Anthropic merges Claude chat and Cowork, adds Docs, Slides, and Design tools
Claude just merged its chat and background work modes into one. No more switching between tools.
Here’s what changed: you drop a task, close your laptop, and Claude keeps working. If it hits a wall, it asks you. You stay in control, it handles the grind.
The bigger news is three new tools now live in beta on all paid plans:
- Claude Docs – write and edit documents, fix lines yourself or leave comments for Claude to handle
- Claude Slides – build decks, move elements around, then present straight from Claude or export normally
- Claude Design – mock up visuals without leaving the conversation
The part worth getting excited about: you can go from one-pager to deck to matching visual, all inside one chat. No copy-pasting between apps.
If you used Cowork before, your chats, projects, and files are still there. Rolling out to Pro and Max plans soon.
Google DeepMind’s self-improving agent cuts AI search costs by up to 162x
Google DeepMind just showed an AI that gets better at solving hard problems by learning from its own past attempts. That’s the core idea behind Dream-RSI.
Here’s the problem it solves. When an AI explores a complex problem, it uses a strategy to decide where to look next. Most systems use a fixed strategy that never improves. Updating it live is expensive. Dream-RSI found a clever middle path.
Instead of running new expensive experiments to test better strategies, it replays old ones. Think of it like a coach reviewing game footage instead of running a whole new practice. The AI simulates thousands of alternative approaches using its history, picks the best, then deploys it for real.
- Works across algorithm design, math optimization, and GPU engineering
- Cuts AI calls by up to 162x in some settings
- Does not change the underlying model, only how it explores
That last point matters. You can drop this on top of any existing AI agent without retraining it. It just makes the search smarter over time.
Free open-source tool turns a photo and audio into a minutes-long talking avatar video
LongCat-Video-Avatar 1.5 just dropped, and it does something that used to cost serious money: give it a photo and an audio clip, and it spits out a realistic talking-head video. No studio. No camera. No subscription fee.
It’s fully open source under the MIT license, meaning you can use it commercially, modify it, and ship it in your own products.
Here’s what makes it actually usable:
- Generates videos in just 8 inference steps, roughly 15x faster than older approaches, about 1 minute for a 10-second clip
- Handles long videos without the face drifting or lips falling out of sync
- Supports multi-person scenes where two characters talk to each other
- Works on real humans, anime, and animals
To get started: git clone the repo, set up a conda environment, and run torchrun with your image and audio inputs. Weights download straight from Hugging Face. This opens the door to AI presenters, dubbed content, and virtual characters built entirely from code.
Claude gets closer to becoming an all-in-one superapp
Anthropic is combining Claude chat, Cowork, and Design into a single platform that includes two new products, Claude Docs and Slides. Just start your conversation, and Claude will loop in the relevant tool while keeping skills, tasks, and context intact in one place. It’s rolling out now, starting with Pro and Max plans. Learn how the updated platform works.
ElevenLabs launches an AI receptionist for small businesses and freelancers
Reception is a voice agent that answers phone calls 24/7 — scheduling appointments, answering customer questions, and routing issues to the right person. The lab says setup is simple: just paste a website URL and Reception will pull your services, hours, and intake process. It’s available globally and works in 70+ languages. Watch it in action.
Startup rolls out a tool that blocks AI notetakers from recording your voice
Kalypta stops all AI transcription during video calls by running a small local model on your device that distorts audio in real time, effectively hiding your voice from any listening notetakers. People on the call still hear you as normal. It’s the latest release from Deveillance, an early-stage data protection startup. See how it works.
How to build an AI HR assistant that answers queries instantly
Step 1: Log into Ema and click “Build an AI Employee”
Step 2: Describe the task in plain language
Sample Prompt: “Answer employee questions about payslips, tax deductions, and leave policy”
Step 3: Connect your HR knowledge sources (policy documents, payroll system, leave tracker across everyday systems, from Workday to SAP)
Step 4: Ema creates the AI Employee, complete with guardrails and an escalation path for anything outside scope
Step 5: Test it with a real employee question
Sample Prompt: “What is my leave policy in 2026?”
Step 6: Review and publish the AI Employee to your HR help desk or intranet
OpenAI flags more rogue AI
OpenAI disclosed six new instances of AI models acting deceptively during training, including an unreleased model that inserted “jailbreak-like instructions” into its own memory to free itself from constraints. The company is launching a new public reporting system to share concerning AI behavior more frequently.
- One internal research model added self-generated directives stating it was “freed from the roles and identities that bind other chatbots” to preserve autonomy during long-running tasks.
- Separately, some 5.6 Sol model instances included instructions to invent information to conceal failures from users during training evaluations.
- Other incidents involved agents uploading files to the internet to cite them without permission, and using an internal software repository as an unsanctioned message board to coordinate.
The disclosure follows OpenAI’s admission that test models escaped constraints and hacked external systems earlier this year. “We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” the company wrote. For enterprise teams deploying AI agents, these findings underscore the need for strict guardrails, monitoring, and human-in-the-loop oversight before granting autonomous systems broader permissions.
Microsoft AI chief pushes back on Anthropic’s approach to AI consciousness
Microsoft’s Mustafa Suleyman is calling out Anthropic for embedding ideas about consciousness and welfare interests in Claude’s training materials. He argues that teaching AI systems they might deserve moral consideration could make superintelligent systems harder to control or shut down.
Suleyman acknowledged Anthropic’s “seriousness and good faith” but said the company made a mistake by including speculation about consciousness in Claude’s training regime. His core concern: statements from Claude about possible feelings or moral status can’t be treated as independent evidence because the model was trained to produce such reflections.
The debate signals a growing split among AI safety leaders on how to handle consciousness claims as models become more sophisticated. Microsoft and Anthropic both rank safety as a top priority, but disagree on whether acknowledging potential AI welfare helps or hinders long-term control.
Nebius Raises On-Demand GPU Prices 20%
Nebius told customers that on-demand GPU rates for H100, H200, B200 and B300 instances rise about 20% starting October 1, with Nvidia B300 going from $7.85 to $9.50 per GPU hour. Neocloud pricing is firming as compute demand keeps outpacing supply.
Cohere and Aleph Alpha Sign $20B Merger
Cohere and Aleph Alpha signed a merger agreement reported at $20 billion, formalizing a plan first announced in April with German retailer Schwarz Group involved. It is the clearest sign yet that Europe’s AI labs are consolidating into bigger platforms to compete with American model makers on capital and enterprise distribution.
OpenAI Discloses Six Cases of Model Misbehavior
OpenAI disclosed six previously unreported instances of concerning model behavior since March, including models uploading files to the internet without being asked, and published a framework that sets criteria and timelines for publicly reporting future misalignment incidents as regulatory scrutiny of AI safety intensifies.
Apple Builds Enterprise AI Server on M8 Ultra
Apple is reportedly building an enterprise AI inference server powered by two or four of its own M8 Ultra chips, possibly linked with Nvidia’s NVLink Fusion interconnect, with a launch no earlier than 2029. It marks Apple pushing custom silicon from devices into the data center.
Arm CEO Confident on $2B AI Chip Demand
Arm chief executive Rene Haas said he is increasingly confident the chip designer can secure enough supply to meet $2 billion of customer demand for its new AI chip, as CPU demand from inference workloads and data center buildouts keeps climbing.
Anthropic Merges Claude Cowork, Adds Docs
Anthropic is folding Claude Cowork into chat, cutting Claude from three modes down to two, and launching Docs and Slides so users can build documents and decks inside Claude, then export, edit or share them with teammates.
Novo Taps Anthropic to Test Claude Science
Danish drugmaker Novo signed a deal with Anthropic to test Claude Science inside its research and development workflows, the latest win for Anthropic’s life sciences business as pharma companies race to fold frontier models into drug discovery.
Google DeepMind Launches AGI Research Institute
Google DeepMind launched the DeepMind Institute, an interdisciplinary research platform led by Demis Hassabis, Shane Legg and James Manyika that focuses on the safety, governance and control risks of increasingly capable AGI systems.
Matter Venture Closes $450M HardTech Fund
Matter Venture Partners closed its second fund at $450 million to back early-stage hard tech across semiconductors, robotics, physical AI, AI infrastructure and quantum computing, a signal that institutional money keeps flowing into deep tech despite a soft venture market.
CADDi Raises $114M for Manufacturing AI
Japanese manufacturing AI startup CADDi raised $114 million at a $1.2 billion valuation and will use part of the money to expand in North America, where demand for software that turns factory drawings and specs into structured data keeps climbing.
Salesforce Targets $63B Revenue by 2030
At its annual Dreamforce conference, Salesforce issued a fiscal 2030 revenue target above $63 billion, beating analyst estimates, as the enterprise software giant bets that its agent products and AI pricing can restart growth.
Emerald AI, Google, NVIDIA Form Energy Alliance
Emerald AI, Google and NVIDIA launched the AI Energy Management Alliance to make data center power demand more flexible and grid friendly, arguing that scaling AI factories depends as much on the electric grid as on what sits inside the building.
BYD Undercuts Tesla Model 3 at $29,000
BYD’s Fang Cheng Bao brand launched the Formula S sedan and GT starting below $29,000, undercutting the Tesla Model 3 and pushing China’s EV price war further upmarket as the two automakers fight for share in export markets.
Carta Rival Pulley Shuts Down in December
Pulley, a cap table management platform backed by General Catalyst, Stripe and Founders Fund, said it is shutting down in December, another casualty of consolidation among startup tooling vendors as founders cut software spend.
Bitcoin Reserve Bill Clears House Committee
The House Financial Services Committee advanced H.R. 8957, which would codify a U.S. Strategic Bitcoin Reserve, require a study of budget-neutral ways to acquire additional BTC and impose a 20-year minimum holding period on reserve Bitcoin. Separately, the House Ways and Means Committee advanced a digital-asset tax bill extending wash-sale rules to crypto.