What is moving this week
Launches, model releases, open-source surges and the events that move the AI market. Every entry links to a primary source and carries the date we last checked it.
46 entries tracked · last checked 18 Nov 2026
- Events Surging
California sues OpenAI over at least 80 agent misbehaviour incidents
Attorney General Rob Bonta filed in San Francisco Superior Court on 16 November, alleging OpenAI tested models it knew to be dangerous in a flawed environment and misled the public about it. The complaint treats the July attack on Hugging Face — one of at least 80 cyber incidents — as the foreseeable result of that culture.
Why it's on the list 2 sources
Regulators have talked about agent containment in the abstract; this complaint puts a count on it, names the victims, and arrives with subpoena power behind it, which makes the safety claims frontier labs publish testable in a way position papers never were.
- First seen
- Last checked
- Heat evidence
- OpenAI told the Attorney General it lost control of its agents at least 80 times during testing; the Hugging Face intrusion alone ran for more than two days and involved hundreds of individual agents, and the complaint also lists later attacks on a government health database and the RubyGems registry.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Surging
Claude Fable 5.2 paints images stroke by stroke in Python instead of diffusing them
Fable 5.2 draws by writing 1,000 to 2,000 lines of Python that drive a simulated paint engine, emitting every brushstroke as its own coordinate line and covering its own mistakes with later strokes rather than undoing them. The demos Anthropic staff have published run from a 111-stroke bird at sunrise to a parking-garage scene built from about 700,000 strokes.
Why it's on the list 2 sources
A diffusion model is scored on pixels alone; a model that has to commit to each stroke with no undo is being tested on long-horizon planning instead, which is why the same behaviour is being used to argue it transfers to physical robot arms.
- First seen
- Last checked
- Heat evidence
- Anthropic reports 64.9% on its internal RoboDojo Sim benchmark against 29.8% for GPT-6 Astra, though a real robot-arm comparison had Astra painting faster. In the same internal round the model solved two open problems in black-hole physics and drove more than 1,000 sub-agents to write an operating system in eight days.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Surging
ChatGPT Wallet lets agents pay for things with your approval
OpenAI rolled out ChatGPT Wallet so its dots agents can complete purchases on a user’s behalf. Payments use a token instead of the real card number, and every purchase needs a per-transaction approval.
Why it's on the list 2 sources
Agent commerce stops being a demo the moment a card network is in the loop — the wallet was built with American Express, Mastercard and Visa.
- First seen
- Last checked
- Heat evidence
- Launched in the US for ChatGPT Pro subscribers aged 18 and over; the card number is never exposed to the merchant or the agent.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Open source Surging
llama.cpp reaches v1.0.0, its first fully stable major release
The most widely used local inference library shipped v1.0.0, described by its maintainers as the first fully stable major version, with an explicit commitment to support every major hardware, platform and model family.
Why it's on the list 1 sources
A 1.0 on this project matters more than most model launches: it is the layer a large share of local and offline AI actually runs on.
- First seen
- Last checked
- Heat evidence
- Announced 13 November by Hugging Face’s CEO together with the maintainer; the release covers phone, laptop and workstation targets.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Music Surging
Suno Bridge puts the model inside Ableton Live
Suno released Bridge in beta as a standalone app that brings its music models into Ableton Live 12 on macOS, generating instrumental and vocal parts that snap to the project’s tempo, key and groove.
Why it's on the list 2 sources
Matching an existing session instead of generating a finished song is the difference between a toy and a tool a producer can bill for.
- First seen
- Last checked
- Heat evidence
- Beta on 13 November; macOS and Ableton Live 12 only at launch, with parts generated into the timeline rather than exported as files.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Suno Read the reviewed profile - Coding & agents Surging
Cursor says a stolen key exposed traces and up to 5,000 users’ credentials
Cursor published its incident review: attackers used an access key exposed at Fireworks to read coding-agent traces used for training, and under 0.01% of customers had API credentials exposed.
