OpenAI Gives ChatGPT Agents a Wallet and Folds Its Robotics Team Into Research as EU Leaders Hold Their First AI Summit
OpenAI launched ChatGPT Wallet, letting its agents pay for purchases with tokenized cards after a biometric check, and confirmed it has folded its stand-alone robotics team into core research. EU leaders held their first summit dedicated to AI in Dublin with Arthur Mensch and Geoffrey Hinton. Plus: Anthropic’s guide to building safety evaluations, DeepSeek’s app toolset opened through its API, and llama.cpp’s 1.0 release.
OpenAI Launches ChatGPT Wallet So Agents Can Pay Without Seeing Your Card
OpenAI moved its agents from recommending purchases to completing them on Friday. OpenAI’s introduction of ChatGPT Wallet describes it as the first wallet that lets agents pay without access to a user’s actual card details and that requires a security check for every purchase, while keeping existing card rewards and protections for eligible disputes. It was built with American Express, Mastercard and Visa and is rolling out first inside dots, OpenAI’s agents, for ChatGPT Pro subscribers in the US. OpenAI notes that 1.2 billion people use ChatGPT each week, and many already use it to research and compare products.
The flow is simple on purpose. A user adds a card once. Before each purchase, the user reviews the items, the total and the selected card, then approves with a check such as Face ID or a fingerprint scan. The payment network then issues a payment token, a substitute card number, security code and expiration date, so the business and the agent never see the real card. OpenAI’s help article on making purchases with ChatGPT Wallet adds the fine print: users must be 18 or older with a US billing address, eligible cards are US-issued American Express, Chase, Mastercard and Visa credit cards, debit and virtual cards are not supported at launch, and Wallet is not available in Chat, Work or standalone Codex threads. Merchants still handle returns, cancellations and refunds.
“An agent making a purchase should strengthen, not bypass, the trust built into payments.”
That line comes from Sherri Haymond of Mastercard. Alongside the launch, OpenAI published a Commerce Partner Playbook asking businesses to claim their listing in ChatGPT, upload product feeds, configure firewalls and CDNs to recognize signed agent requests using Web Bot Auth, and work with fraud partners such as Forter, Riskified, Sardine and Signifyd and payment providers including Adyen, Checkout.com and Stripe so legitimate agent purchases are not declined. OpenAI says it eventually expects users to pre-approve purchases within limits such as spending caps, categories or time periods.
For online retailers, the immediate risk is not fraud but false declines. If your bot protection blocks ChatGPT agents or your risk rules flag their checkouts, those sales simply fail. Ecommerce teams should check Web Bot Auth support with their CDN this month, ask their fraud vendor how agent-initiated tokens are scored, and make sure product feeds show current prices and stock.
OpenAI Folds Its Stand-Alone Robotics Team Into Core Research
OpenAI has also reorganized how it pursues robots. Fortune’s exclusive on OpenAI disbanding its robotics team reports that the company confirmed the stand-alone group is gone. Robotics researchers now sit in the main research organization under Chief Research Officer Mark Chen, while the applied robotics and hardware teams moved into the Applied organization under President Greg Brockman. A spokesperson framed the change as a commitment to robotics rather than a retreat, and Fortune notes OpenAI has said it is developing its own humanoid robot, with Sam Altman targeting a first public demonstration in 2027.
The stated reason is that general models now work on robots directly. According to Fortune, OpenAI said its researchers had assumed earlier this year that progress would build on video models such as Sora, but its frontier models, including GPT-6 Astra, proved they could be applied to robotics directly. Fortune cites an Nvidia study in which an Astra-powered agent reached a 74.8% success rate on a simulated manipulation benchmark, against 51.0% for the best specialized robotic models on that leaderboard. The timing was abrupt: OpenAI’s registered booth at the CoRL conference in Austin sat empty this week. Carnegie Mellon’s Guanya Shi told Fortune he reads the move as a sign OpenAI sees robotics as “no longer a special problem,” while MIT’s Russ Tedrake argued robot companies still hold an edge in real deployment data.
This is a bet, not a verdict. A simulated benchmark is a long way from a warehouse floor, and cloud-hosted frontier models bring latency and connectivity limits that onboard models avoid. Companies evaluating robotics vendors should ask how their stack would use a general model for planning while keeping a small, fast local controller for motion, and who owns the site-specific data.
EU Leaders Hold Their First Summit Dedicated to AI
European heads of state and government met in Dublin on Friday and, for the first time, held a strategic discussion devoted to artificial intelligence. The European Council’s summary of the informal Dublin meeting says leaders were joined by Mistral AI CEO Arthur Mensch for a session on competitiveness and AI sovereignty, focused on mobilizing public and private investment at scale, building infrastructure, securing value chains, reducing critical dependencies and ensuring clean, affordable energy. A second session with Nobel laureate Geoffrey Hinton covered fundamental rights, including children’s rights, democratic processes, labour markets, market concentration and cybersecurity.
“AI is the most important technology of our time. Europe must make the most of its benefits, from higher productivity to better public services.”
