, plus a thinking mode similar to OpenAI's o-series. Available throu…
-
EU AI Act phase 1 took effect
The first chunk of the EU AI Act became enforceable. Eight specific AI practices are now flat-out banned in the EU — social-scoring, untargeted face scraping for databases, emotion-recognition at work or…
-
OpenAI launched Deep Research
A ChatGPT mode that spends 5-30 minutes browsing dozens of websites to write a long research report on your question, with citations. The first 'slow but thorough' AI feature aimed at humans, not other A…
JANUARY 2025
2 changes-
OpenAI launched Operator (research preview)
OpenAI's first AI that opens a web browser and does tasks for you — book a flight, fill a form, gather research. Worked in a sandboxed cloud, not on your real computer. ChatGPT Pro subscribers in the US …
-
DeepSeek launched R1 · era-defining
A Chinese lab released an AI that thinks before it answers — about as well as OpenAI's best — and you can download it for free. It briefly tanked Nvidia's stock and got everyone asking whether AI is stil…
DECEMBER 2024
5 changes-
Microsoft released Phi-4 (open weights)
A 14-billion-parameter open-weights small model from Microsoft Research, trained on heavily curated synthetic data, with benchmark performance matching much larger models on math and reasoning tasks.
-
Google launched Gemini 2.0 Flash
Google's fast Gemini model that can speak, draw, and see — plus a real-time mode that watches your camera and screen while you talk to it (called Project Astra). Free in Google AI Studio at launch.
-
OpenAI made Sora generally available · era-defining
Sora — OpenAI's text-to-video AI — left research preview and became available to ChatGPT Plus and Pro subscribers. Up to 20-second 1080p clips, with editing tools and a remix feature.
-
Meta released Llama 3.3 70B
A free-to-download AI from Meta that matches the much larger Llama 3.1 405B on most tests, while costing about a quarter as much to run. The 'good and small enough' middle for open-source AI.
-
DeepMind's GenCast beat the world's best weather model · era-defining
DeepMind's GenCast — an AI weather model — outperformed the European Centre for Medium-Range Weather Forecasts on 97% of evaluated tests. ECMWF is what your daily forecast already comes from. The AI vers…
NOVEMBER 2024
2 changes-
Anthropic launched the Model Context Protocol (MCP) · era-defining
Anthropic released MCP — a free, open standard for hooking AI assistants up to your data, tools, and apps. Often described as 'USB for AI'. Within a year, OpenAI, Google, and most coding tools were speak…
-
Microsoft relaunched Recall (privacy-overhauled)
After being delayed and re-architected following the May 2024 security backlash, Recall returned to Copilot+ PCs as an opt-in preview with on-device encryption, Windows Hello gating, and exclusion contro…
OCTOBER 2024
3 changes-
Apple Intelligence began rolling out (iOS 18.1)
Apple's first wave of AI features that mostly run on your phone arrived in iOS 18.1, iPadOS 18.1, and macOS Sequoia 15.1 — writing tools, smarter notifications, photo cleanup, and a new look for Siri. Th…
-
Claude 3.5 Sonnet (new) + Computer Use · era-defining
Anthropic upgraded its mid-tier Claude and added Computer Use — the ability for Claude to take screenshots and click around your computer on its own. The first believable 'AI uses your computer' demo fro…
-
AI changed daily life for 1 million blind users · era-defining
Be My AI — the AI feature inside the Be My Eyes app — went mainstream. Built on OpenAI's image-understanding model, it lets blind and low-vision users point a phone at anything and get a description in s…
SEPTEMBER 2024
2 changes-
OpenAI launched o1-preview · era-defining
OpenAI's first AI that pauses to think before answering. It takes longer, costs more, and is much better at hard maths, science, and coding problems where the older models would confidently make things u…
-
NotebookLM added Audio Overview
Google's research tool got a feature where it turns any document, slide deck, or PDF into a two-host podcast between AI voices. The surprise hit of late 2024 — most people couldn't tell it wasn't real on…