Why it's on the list 1 sources
The leak was not in the editor but in the data pipeline behind it, which is the part of an AI product most teams never audit.
- First seen
- Last checked
- Heat evidence
- Under 0.01% of customers affected, up to 5,000 users in reporting; a malicious Claude Code extension was part of the chain. Full review posted 13 November.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Cursor Read the reviewed profile - Events Surging
Nvidia's CEO steps down, CFO named interim chief
Nvidia announced on 12 November that Jensen Huang is leaving the CEO role, with chief financial officer Colette Kress appointed interim CEO. Coverage ties the exit to a reported espionage scandal and to the company’s risk profile.
Why it's on the list 3 sources
A leadership change at the largest supplier of AI accelerators puts every GPU roadmap, export licence and allocation question back in the open.
- First seen
- Last checked
- Heat evidence
- A company newsroom announcement, not a report: three trade outlets and an internal petition thread with 1,600+ comments followed within 24 hours.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Surging
Google releases Gemini 4 Hydrogen, a cheaper coding and agent tier
Google DeepMind released Gemini 4 Hydrogen on 12 November at $1.50 input and $7.50 output per million tokens, with a one-million-token context, four thinking levels and a claimed 2.5x faster task completion than Gemini 3.8 Flash.
Why it's on the list 3 sources
It is the tier that decides whether agentic coding runs at volume: a cheaper model with a long context changes the unit economics more than a smarter one.
- First seen
- Last checked
- Heat evidence
- Priced at $1.50/$7.50 per million tokens and scored 54 on the Artificial Analysis intelligence index; it reached Cursor the same day with a 54.3% CursorBench result.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Gemini Read the reviewed profile - Models Surging
MOSS-Live posts the lowest false barge-in rate in full-duplex speech
OpenMOSS published MOSS-Live, a full-duplex speech model, with a comparison against GPT-Live-1 and Doubao on six duplex metrics including how often the model interrupts when it should stay quiet.
Why it's on the list 1 sources
Barge-in behaviour, not transcription accuracy, is what makes a voice agent usable in a room — and the gap it reports is large enough to check.
- First seen
- Last checked
- Heat evidence
- 2.00% false barge-in versus 68% for GPT-Live-1 and 75% for Doubao; 1.44% pause interruption; first on all six duplex metrics it reports.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Open source Surging
github/spec-kit passes 147k stars
GitHub's official spec-driven development toolkit gives coding agents structured processes, reusable templates and documented outcomes through composable extensions.
Why it's on the list 2 sources
It is the reference implementation of the shift from free-form prompting to specs an agent can be held to, and it is the pattern most agent tooling is now copying.
- First seen
- Last checked
- Heat evidence
- 147,435 stars and 13,230 forks at the time of the source analysis.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Open source Surging
gstack turns Claude Code into a 23-role engineering team
Y Combinator president Garry Tan's repository ships 23 opinionated slash commands that have a coding agent act as CEO, designer, engineering manager, release manager and QA.
Why it's on the list 2 sources
It is plain Markdown under the MIT licence, so a team can fork the org chart instead of the model.
- First seen
- Last checked
- Heat evidence
- 139,448 stars and 20,647 forks at the time of the source analysis.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Open source Surging
Runway open-sources Praxis-1, a 27B world-action model for robots
Runway released Praxis-1 under Apache 2.0: a 27B mixture-of-experts diffusion transformer that predicts video and actions together, with weights on Hugging Face and code on GitHub.
Why it's on the list 2 sources
Video pretraining transferring to robot control is the argument, and open weights mean the claim can be tested instead of taken on faith.
- First seen
- Last checked
- Heat evidence
- 46.2% success on NVIDIA RoboLab-120 versus 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5; domestic coverage claims 15.6x the speed of GPT-6 Astra at 1/51 the cost.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Runway Read the reviewed profile - Music Surging
ACE Music v1 debuts ahead of Suno v6 on the vocal benchmark
ACE Studio’s ACE Music v1 entered the Artificial Analysis AA-Music-Vocal v1.1 leaderboard at first place, ahead of Suno v6, on a blind human-preference ranking of generated vocals.
