GPT-6 Astra lands everywhere, and voice stops being a side channel

OpenAI had one of its heaviest weeks of the year, and the releases rhyme: the company is collapsing the distance between what a model can do in text and what it can do in a live conversation, an agent loop, or a vertical product.

Astra arrives, and voice becomes a first-class surface

The centerpiece is GPT-6 Astra, billed as OpenAI's most capable model for business, with stronger reasoning, computer use, and — notably — better writing and design judgment. That last pair is a tell: this is a model pitched at deliverables, not benchmarks.

It shows up quickly in voice. ChatGPT Voice now runs GPT-5.6 or GPT-6 Astra for search and harder reasoning, with model and reasoning effort chosen through the same controls as text chat; the separate Instant/Medium/High Voice intelligence levels are deprecated. Enterprise and Edu workspaces got the same treatment alongside new Live pricing, where credit-based billing drops to 1 credit per minute from 5.

On the API side, GPT-Live 1 is generally available for full-duplex voice, letting conversations continue while a backend model or agent handles reasoning and tools — via Responses delegation or delegation to your own backend. Voice sessions run $0.05 per minute billed per second, with backend usage charged separately. A companion product announcement covers custom voices and telephony support. The SDKs caught up immediately: Python, Node, and Go all published Live API support on the same day.

Agents become a managed service

The other structural release: the Agents API entered public beta, built on a managed Codex harness that handles session orchestration, context compaction, and recovery. Durable sessions carry work across turns with progress streaming, custom tools and MCP servers, and sandboxes either OpenAI-hosted or from your own infrastructure — with a matching product-side launch. It's supported in the Python SDK's v3.13.0 and the Go SDK's v3.61.0.

Agents need guarded credentials, and OpenAI answered that too: project API keys now support configurable expiration dates, with org- and project-level maximum lifetimes admins can enforce. The Node SDK and Go SDK followed with the corresponding controls.

Verticals, workspace plumbing, and Claude Code's cleanup

ChatGPT for Financial Services pairs GPT-6 Astra with built-in financial data for research, modeling, and client-ready materials — the clearest signal yet that OpenAI is packaging models as industry products, not just endpoints. Alongside it came ChatGPT Images 2.5 and the API-side GPT Image 2.5 Sunburst and Flare, with new xhigh and max quality settings.

Most of the remaining churn is workspace-facing: the Data plugin for analyzing connected business data, Deep Research coming to Work and Codex, Box, Dropbox, and SharePoint in ChatGPT Library, plus group managers and Groups Admin API automation. The bulk of it was duplicated across Business, Enterprise, and Edu editions.

Anthropic's week was quieter but substantive. Claude Code v2.1.269 added claude plugin eval for scored, reproducible plugin test suites and /output-style switching that works over Remote Control and headless sessions. Before it, v2.1.268 fixed an HTTP 400 that broke every turn on third-party Anthropic-compatible endpoints since 2.1.265, and v2.1.267 closed a sandbox containment bypass on macOS and Linux while adding a maxEffortLevel cap that applies across Bedrock, Vertex, and Foundry.