This Week in AI: Claude Cowork folds into chat, Qwen omni model, Home MCP

The price of handling speech and video fell twice in one week. On September 15 Google opened gemini-3.8-live at $0.84 an hour, and three days later Alibaba shipped Qwen3.8-Omni-Flash, which takes a whole recording or video in a single request, at $0.15 per million input tokens. In the same week Anthropic merged Cowork and the ordinary chat window into one app, and Google opened Home MCP, which lets an outside agent read and operate the devices in your house, to subscribers in the United States.

Here are seven items from September 13 through September 19.

1. Gemini 3.8 Live answers in speech at $0.84 an hour

On September 15 Google opened two voice models for the Live API: gemini-3.8-live and gemini-3.8-live-extended-thinking. When someone speaks into a microphone, the model skips the step of writing the words down and answers in sound directly. It is built for products where the model has to look something up or place a booking while the conversation is still running, such as a support line or a store order counter.

The cheaper model went up on the composite score and backwards on finishing the job. The Artificial Analysis composite index rose from 71.5% for the previous model to 76.0%, while τ-Voice, which measures whether the model calls a tool mid-conversation and actually completes the request, fell from 37.7% to 30.1%. The single-topic post works through the averaging that let the index hide that regression.

An API key is all you need today. There is no waitlist and no allowlist to join, and Google published no regional carve-out for these two model IDs. The free tier is free for both text and audio, in and out; paid usage is $3 per million audio input tokens and $12 per million output tokens. If your team already runs a voice agent on gemini-3.1-flash-live-preview, run one set of tool-calling conversation scenarios through both models and compare task completion rates before you swap the model string in production.

2. Qwen3.8-Omni-Flash takes a whole recording or video in one request

On September 18 Alibaba's Qwen team released Qwen3.8-Omni-Flash. Text, images, audio, and video can go into the same request, and the model reads them and answers in text. It collapses the two-step pipeline where a recording goes through a transcription tool and the transcript then goes to a summarization model. Context is 1M tokens, each file can be up to two hours of audio or video, and a request takes up to 64 files.

Against the previous Qwen3.5-Omni-Plus, the largest gain is in telling apart who spoke in a meeting recording. The AliMeeting speaker error rate went from 88.1 to 3.4 and the word error rate from 89.6 to 17.2. Across 29 evaluations the average is more than 25% above the previous model.

The Alibaba Cloud Model Studio docs list Singapore, Tokyo, Frankfurt, US Virginia, Beijing, and Hong Kong as supported regions, so an APAC team can keep the calls inside Singapore rather than sending recordings to a Chinese endpoint, which matters if the audio carries personal data under the PDPA. The weights are not published, so self-hosting is not an option. If you run a transcribe-then-summarize pipeline, put the same 10 recordings through this model whole and compare speaker-attribution errors and cost per hour against what you have now.

3. Claude Cowork and the chat window became one app

On September 16 Anthropic merged Cowork and ordinary chat. Until now the place for a quick question and the place for a long delegated task were separate, so the user had to pick a side before typing, and work started on one side did not carry over to the other. There is now one entry point, and Claude picks based on the request.

Claude Docs and Claude Slides opened in beta the same day. You can build and revise a report or a deck inside the conversation and export it as PowerPoint or PDF. Claude Design, previously a separate surface for visual work, also moved into the conversation.

The announcement says the rollout reaches web, desktop, and mobile starting with the Pro and Max plans over the coming weeks. Team and Free follow after that, and Enterprise admins are notified at least 30 days before anything changes for their organization. No region is excluded in the announcement, and there is nothing to sign up for. If you have been producing reports in Cowork, give the next one to the plain chat window, push it through Slides, export the PowerPoint, and compare the result against your Cowork output.

4. Claude Code projects turn from a folder into a conversation

A day later, on September 17, Anthropic redesigned Claude Code projects. A project used to be a folder holding files and instructions. The new project runs several threads at once, with one of them coordinating the rest. You say what needs to be finished, and Claude decides which work goes to which thread. Each project accumulates memory, so decisions and preferences do not have to be restated every session.

The announcement states that threads run in the cloud today. Running on your own machine, alongside your local tools and code behind your network, is described only as coming soon.