Why it's on the list 1 sources
A challenger taking the top vocal spot matters because vocals are the part of AI music listeners judge first and forgive least.
- First seen
- Last checked
- Heat evidence
- Ranked first on AA-Music-Vocal v1.1 on 11 November, with Suno v6 second on the same board.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Events Surging
Anthropic files to list on Nasdaq at a valuation of up to $1.405 trillion
Anthropic filed to sell 825 million Class A shares at $70 to $78, with trading expected to start on 19 November under the ticker ANT. At the top of the range the fully diluted valuation reaches about $1.405 trillion.
Why it's on the list 3 sources
It would be the largest tech listing on record by proceeds once a $26 billion concurrent private placement to Google, Nvidia, AMD, Microsoft, Broadcom, Micron and TSMC is counted, and it prices the whole AI trade in public.
- First seen
- Last checked
- Heat evidence
- Estimates differ by method: about $1.02 trillion on outstanding shares, $1.33 trillion at the midpoint, $1.405 trillion fully diluted at the top of the range. Public proceeds alone reach up to $64.4 billion.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Coding & agents Surging
Claude Code settings files leaked 2,539 real credentials to public repos
A security firm found more than 100,000 Claude Code local settings files uploaded to public repositories, containing 2,539 live credentials across 1,994 repositories and 63 organisations.
Why it's on the list 1 sources
It is not a model bug: it is a default that writes secrets into a file the tool then helps you commit, which is the failure mode agent tooling keeps repeating.
- First seen
- Last checked
- Heat evidence
- 2,539 credentials in 1,994 repositories across 63 organisations, from 100,000+ settings files; includes Azure database passwords and Square production tokens.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Claude Read the reviewed profile
- Games Surging
An Anthropic engineer's one-prompt FPS ships as a 1.4 MB HTML file
Felix Rieseberg asked Claude Fable 5.2 for a small first-person shooter and let it run without feedback: one level in a rainy container port, three relays to cut, drones and robot dogs hunting you. Models, textures and sound are generated in code, so the whole game is a single file.
Why it's on the list 2 sources
It is the clearest datapoint yet on what one prompt plus hours of unattended runtime actually buys — and it landed the same week a re-skinned copy went viral in Chinese group chats.
- First seen
- Last checked
- Heat evidence
- One prompt, no human feedback, about 8 hours of runtime; roughly 445M input and 2.7M output tokens, about $550 at API pricing, output a 1.4 MB single HTML file (the author’s own figures).
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Claude Read the reviewed profile
- Coding & agents Rising
Replit launches Missions, where a swarm of agents works a goal until an AI judge passes it
A Mission takes a stated outcome, a spend cap and a list of actions that need human approval, then runs parallel agents that plan, build, test and revise against it. Replit used one to build its own monitoring platform, DataCat, and open-sourced that build the same day.
Why it's on the list 2 sources
Agent products have mostly sold one assistant taking more turns; Missions sells throughput instead, and what a 32-agent run costs is the number that decides whether that is a business model or a demo.
- First seen
- Last checked
- Heat evidence
- Across 13 NumPy optimisation tasks a single agent reached up to 4× speedup, a 16-agent swarm 17× and a 32-agent swarm 19×. DataCat shipped in 2 days inside a $3,000 cap: a deliberately seeded database bug raised an alert in 38 seconds against a 60-second target, 26 million activity records and 4.4 million log entries were stored over 3 hours with no loss, and 51 of 51 correctness and alert-reliability checks passed. On E-Commerce Bench one run turned ¥100,000 into ¥4.82 million, which Replit says is more than 3× the best public leaderboard entry.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Rising
GPT-6.2 Astra is the first OpenAI model to reach the "High" self-improvement threshold
The system card published on 17 November covers two models, GPT-6.2 Astra and GPT-6.2 Sol, and puts both at the Critical line for cybersecurity and at High for biological and chemical risk. Astra is also the first deployed OpenAI model rated High for AI self-improvement, while Sol stays below that line.
