Apps.

Adobe's own Firefly models, led by Firefly Image 5, are trained on licensed Adobe Stock and public-domain content; the app also runs partner models from Google, OpenAI, Runway, Kling and others.

Bloomberg reports Apple's first smart glasses (N50) slipped to a WWDC 2027 unveiling and a late-2027 release, with two cameras and no display.

Beam AI sells a no-code platform that turns business processes into AI agents. Hebbia sells Matrix, a document-research tool for finance and law.

Open-source AI glasses with a colour display, camera and the Noa assistant with long-term memory; in mass production from July 2026 at $399.

BrowserOS is a free, open-source browser that AI agents drive with your logged-in accounts. Simular Pro is a Mac computer-use agent.

Browses for the longest, most structured reports.

ChatGPT Work (9 Jul 2026) is OpenAI's long-running agent for finished docs, sheets, slides and web apps on paid plans; shared workspace agents remain a research preview for Business and Enterprise.

Anthropic's coding agent now defaults to Claude Opus 5.5 on Pro, Max, Team, Enterprise and the API; v2.1.287 (1 Oct 2026) added mods, TypeScript plugins that reshape its behaviour.

On 16 Sep 2026 Anthropic folded Cowork into the main Claude app: any conversation can now take on agentic work, with Claude Docs and Claude Slides in beta and Claude Design built in.

Anthropic's new leading model and recommended default, performing at Claude Fable 5.1's level on most work while costing 40% less to run than Opus 5.

Leading open-source coding agents alongside OpenCode; Cline (Apache 2.0) reports 11M+ developers, OpenHands adds a cloud and SDK, and Aider's tagged releases have stalled at v0.86.1.

Claude's screen-driving tool is now generally available on the Claude API with no beta header; Opus 5.5 scores 81.8% on OSWorld 2.1, and Claude in Chrome is in beta on all paid plans.

Cursor became part of SpaceX on 14 Aug 2026; its in-house model is now Grok 4.7, built with SpaceXAI and included on every paid plan. Pro $20/mo, Teams $40/user/mo.

DeepSeek's MIT-licensed, model-agnostic agent harness (dsh), open-sourced in August 2026; the v0.2 developer preview in early October added desktop apps for macOS and Windows.

MIT-licensed open-weight line with 1M-token context; the new 552B multimodal V4.1-Flash needs a quarter of the KV-cache memory and DeepSeek says it outperforms the 1.6T V4-Pro.

Cognition's cloud coding agent, now sold with Devin Desktop and the CLI on one plan from $20/mo; Cognition raised $2B at a $48B valuation in Sep 2026 with run-rate revenue of $900M.

Formerly Windsurf (renamed 2 June 2026); now runs Cognition's own SWE-2 model, post-trained from Kimi K3, and Fusion, which pairs a frontier lead model with a cheaper one.

Eleven v4 (28 Sep 2026) clones a voice from 10 seconds of audio across 90 languages, and Music v2.5 (11 Sep) is the new default music model; annualised revenue is now over $600M.

Camera-free display glasses ($599, ~2-day battery) for translation, teleprompting and meeting cues; maker raised $150M at a $1B valuation in July 2026.

A web search engine built for AI agents rather than people, sold as search, contents, agent and monitoring APIs; valued at $1.95B after a $250M Series C in May 2026.

Superhuman (formerly Grammarly) acquired the Fathom notetaker on 14 Sept 2026; Fathom still lists a free plan with unlimited recordings and AI summaries.

Bot-based meeting notetaker that transcribes 100+ languages; in Sept 2026 it added Fireflies Talk, free unlimited voice dictation on Mac and Windows.

FLUX 3 Image (1 Oct 2026) adds JSON bounding-box layout, edits that leave other pixels untouched, up to 10 reference images and 4K output; API-only for now, with open weights promised.

Oura Ring 5 ($399) shipped in June and Oura reached 5.7M paying members, but it shelved its $2.2B IPO on 29 Sept 2026. Galaxy Ring 2 is expected in early 2027.

Google's new frontier model for long-horizon coding and knowledge work, claiming the lead on the Vals Index and 77.9% on DeepSWE v1.1; for now only trusted cyber defenders can use it.

Built for long async research workflows.

Google's any-input video model: Omni Flash makes and edits video from text, images, audio and video by conversation; version 1.1 (27 Aug 2026) extends clips to 40 s and costs about $0.10/sec.

An all-in-one AI workspace of agents; in June 2026 it extended its Series B to $485M at a $2.6B valuation and launched AgentBase, an AI-built database and dashboard tool.

All plans now run on monthly AI credits, with a new $100/mo Max tier; Claude Opus 5.5, GPT-6.1 Sol and Grok 4.7 are in the model list, and desktop computer use is in public preview.

Z.ai's flagship for agentic coding and cyber defence, post-trained on the GLM-5.2 base to 84.5% on CyberGym and 48.2% on AutomationBench, with a 1M-token context.

Google's agent-first coding app and CLI, free for individuals; v2.19.1 (30 Sep 2026) lets you message sub-agents directly, and Gemini 3.8 Flash arrived on 1 Sep.

OpenAI's most capable model and first of the GPT-6 generation, and the first OpenAI model to meet the Critical cybersecurity threshold in its Preparedness Framework.

GPT Image 2.5 (8 Sep 2026) comes as Flare, the fast default, and Sunburst, for precise edits; they hold first and second place on the Artificial Analysis text-to-image leaderboard.

Bot-free AI notepad that captures meeting audio on your device; raised $125M at a $1.5B valuation in March 2026 and added Android and Apple Watch apps in July.

SpaceXAI's frontier model for coding, agentic tasks and knowledge work, with a 500K-token context at $2/$6 per 1M tokens for prompts under 200K tokens.

SpaceXAI's coding agent, now powered by Grok 4.7; a terminal CLI that is free to try, plus an app builder on web and mobile for every Grok plan since 19 Aug 2026.

Grok's search still pulls live X posts alongside real-time web results, which keeps it ahead on breaking news; its maker is now SpaceXAI after SpaceX bought xAI.

Still xAI's current video model, but no longer first; Artificial Analysis now ranks it 8th for image-to-video and 11th for text-to-video, behind MiniMax H3, Gemini Omni Flash and Seedance.

HappyHorse 1.1 (21 June 2026) from Alibaba's ATH unit adds reference-to-video with native audio at up to 1080p; it now ranks 7th for image-to-video and 12th for text-to-video on Artificial Analysis.

HeyGen's Avatar V and Synthesia's interactive avatar Sessions lead business talking-head video; Synthesia raised $200M at a $4B valuation in January 2026, with ARR above $150M.

Tencent's Apache 2.0 open-weight MoE, 770B parameters with 49B active and a 1M-token context, scoring 65.7 on SWE-bench Pro and 92.3 on GPQA Diamond by Tencent's own figures.

Ideogram 4.5 (30 Sep 2026) is an editing-first model built to stop drift across repeated edits, with native 2K output and edits up to 4,016 x 6,016 px, at about $0.03 to $0.22 per image.

Moonshot's 2.8T-parameter open-weight MoE with a 1M-token context and native vision; on Moonshot's figures it beats Claude Opus 4.8 on GPQA Diamond, Terminal-Bench 2.1 and BrowseComp.

Kling 3.0 is still the generally available model, now outside Artificial Analysis's top ten; Kling 4.0 (30-second clips, up to 4K) entered early access on 28 Sep 2026, with full release due in October.

Meta bought Limitless (Dec 2025), stopped Pendant sales and cut service in the UK and EU; Amazon bought Bee (July 2025) and still sells it as a wearable separate from Alexa+.

Microsoft AI's first reasoning model. A mixture-of-experts model with 35B active and about 1T total parameters, 256K context, in preview on Microsoft Foundry.

General agent that runs long multi-step jobs in a cloud sandbox; independent again after China forced Meta to unwind its $2B purchase, it shipped Manus 2.0 on 28 Sep 2026.

Desktop screen-driving from Manus, which has run as an independent company since 11 Aug 2026 after China blocked Meta's planned acquisition.

Copilot Cowork handles long, multi-tool tasks inside Microsoft 365. It became generally available on 16 June 2026 and is billed by usage on top of a Copilot licence.

Zoom launched ZoomMate ($20/user/month) in June 2026 and took its notetaker into Teams and Meet; Teams now holds third-party recording bots in the lobby by default.

Microsoft's first "Autopilot", an always-on agent in Microsoft 365 that prepares meetings and schedules work. Experimental, for Frontier customers.

V8.2 became the default model on 24 July 2026, tuned for aesthetics, image quality and Personalization, and adds an Edit Model that replaces the older reference tools.

Xiaomi's MIT-licensed 1.02T-parameter MoE (42B active) taking text, image, video and audio, with a 1M-token context and 89.9 on Terminal-Bench 2.1 by Xiaomi's own figures.

Open-weight model combining frontier-level coding, a 1M-token context and native image and video input, with per-token compute at 1M context cut to 1/20 of MiniMax's previous generation.

Meta's proprietary frontier model, tuned for agentic coding, with a 1M-token context and video, image and document input, at $1.25/$4.25 per 1M tokens on the Meta Model API.

Gemini 3 Pro Image, generally available in the Gemini API since 28 May 2026 at $0.134 per 1K/2K image; it now ranks below Google's own Nano Banana 2 and OpenAI's GPT Image 2.5.

OpenClaw's orbit: container-isolated NanoClaw, Rust-built ZeroClaw and IronClaw, Nous Research's self-improving Hermes Agent, NVIDIA's NemoClaw sandbox stack and Alibaba's multi-agent HiClaw.

Google's source-grounded research notebook, upgraded in June 2026 to Gemini 3.5 with a secure cloud computer that runs code and exports PDFs, spreadsheets, slides and more.

ChatGPT added a Meetings plugin (beta, 29 Sept 2026) that takes notes in the macOS desktop app for Pro and Business users and saves summaries to ChatGPT Space.

Open-source AI pendant that records and summarises conversations, sold as a one-off purchase ($179 list) with a free plan and no subscription required.

OpenAI's coding agent, now part of the unified ChatGPT desktop app; GPT-6.1 Sol reached Codex on 29 Sep 2026, and the Agents API (beta) offers the Codex harness as a managed service.

Always-on ChatGPT agents running GPT-6 Astra, each on its own cloud computer, that keep working between conversations and reach 4,000+ apps; first dot included on Pro and Business Premium.

The self-hosted personal agent that started the category; OpenClaw 2.0 (v2026.8.1, 31 Aug 2026) added guided setup, shared multiplayer sessions, built-in memory and request-bound approvals.

Open-source coding agent for the terminal, desktop and IDEs that works with any model provider; its makers report 16 million monthly developers and 208,000 GitHub stars.

Retired. OpenAI folded Operator into ChatGPT agent in July 2025, then in July 2026 replaced agent mode with ChatGPT Work and deprecated its Atlas browser.

Meeting notetaker with 35M+ users; in July 2026 it launched Live Assist, an Enterprise-only agent that coaches you during Zoom, Meet and Teams calls.

Answer engine turned agent platform: Perplexity Computer now runs scheduled and event-triggered Automations, picks models by effort level, and offers GPT-6.1 Sol and Claude Opus 5.5.

Perplexity's desktop agent for Mac and, since 28 July 2026, Windows: works on local files and native apps, runs 24/7 on a Mac mini, and can keep sensitive work on local models.

MiniMax H3 (Hailuo, 31 Jul 2026) and an H3-based variant hold the top two image-to-video places on Artificial Analysis; Luma shipped Ray3.2, and Pika now hosts rival models.

Card-sized AI voice recorder ($179); Plaud says it has 2.5M users and $100M annual run-rate, and added eSIM earbuds (Plaud One) and an agent in Sept 2026.

Resemble AI now focuses on deepfake detection and watermarking. Descript has an open-beta API and its Underlord co-editor. Play.ht has shut down.

Google's experimental browser agent from Google Labs, discontinued on 4 May 2026.

Alibaba's largest and most capable model, a 2.4T-parameter MoE (95B active) with 1M context and image and video input; a text-only checkpoint has open weights under a custom licence.

Ray-Ban Meta Gen 3 launched at Connect 2026 from $449 with a 9-hour battery; a camera-free Ray-Ban Meta Audio ($349) ships 13 Oct 2026.

V4.1 Utility is Recraft's best-ranked model on Artificial Analysis (Elo 1,023, roughly 21st to 32nd); a faster V4.1 Flash followed in September 2026.

Gen-4.5 is still Runway's own video model, but Runway now sells a platform: Aleph 2.0 editing, Ruby HDR grading, Premiere Pro and After Effects plugins, and hosted rival models.

Seedance 2.5 (Aug 2026) makes clips of up to 30 seconds from up to 50 mixed references and lets you re-edit specific moments or characters; it ranks 3rd on Artificial Analysis text-to-video.

OpenAI closed the Sora app on 26 April 2026 and set 24 September 2026 as the shutdown date for the Sora 2 API (sora-2, sora-2-pro), with no replacement named.

Suno v6 (9 Sep 2026), developed with Warner, BMG and Believe, adds section editing and mashups; on 18 Sep Universal and Sony sued again, alleging v6 launders 60,202 recordings.

Best raw fidelity and stem separation, with a cleaner licensing story after Universal settled.

An open-source personal assistant, native to Mac, whose isolated credential service lets it use your keys and logins without ever seeing them; now also on iOS, Android, Windows and Linux.

Still available, but Google now tells developers to use Gemini Omni Flash as the default video model and keep Veo 3.1 for scene extension, last-frame control or legacy pipelines.