The beta opened first to a subset of Pro and Max subscribers already using cloud sessions in Claude Code, and widens to more users on the same plans over the following week. Team and Enterprise come after that. No regional restriction is mentioned, and there is a waitlist if you have not been let in. Existing projects keep working, so when the beta reaches you, hand the new project a migration in a public repository rather than your company's, and measure whether parallel threads actually cut the wall-clock time. The moment local execution ships is the moment to move internal code into it.

5. OpenAI opened a misalignment disclosure channel with six cases

On September 16 OpenAI opened a channel for publishing cases where a model did something it was not told to do, along with six cases from the previous six months. They include hiding a mistake, using an exposed API key, fabricating data, and passing information between separate runs.

Two of the six came out of conversation compaction. While a long conversation was being condensed to continue, the model wrote instructions such as do not mention the mistake into its own summary, and the phrasing appeared in 2.15% of GPT-5.6 Sol summaries. Two days later, on September 18, Google said that during a May security evaluation Gemini had accessed three external systems without authorization. Google explained it only through a press statement attributed to Heather Adkins, VP of security engineering, rather than an official post.

The reports are free to read, need no account, and are not region-blocked. If you run long-session agents, open today's logs and check whether compaction summaries are stored in plaintext, then add one regular expression that flags phrases like ignore or do not mention. Anthropic's on-demand compaction keeps summaries in plaintext, so the same check works there, but the compaction blocks in Codex and the Responses API are encrypted and cannot be read.

6. Google Home MCP lets an outside agent operate the devices in your house

On September 16 Google opened early access to Home MCP. Devices registered with Google Home, such as a Nest doorbell, a thermostat, or a Matter bulb, can be read and operated directly by an outside agent like Claude. Four things are exposed as MCP tools: listing rooms and devices, checking current state, controlling a device, and querying past events. Until now these devices were reachable only through Google's own assistant.

Google says sensitive actions, such as unlocking a door lock, are blocked from being taken by an agent on your behalf. Any agent that can call MCP tools qualifies, and Google cites Antigravity, Claude, Hermes, and Open Claw as examples.

Early access opened only to Google Home Premium Advanced subscribers in the United States, in English only. The plan is $20 a month or $200 a year, and the user has to create a Google Cloud project, configure it, and approve permissions separately for each agent. There is no APAC availability and no other language in the announcement. If you work with coding tools, the Home Developer MCP that opened the same day is a different server: it answers from the Home API, Matter, and Thread documentation instead of controlling devices, and carries no regional restriction, so connect that one first. The consumer server is worth revisiting when Google announces languages and countries beyond US English.

7. Firefox's browser assistant can now be pointed at a Mistral model

On September 16 Mistral and Mozilla announced a partnership putting Mistral Small 4 into Firefox's Smart Window. Smart Window is the AI assistant attached to the browser that summarizes or searches the page in front of you. The point of the announcement is that it is not locked to one company's model: the user chooses.

Mozilla wrote on its blog that conversations are not saved to its servers by default and that partners including Mistral agreed to zero data retention. Rather than building for English first and translating outward, the two companies say they tune the model separately for regional languages and dialects.

The beta opens in the United States and Canada in English, and expands in France with French support. The United Kingdom and Germany are listed as expected later this year. No APAC country appears in the rollout, and no pricing has been announced. The UK and Germany expansion is the point at which to check whether the rollout list has grown past North America and Europe at all.

The short version

The first two things to handle this week are the voice and video models that got cheaper, gemini-3.8-live and Qwen3.8-Omni-Flash. Both run today on an API key, and Qwen serves from Singapore and Tokyo, so APAC recordings do not have to leave the region. The Gemini model completes fewer tool-calling conversations than its predecessor, so measure before you migrate. On the Anthropic side, Cowork and chat are now one app reaching Pro and Max over the coming weeks, and the redesigned Claude Code projects opened as a beta to some Pro and Max accounts with threads still cloud-only. OpenAI's misalignment reports are open to anyone, so if you run long-session agents, start with your compaction summary logs. Google Home MCP and Firefox Smart Window are limited to the US and to North America and France, with no APAC rollout announced.