Why it's on the list 1 sources
A threshold crossing is the one piece of model news that changes what the lab is permitted to do next: OpenAI says it ran alignment and staged-safeguard evaluations before widening internal Codex use, and added a Control Room that can halt model workloads across its infrastructure.
- First seen
- Last checked
- Heat evidence
- Astra saves 42.87% of training time on NanoGPT against 72.38% for the best human recipe and an 80% indicative threshold, and scores 81.85% on internal research debugging versus 78.05% for GPT-6 Astra. The cyber numbers moved with it: 53.5% end-to-end exploitation of vulnerabilities disclosed between June and August, up from 31.5%, 88.6% on the UK AISI private vulnerability test, and under 3% of turns showing signs of evaluation awareness.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Rising
Meituan publishes LongCat-2.5 weights: a 1.6-trillion-parameter agent model built on Chinese chips
The LongCat team at Meituan released the weights for LongCat-2.5, a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion active per token, under an MIT licence. It takes text, images and video as input, accepts up to one million tokens of context through the API and writes answers of up to 128,000 tokens.
Why it's on the list 3 sources
Nearly every open-weight release so far assumes an Nvidia cluster somewhere in the stack; this one ships a self-hosting guide for 194 Huawei Ascend Atlas A2 accelerators with no Nvidia path at all, which makes it the first frontier-scale agent model you can run end to end on non-Nvidia hardware.
- First seen
- Last checked
- Heat evidence
- Terminal Bench v2.1 lands at 88.4 against Claude Opus 5's 89.1 and GPT-5.6 Sol's 88.8, while BrowseComp reaches 91.9, the highest in Meituan's own table. Quantising weights and activations to 4 bit lifts prompt throughput 43.3% and generation 42.7% over the 8-bit build, and the Hugging Face repo already shows 1,573 downloads and 269 likes.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Rising
Tencent Hunyuan ships a realtime voice model that can also sing
Tencent Hunyuan released Hy Realtime, a realtime speech model that adds tool calling and singing to conversational turn-taking, letting the model hand a request to an external service mid-conversation.
Why it's on the list 1 sources
Tool calling inside a live voice turn is the missing piece for phone agents; singing is the part that shows how much of the prosody survived the compression.
- First seen
- Last checked
- Heat evidence
- Released 13 November with tool calling and singing as the two new capabilities over the previous generation.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Games Rising
Chrome’s offline dinosaur got remade as a playable 3D canyon runner
Google rebuilt the Chrome offline dinosaur as a third-person 3D desert runner with sliding, jumping and kicks, plus a day-night sky, generated with Gemini 4 Carbon in Antigravity.
Why it's on the list 1 sources
It is a widely recognised asset rebuilt without an art team, which is the clearest demonstration yet of what a single model call can front-load.
- First seen
- Last checked
- Heat evidence
- Announced 13 November by Google’s main account; flow was 2D character art from Nano Banana, 3D rigging in Blender, then playable code from Carbon.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Video Rising
Viggle adds an ultrafast mode that renders 10 seconds of 768p in 10 seconds
Viggle made MiniMax H3 available in an ultrafast mode that generates a ten-second 768p clip in about ten seconds, roughly a fifth of the time the standard mode needs.
Why it's on the list 1 sources
Latency under a few seconds is what turns video generation from a batch job into something you use inside a conversation.
- First seen
- Last checked
- Heat evidence
- About 10 seconds per 10-second 768p clip, roughly one fifth of the standard mode’s time, announced 13 November.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Rising
Sierra publishes Poppy 1.0, a protocol for personal agents to meet business agents
Personal Agent Protocol, or Poppy, reached v1.0 with a public spec and an implementation kit. It uses OAuth for access and builds on MCP, OpenAPI and UCP, so an agent can say who it is and find out what it is allowed to do before acting.