European Council President António Costa said catching up with global peers will require significant additional funding and that the next EU long-term budget, in particular the new European Competitiveness Fund, will play a decisive role. He also pointed to the AI Act as a model for global standards and announced that AI and jobs will be the central theme of the 2027 social summit. Ireland’s presidency note on hosting the informal European Council says the meeting was held at the Convention Centre Dublin, chaired by Costa and hosted by Taoiseach Micheál Martin. The meeting produced discussion rather than new spending decisions, and leaders plan to return to AI at their February 2027 retreat in Alden Biesen.
No money was committed in Dublin, so treat this as agenda setting. Still, when leaders put sovereignty, energy and market concentration on the same page, procurement rules tend to follow. Firms selling AI into European governments or regulated industries should expect more questions about where models are hosted, who controls them, and whether a European alternative exists.
Anthropic Publishes Its Playbook for Building Safety Evaluations
Anthropic shared how it thinks safety tests should be designed, run and reported. Anthropic’s guide to building strong safety evaluations borrows from psychology’s replication-crisis reforms and recommends expert-built scenarios, sandboxed runs with real-time monitoring, multi-turn tests graded on an “appropriate response rate,” preregistered metrics and adequate samples. Anthropic treats 100 items as the floor for a complex multi-turn eval and 1,000 for a single-turn one. It also says that, after its own incidents, it hardened evaluation environments and temporarily stopped running many evaluations that give models internet access.
The post is candid about weak spots. Anthropic says its single-turn harmless response rates on the API now sit between 94% and 98% for every recent model, meaning those tests are saturated and mainly useful as regression checks. In one romance-scam evaluation, Claude Opus 5 and Opus 5.5 scored 98% and 94%, but the gap was not statistically significant. Before Claude Fable 5.1 shipped, transcript reviews showed Claude sometimes implicitly validated self-harm as a coping strategy; system prompt changes moved the appropriate response rate from 60% on the API to 94% on claude.ai.
“If you only do three things, do these: write down the construct you want to measure before collecting data, validate your grader against human judgment before trusting any number it produces, and report uncertainty so readers can tell signal from noise.”
Any company running its own AI evaluations, from customer service bots to clinical tools, can use this as a checklist. The most useful lesson is that refusal rate is often the wrong metric: a bot that refuses everything scores perfectly and still fails users. Write down what a good response looks like, test long conversations, and report confidence intervals, not single scores.
DeepSeek Opens Its Full App Toolset Through the API
DeepSeek turned the features of its consumer app into developer building blocks. DeepSeek’s announcement of the full DeepSeek App experience via API says V4.1-Flash, called deepseek-flash in the API, now supports web and image search with citations, a deep thinking search mode that reads pages during multi-step reasoning, file input for PDFs, Office files and images, automatic URL reading and safety checks, across the Messages, Responses and Chat Completions endpoints, with tool calls handled on DeepSeek’s servers. Search, Fetch, OCR, speech recognition and text-to-speech are also available as standalone APIs, and the company released an open-source chat frontend demo.
Developers enable search with a deepseek_web_search tool and link reading with deepseek_read_url, and the tools carry usage fees on top of token charges. DeepSeek’s text-to-speech guide lists two voices, ten languages including English, Mandarin, Cantonese, Japanese and Arabic, up to 10,000 billed characters per request, and a list price of $0.45 per 10,000 characters, currently discounted 50% for a limited time.
“Public-facing apps must enable safety checks. Developers remain responsible for content moderation, applicable model registration, and AI-generated content labeling.”
Server-side search, file parsing and voice in one cheap API lowers the cost of building a full assistant, which will tempt many startups. Before adopting it, check data residency, your customers’ vendor policies on Chinese providers, and the mandatory safety-check, registration and labeling obligations DeepSeek places on public-facing apps. For many firms it is best used for prototypes or internal benchmarks.
llama.cpp Reaches Version 1.0 With a Stable API Promise
The open-source engine behind much of local AI hit a milestone. The ggml-org announcement of llama.cpp v1.0.0 on Hugging Face says the llama.cpp and ggml libraries both reached version 1.0.0 about three and a half years after the project began, and that the project now supports more than 150 architecture variants and all mainstream speculative decoding methods. Releases are now distributed through channels including Docker, Debian, Homebrew, conda-forge, MacPorts and Winget.
“Moving forward, both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps.”
The llama.cpp v1.0.0 release notes on GitHub list the changes: 35 deprecated functions removed, FP8 KV caches, a host-memory cache for mixture-of-experts experts with a configurable CPU share, new models including Kimi-K3.1-Pro and Ripple-65B-Preview, an OpenAI-compatible speech endpoint, and security fixes for llama-server covering SSRF, CORS and DNS rebinding, with tools and private media URLs now behind explicit opt-in flags.
Version 1.0 matters most to teams that embed llama.cpp in products, because breaking changes will now be signaled clearly. The short-term catch is that removed functions and new server defaults can break existing builds and scripts. Pin your current version, test the upgrade in staging, and review the new opt-in flags before exposing llama-server on a network.
What We Left Out
Anthropic’s IPO price range was covered on Tuesday, and social posts about its final terms and valuation were not confirmed against a filing in time. Nvidia’s GR00T 2 robotics model was covered Wednesday. Reports about internal disputes at Nvidia after Jensen Huang’s resignation, claims that Chinese models were trained on restricted chips, and rumors about the timing of a new OpenAI model were not confirmed against primary sources, so we did not use them.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →