On November 13, 2026, OpenAI launched ChatGPT Wallet — a payment service for purchases made through AI agents, created in partnership with American Express, Mastercard, and Visa: users link a card once and confirm each purchase with biometrics, while card details are replaced by a network token that neither the merchant nor the agent sees; the service currently works only for ChatGPT Pro subscribers in the US. Simultaneously, Anthropic opened a closed beta for self-hosted Claude Managed Agents, where the entire agent stack is deployed within the client's perimeter, and Perplexity published open-source Decider v1.3 classifiers under the Apache-2.0 license. Together, these three releases provide the missing production layers for agent pipelines: payments, controlled deployment, and cheap classification.

What happened
On November 13, 2026, OpenAI launched ChatGPT Wallet — a payment service for purchases made through AI agents, created jointly with American Express, Mastercard, and Visa. The mechanics work as follows: the user links a card once, confirms the order and amount via Face ID or fingerprint before each purchase, and the payment network issues a token replacing the card details — which neither the merchant nor the agent sees, while card rewards and transaction dispute protection are preserved. The service is currently available to ChatGPT Pro subscribers in the US with cards from American banks; a Commerce Partner Playbook guide has been published for merchants in parallel. On November 12, Anthropic opened a closed beta for self-hosted Claude Managed Agents: the entire agent stack, including the harness, sandboxes, credential vaults, and session history, is deployed in the client's cloud or data center, while Claude model inference still runs in Anthropic's cloud, and step audits are exported via OpenTelemetry; participation in the beta requires an application. On the same day, Perplexity released open-source classification models Decider v1.3 and Decider Flash v1.3 based on open-source Qwen models: weights are published on Hugging Face under the Apache-2.0 license, models are available via OpenRouter and the Perplexity Decisions API, with a claimed average accuracy of 91.6% on 63 classification tasks and a 6.54 percentage point improvement, and the locate function returns precise text fragments with offsets.
Context
Before these two announcements, AI agents faced three barriers: people were reluctant to trust programs with payment data, merchants could not distinguish legitimate signed agent requests from regular bots and scrapers, and corporations could not deploy the agent stack on their own territory when responsible for personal data before regulators. ChatGPT Wallet addresses the first barrier through infrastructure that already works in regular online shopping — tokenization by Visa, Mastercard, and American Express card networks — so a purchase made by an agent remains legally and technically a standard card transaction with familiar protection. The launch itself is not a breakthrough in model capabilities: its significance lies in creating a testable loop of real transactions where it is possible to observe for the first time how agents behave as buyers. Demand for self-hosted deployment reflects a different pressure: for a company, it is important where orchestration, secrets, and logs reside, not just the models themselves, so Anthropic offers to divide responsibility zones between the client and the vendor. The Decider release fits into a broader trend of moving away from the "one big LLM for everything" scheme: request routing, classification, and entity extraction do not require free generation, and Perplexity replaces the generative model with the probabilistic output of a narrow, cheap classifier.
Why this matters for the industry
Agent commerce gains an infrastructure layer that merchants did not have: payment tokenization via Visa, Mastercard, and American Express, combined with Web Bot Auth, a standard for recognizing signed agent requests, lowers the barrier to entry — instead of their own payment logic, merchants only need to refer to the Commerce Partner Playbook guide and prepare their storefronts for agent purchases. Self-hosted Claude Managed Agents gives companies the option to keep agent orchestration, sandboxes, and credential storage within their own security perimeter, which is critical where personal data is processed; however, request and response content still leaves the client's perimeter because inference remains in Anthropic's cloud, meaning orchestration and secrets are localized, but not the data of the conversation with the model. Open-source Decider classifiers compete with large generative models in routing, classification, and entity extraction pipelines, and open weights allow any team to re-verify the claimed accuracy on their own data. Layers are clearly standardizing: agent payments are anchored to OpenAI and card networks, the corporate runtime to Anthropic's self-hosted stack, and the defensive advantage of startups is shifting up the stack — products must be built on top of these platforms.
Why this matters for users
Only ChatGPT Pro subscribers in the US with cards from American banks can currently try purchasing through an agent, so the feature is not directly available to most Russian-speaking readers; however, the scheme itself sets expectations for future shopping: purchasing through an agent does not cancel familiar guarantees — card rewards and transaction dispute protection are preserved, and a person explicitly confirms the order and amount before each charge. Developers have more options: they can apply for the closed beta of self-hosted Claude Managed Agents and practically examine how the separation of the harness and sandboxes, credential storage, and step audits via OpenTelemetry work, to understand if such an architecture is suitable for migrating their own agents. The most accessible component is the open-source Decider classifiers: their weights are already on Hugging Face, they connect via OpenRouter and the Perplexity Decisions API, and they can be put into a pipeline for request classification, routing, and entity extraction today — run their own eval on their data, compare with previous solutions, and, if the accuracy is satisfactory, replace the expensive generative model in these tasks with a cheap, narrow component.
What is still unknown / limitations
There are currently more unknowns than answers. For ChatGPT Wallet, the reliability of the end-to-end payment pipeline, failure rates, fees, and actual dispute practices have not been disclosed, and nothing has been announced about expansion beyond the US and Pro subscription. The claimed average accuracy of 91.6% on 63 classification tasks and a 6.54 percentage point improvement for Decider v1.3 have not been independently verified: the list of tasks has not been published, there is no definition of "average accuracy" or comparison conditions with the baseline — although open weights do allow such a check to be performed. Self-hosted Claude Managed Agents exists only in a closed beta without prices or SLAs, the timeline for its expansion has not been named, and request and response content still goes to Anthropic's cloud, so this option does not provide full data localization. Finally, expectations such as the geographic expansion of Wallet, the transformation of classifier-routers into a default layer before LLMs, and the standardization of tokenized agent payments with Web Bot Auth are interpretations of the three announcements, not established facts.
Sources
- Introducing ChatGPT Wallet | OpenAI
- Operate Claude Managed Agents inside your security perimeter | Claude by Anthropic
- Decider v1.3 — perplexity-ai Collection | Hugging Face
Author
Look at AI, editorial team