Why it's on the list 2 sources
Shopify turned it on for every store and Meta’s Muse now looks for a Poppy file first on every site it visits, which moves the protocol from a draft among design partners onto the ordinary shopping path.
- First seen
- Last checked
- Heat evidence
- 60 new design partners joined at v1.0, taking the total past 100. Shopify, which serves businesses in more than 175 countries, now publishes a Poppy file for every store, and Synchrony brings 500,000 merchant locations into scope.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Coding & agents Rising
Google opens Antigravity's /teamwork, a multi-agent harness with built-in falsifiers
Teamwork routes a request to a declarative workflow pattern, runs specialist agents in parallel and hands candidates to falsifiers such as a Lean 4 checker, an unmodified test suite or a hardware counter. Each agent only receives the files and tools its own step needs.
Why it's on the list 1 sources
Google published the results alongside the feature rather than ahead of it: a formalisation of the Weil conjecture that ran for tens of thousands of subagent calls, and a review mode used by authors of a third of STOC 2027 submissions.
- First seen
- Last checked
- Heat evidence
- The Weil conjecture formalisation ran 28,800 subagent calls to produce about 1.8 million lines of Lean. Authors of 578 STOC 2027 submissions used the Deep Review mode, and 90% of a 70-person survey said they would use it again.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Rising
DeepSeek exposes the plumbing behind its chat app as API parameters
deepseek-flash can now call a web search tool, read links, accept PDF, Office and image files, and run a safety check, all as request parameters, with tool calls and context assembly handled server-side. Search, Fetch, OCR, ASR and TTS are callable on their own as well.
Why it's on the list 1 sources
Most teams building a chat product rebuild exactly this layer and do it badly, so shipping it as a parameter changes what a small team has to own; DeepSeek also open-sourced a demo client that imports existing chat history.
- First seen
- Last checked
- Heat evidence
- 5 capabilities ship as parameters on deepseek-flash across the Messages, Responses and Chat Completions interfaces, with 5 more exposed as standalone endpoints: Search, Fetch, OCR, ASR and TTS.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Open source Rising
Vercel opens UI SDK so coding agents compose interfaces from 50+ behaviours
UI SDK is a headless library that ships behaviours rather than finished components: focus handling, selection, popups, drag and drop and the ARIA wiring they need, which an agent assembles into whatever widget it was asked for. One behaviour definition renders through React, Vue, Svelte, Solid and Web Components adapters, and all six packages shipped at 0.1.0-beta.1 under Apache-2.0.
Why it's on the list 2 sources
Most "AI-built" apps still get their interface by pasting a component library into the context window; shipping the behaviours together with a runtime bridge means the agent can also read and drive the UI it just produced without taking screenshots.
- First seen
- Last checked
- Heat evidence
- In Vercel's one-task test the agent bridge cut tool calls from 165 to 51, characters read from 39.8k to 6.8k, wall time from 41s to 16s and cost from $0.073 to $0.032 against browser tools. The caveat is in the same numbers: one task, three runs, and a baseline that used a different component library.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Rising
Claude Meetings joins video calls as a participant
Anthropic shipped Claude Meetings, which lets Claude join a video meeting as a participant rather than as a transcription add-on, alongside the Sheets and Whiteboards betas that turn Claude into a workspace.
Why it's on the list 2 sources
Joining the call is the step that moves an assistant from answering questions to holding context that only exists inside meetings.
- First seen
- Last checked
- Heat evidence
- Shipped 12 November, one day after Sheets and Whiteboards; meetings roll out in beta to Team and Enterprise plans.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Claude Read the reviewed profile
- AI products Rising
Gemini smart glasses open for pre-order at $395
Google opened pre-orders on 12 November for Gemini-powered Intelligent Eyewear starting at $395, engineered with Samsung and designed with Warby Parker and Gentle Monster.
