Three AI stories from last week. Anthropic gave Claude read-only access to bank and card accounts. Thinking Machines put out Inkling 2 Small, its first second-generation model, with open weights coming. And Perplexity launched Teammates, personal agents built for long-running work.
Claude Gets Read-Only Access to Your Bank Accounts
Image: Anthropic
Anthropic launched Personal Finance in Claude on November 5. It is a beta feature available on Pro, Max, and Ultra plans in the US. The core premise: connect your bank and card accounts and ask Claude about your money in natural language.
The connection goes through Plaid. You sign in to your bank yourself through Plaid’s interface, approve read-only access, and Claude never sees your password. What Claude can see: the bank’s name, each account’s name, type, and last four digits, balances and credit limits, and up to two years of transactions. Personal Finance cannot move, manage, or invest money.
Anthropic is specific about what the feature does and does not do. No ads. No sponsored recommendations. No credit card or loan products. The launch post walks through a worked example: asked where the money went this month, Claude adds up $4,309.37 in spending across connected accounts, with more than half going to rent and utilities and $1,024.51 to food and drink, then lists the payments due before payday and what will be left. From there, the user can follow up: “Is that more than usual?” or “What’s left before payday?” Claude can flag charges that hit twice, unexpected fees, or a bill that has crept up all year.
Conversations inside the Personal Finance hub are not used to train Anthropic’s models. That is called out specifically. The launch post also includes the caveat that Claude can make mistakes and users should verify important figures with their bank.
Anthropic calls this a first step. For bigger money decisions, it says it is exploring how Claude could work alongside licensed professionals and regulated firms.
Why it matters: This solves a real friction problem. How many apps do you open to get a clear picture of your finances for the month? Claude’s read-only access does not replace those apps, but it shortens the loop considerably. The Plaid integration is the right trust architecture: you sign in to your bank yourself, you approve access yourself, and Claude can’t move, manage, or invest your money (Source). For a first step, Anthropic kept the scope tight.
Find out more: Anthropic’s Introducing Personal Finance in Claude
Thinking Machines Releases Inkling 2 Small: 594B Parameters, Open Weights Coming
Image: Thinking Machines
Thinking Machines released Inkling 2 Small on November 5. It is the first second-generation Inkling model, with 594 billion total parameters and 17 billion active parameters.
The model is proprietary at launch, available through the company’s Tinker platform and Tinker Playground. Thinking Machines says it intends to release the full weights in the coming weeks, pending additional safety testing. Artificial Analysis lists the planned weights license as Apache 2.0, with the date unconfirmed.
Thinking Machines describes it as built for agentic coding and vision, with a “big jump” in both compared to previous Inkling models. Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, up 18 points from Inkling Small. If the weights release as planned, Artificial Analysis says it would be the highest-scoring open-weight model developed outside China based on current rankings.
The agentic benchmark gains are the most notable numbers. AutomationBench-AA rises from 5% to 46%. FrontierCode 2.0 goes from 5% to 25%. The model averages about 150,000 output tokens per Intelligence Index task, compared to about 37,000 for the previous Inkling Small.
Thinking Machines lists the model’s named capabilities on the product page: general intelligence, agentic coding and tools, controllable effort (adjustable thinking time), forecasting with calibrated confidence, and instruction following. It is fine-tunable on Tinker.
Why it matters: The parameter count is less interesting to me than the agentic benchmark numbers. AutomationBench-AA going from 5% to 46% is not an incremental gain. If that holds in real workloads and the weights ship under Apache 2.0 as planned, I think this could be the open-weight model to watch for agentic work. The Apache 2.0 license is what makes that practically useful: you can run it, fine-tune it, and deploy it without negotiating terms. The safety testing hold before releasing weights is the right approach. That is a door worth opening carefully.
Find out more: Thinking Machines’ Inkling model page
Perplexity Launches Teammates: Agents for Ongoing Work
Perplexity launched Teammates on November 4. The description: personal agents that take on long-running work. They plan and coordinate projects, switch gears as priorities change, and present results for review. They also monitor ongoing work and act without waiting for a fresh query.
Teammates build on Perplexity Computer and work inside it with your tools and files. Each teammate runs in a separate environment, with browser sessions kept separate from other teammates, and saved credentials are encrypted. Users can review completed work, take over the desktop, or redirect a teammate mid-task.
Perplexity walked through use cases across marketing, sales, customer support, and operations. In the marketing example, a teammate manages social posts, email campaigns, and ads, monitors competitor promotions, and drafts fixes when results slip. In the sales example, a teammate finds stalled conversations, prepares follow-ups for review, and flags support issues that need resolution before outreach. Teammates can be added to a Slack channel so colleagues can hand off work or ask for updates.
The coordination layer goes further than a single agent: teammates can pass work to each other. Perplexity describes a scenario where one teammate finds bugs on a website and passes its findings to another to prepare fixes.
Availability: Teammates are launching on the web and soon on iOS for Perplexity Max subscribers. Free and Pro subscribers can upgrade through their account settings.
Why it matters: The difference between a task automation and what Perplexity is describing here is project memory. Teammates remember goals, task history, and shared files across runs. Each new task builds on what the previous ones established. That is a different product category than a workflow builder or a one-shot agent. The Max-subscriber gating makes sense at this stage given the compute required. What I want to understand, and cannot know from an announcement alone, is how well Teammates holds up when a project deviates from the initial brief. That is where human judgment has historically been irreplaceable. If it handles that case, this is worth a close look.
Find out more: Perplexity’s Teammates announcement