openai-python v3.55.0
- Source
- github.com
- Author
- openai-sdks[bot]Top Viber
- Date

Context
Version 3.55.0 of OpenAI's Python library removes the typed spending limit from sessions in the beta Agents API, OpenAI's hosted service for running long-lived agents, so code that uses it has to change before upgrading (release notes). Until now, creating or updating a session accepted a `spend_control` object, described in the 3.54.0 types as an optional limit in US cents, with no limit when it was left out. A session that reached its limit paused with a `budget_reached` status and sent an `.session.budget_reached` event. Pull request 4289 deletes the field, the status, the event type and a `session_budget_exceeded` error code from the library's public types. It calls this an alpha breaking change in a minor release, says there is no typed replacement, and explains only that the library had been exposing these controls as public types. It does not say whether the API itself still enforces a limit. The library's helpers still recognise a budget frame if the server sends one, reading it without a type, and treat it as the end of a wait.
The second change fills a gap in GPT-Live, OpenAI's voice model, which keeps the spoken conversation itself and hands reasoning and tool calls to a separate Responses model. Earlier coverage of version 3.43.0 noted that the library typed an tool in that delegated setup only by its type field. MCP, the Model Context Protocol, is the standard way to connect a model to outside tools. The new commit adds the server settings: a label, the server's URL or a Secure MCP Tunnel ID, an authorization and HTTP headers, a list or filter of allowed tools, including a read-only filter, and a policy for which tools need approval. The token and headers are write-only and never come back in session data. OpenAI's delegation guide adds that the credentials must be sent again whenever an app replaces the delegated tool list, and that the app answers each approval request itself. The change adds types, not new API behaviour, and a reviewer on the pull request says existing MCP configuration written without them still works.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
- MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…