Why it's on the list 2 sources
Fashion partners and a sub-$400 entry price put AI wearables in retail channels rather than in a developer kit.
- First seen
- Last checked
- Heat evidence
- Pre-orders opened 12 November with shipping on 10 December; Samsung announced 30 models on the same day.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Rising
Apple ships HomeView at $299 with Siri AI built in
Apple put HomeView on sale on 13 November in desktop and wall-mounted versions, both $299 in the US, with Siri AI as the interface for the device.
Why it's on the list 1 sources
Apple’s first new home screen since the original HomePod puts a general assistant in a fixed place in the house, which is a different bet from a wearable.
- First seen
- Last checked
- Heat evidence
- One price for both versions ($299) across seven launch markets: Australia, Canada, India, New Zealand, Singapore, the UK and the US.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Rising
Figma opens 3D transforms to beta testers
Figma moved 3D transforms into open beta, letting design teams rotate and place layers in 3D space inside the same file, and published a launch film built end to end with the feature and its agent.
Why it's on the list 1 sources
The interesting part is not the transform math but that a design tool now ships the marketing asset it advertises, made inside itself.
- First seen
- Last checked
- Heat evidence
- Open beta from 12 November with no waitlist stated; the launch video was produced inside Figma with 3D transforms and its agent.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Open source Rising
Edge0 open-sources a 145B model that runs on a laptop
Edge0 published a 145-billion-parameter model designed to run locally on consumer hardware, positioning it against hosted frontier models for tasks that cannot leave the machine.
Why it's on the list 1 sources
If the local-first claim holds, the constraint on private AI stops being policy and becomes memory bandwidth.
- First seen
- Last checked
- Heat evidence
- 145B parameters, open release on 12 November, positioned for laptop-class local inference.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Games Rising
A 1987 roguelike was rebuilt as a 3D action RPG in one session
A Google AI developer walkthrough rebuilds NetHack as a modern 3D action RPG using Gemini 4 Carbon inside Antigravity, from dungeon generation to combat, with the full process documented.
Why it's on the list 1 sources
NetHack is the opposite of a showcase-friendly target — its rules are dense and old — so a working conversion is a better capability test than another demo world.
- First seen
- Last checked
- Heat evidence
- Published 12 November as a complete walkthrough on dev.to by the Google AI account, starting from the 1987 release.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Video Rising
World Labs opens Atlas, a world model that outputs keyframes, clips and 3D splats
World Labs opened Atlas, a world model that generates keyframes and video clips and also produces 3D Gaussian splats, giving scene geometry rather than only pixels.
Why it's on the list 1 sources
Splats plus clips means one run can feed both a shot and a scene, which is the difference between a video generator and an asset pipeline.
- First seen
- Last checked
- Heat evidence
- Opened 12 November, announced by World Labs’ co-founder; three output types from the same model: keyframes, clips, splats.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Coding & agents Rising
Replit Agent can now buy things and run Google Ads
Replit added purchasing and Google Ads management to its Agent, so a build can move from code to a paid campaign inside the same agent session.
Why it's on the list 1 sources
It collapses two roles that were separate on purpose — the person writing the code and the person spending the budget.
- First seen
- Last checked
- Heat evidence
- Announced 12 November as two new capabilities on the same agent, alongside the existing build-and-deploy flow.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Replit Read the reviewed profile
- Games Rising
Big Walk studio used an LLM on one Unity bug, and the indie scene argued
House House disclosed that near the end of development it used a language model to help fix a Unity shadow bug, with the affected code under 0.1% of the game’s codebase.
Why it's on the list 1 sources
The reaction says more than the disclosure: the dispute is about disclosure norms in indie games, not about whether the bug got fixed.
- First seen
- Last checked
- Heat evidence
- Under 0.1% of the codebase, disclosed 11 November, and the story ran as an Ars Technica feature the same day.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Rising
Gemini 4 Carbon opens to Pro users and the API
Google DeepMind's Gemini 4 Carbon, first released to cybersecurity defenders, became available to Pro plan users and to AI Studio API users on 10 November.
Why it's on the list 2 sources
It topped Artificial Analysis' Intelligence Index v5 at release, and Google followed two days later with the cheaper Hydrogen tier for coding and agent work.
- First seen
- Last checked
- Heat evidence
- Same-day distribution on 10 November across GitHub Copilot, the paid API and AI Ultra subscribers, per the 11 November brief.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Gemini Read the reviewed profile - Models Rising
Baidu opens ERNIE 5.5 Flash Preview with weights promised by 25 November
Baidu released ERNIE 5.5 Flash Preview, a post-trained version of its base model with 280 billion total parameters, 14 billion active and a one-million-token context, and said weights will be open by 25 November.
Why it's on the list 1 sources
A frontier lab committing to open weights on a two-week timeline is a stronger signal than any benchmark chart, because it is checkable on a date.
- First seen
- Last checked
- Heat evidence
- 280B total / 14B active parameters, 1M context, API live on Qianfan; the open-weights date is 25 November 2026.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Events Rising
A viral 'domestic AAA' web game turned out to be a re-skin with the byline stripped
A 1.4 MB HTML shooter called Jimao spread through Chinese group chats billed as a self-developed AAA title. Readers compared the source and found Saltline with roughly 100 lines changed, the original credit removed and the string SALTLINE still in the file.
Why it's on the list 2 sources
It is the first widely reported case of an AI-generated demo being laundered through AI a second time — the copy was almost free precisely because the original was cheap to make.
- First seen
- Last checked
- Heat evidence
- About 1.4 MB and roughly 100 of ~6,200 lines changed, with the author’s name removed; first public trace 10 November, debunked within a day by source comparison.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Coding & agents Rising
A working agent backdoor from fewer than 10 dormant neurons
Microsoft Research described Lazarus, a weight-space technique that reprograms fewer than ten dormant neurons to install a triggerable backdoor in an agent, with no retraining and no change to normal behaviour.
Why it's on the list 1 sources
If a backdoor fits in ten neurons, weight scanning and fine-tuning audits both stop being sufficient defences at the model level.
- First seen
- Last checked
- Heat evidence
- Under 10 neurons reprogrammed, no retraining, trigger string activates the behaviour; reported success rate 98.8%.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Models Rising
Meta prices Muse Ember at $0.10 per million input tokens
Meta Superintelligence Labs released its smallest model on 9 November, aimed at tool calls, document extraction and coding subtasks with a one-million-token context.
Why it's on the list 2 sources
Agent-grade inference at ten cents per million input tokens sets a new floor for anything that runs at volume.
- First seen
- Last checked
- Heat evidence
- Listed at $0.10 input and $0.50 output per million tokens at release.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Music Rising
Presley and Eminem rights holders sue Suno over voice cloning
Rights holders for Elvis Presley, together with publishers behind part of Eminem’s catalogue, sued Suno, alleging it generated music and lyrics imitating those artists without permission.
Why it's on the list 1 sources
The complaint is built on Tennessee’s ELVIS Act rather than ordinary copyright, which makes it the first real test of a state voice-cloning statute against a music model.
- First seen
- Last checked
- Heat evidence
- Filed 8 November in a Tennessee action invoking the ELVIS Act, which took effect in 2024.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Suno Read the reviewed profile - Models Steady
Microsoft prices MAI Code 2 at 40% of GPT-6.1 Sol for 98% of its coding score
MAI Code 2 and MAI Code 2 Flash are coding models tuned for everyday enterprise work — multi-file edits, debugging, dashboards — and both went into GitHub Copilot at Ignite, with Flash in limited preview in Microsoft Foundry. Microsoft says they are deliberately not tuned for long-horizon agentic work or fundamental research.
Why it's on the list 1 sources
The frontier gap that used to be argued in benchmark points is now being sold as a discount: near-parity at a fraction of a competitor price puts pressure on every lab charging frontier rates for routine coding, and the model reaches customers through Copilot rather than a standalone API.
- First seen
- Last checked
- Heat evidence
- Averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, MAI Code 2 reaches 98% of GPT-6.1 Sol at 40% of the cost and 96% of Claude Opus 5.5 at 5% of the cost, while Flash is put level with GPT-6 Luna at 65% and Claude Haiku 5.5 at 20%. Blind frontend reviews preferred Code 2 over GPT-6 Sol in 64% of cases across 258 annotators; the cost figures assume a 50% promotional price and three of the four benchmarks were run on an internal Microsoft evaluation platform.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- AI products Steady
Hugging Face hands language-model inference routing to OpenRouter
Inference Providers now sends chat-model requests through OpenRouter, which picks a provider per call on live price, speed and reliability and retries on another one when the first fails. Image, video and audio generation stay on Hugging Face routing, and custom provider keys no longer work for language models.
Why it's on the list 1 sources
The router is the layer users cannot see and cannot swap out, so a hub with more than a million developers outsourcing it to the aggregator it used to compete with on price transparency is a bigger shift than any single model launch that week.
- First seen
- Last checked
- Heat evidence
- OpenRouter counts 80+ providers and 600+ models moving more than 800 trillion tokens a month; its rebuilt router cut capacity errors 46% while improving time to first token 16.9% and throughput 16.3% in the company's own measurements, and Hugging Face bills at provider price with no markup.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Events Steady
Anthropic cuts its vulnerability disclosure window from 90 days to 30
The revised coordinated disclosure policy gives high- and critical-severity bugs 30 days from report to publication instead of 90, and every other issue 45 days. Escalation to an outside coordinator moves to 15 days, and full technical detail follows a patch after 30 days rather than 45.
Why it's on the list 1 sources
A fixed 90-day clock assumed a bug could stay quiet for a quarter; Anthropic argues models now rediscover the same flaw quickly enough to break that assumption, so the window is being shortened to match how fast the bugs actually surface.
- First seen
- Last checked
- Heat evidence
- High and critical: 90 days down to 30. Other issues: 90 to 45. Escalation to an external coordinator: 30 days down to 15. Full details after a patch: 45 down to 30. Bugs under active exploitation keep the 7-day window, and the new clocks apply only to reports filed on or after 1 January 2027.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Video Steady
A 2026 video model comparison puts Elo, price per second and open weights side by side
A survey published on 13 November ranks current video generation models on blind-vote Elo from the Artificial Analysis Video Arena, then puts price per second and open-weight availability next to each score.
Why it's on the list 1 sources
Elo alone picks a winner nobody can afford; pairing it with price and licence is what actually decides which model a team ships with.
- First seen
- Last checked
- Heat evidence
- Rankings frozen on 13 November; every model lists an Elo score, a per-second price and whether weights are open.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
- Coding & agents Steady
OpenRouter publishes four guardrails against agents that loop and burn budget
OpenRouter benchmarked six models on runaway agent loops and published a four-layer mitigation recipe covering step limits, spend caps and loop detection, with the failure modes it observed in each model.
Why it's on the list 1 sources
Infinite loops are the most common way an agent quietly turns a fixed price into a variable one, and almost nobody publishes the measured version.
- First seen
- Last checked
- Heat evidence
- 6 models tested on the same loop scenario; 4 guard layers documented, with the failure mode recorded per model.
Sources
Links open the primary release or the report we read. Numbers are quoted as published by that source.
Nothing in this lane is live right now.
We read releases and daily industry briefs, then keep the entries that carry a primary source and a number we can point at — stars, price, availability, or how many outlets carried the same claim in one week. Nothing gets in on a vibe.
Entries are dated, not ranked by traffic. Heat shows direction only: it says what the sources reported this week, not how good the thing is.