Speed tiers, browser agents, and the great SDK migration

The need for speed

OpenAI made speed the headline this week with Ultrafast mode for GPT-5.6 Sol, previewed first in the API and generating up to 750 output tokens per second — a 14× jump over Standard processing. Powered by Cerebras silicon, the tier targets workflows where latency is the bottleneck: incident response, reliability engineering, anything that can't wait for a typical frontier-model round trip. The API announcement confirms it's in limited preview for select customers, with sign-ups open for broader access. It's a notable shift in how frontier labs are competing: not just on benchmark scores but on wall-clock time, treating inference speed as a first-class product feature rather than an infrastructure footnote. The SDKs are already catching up — the Python SDK v3.1.0 adds an Ultrafast tier alongside WebSocket stream IDs.

Agents move into the browser

Anthropic's big play this week was Claude Cowork as a Chrome side panel, letting Claude read, click, and fill forms on the page you're already signed in to. The notable part is that skills, plugins, and connectors work in the browser for the first time, and sessions sync across desktop and phone — so a task started in a tab can be picked up elsewhere. OpenAI answered on the same front with smaller steps: Google Drive browsing in Library lets ChatGPT open Docs, Sheets, and Slides beside a conversation and update source files directly, while a new app directory with connectors as apps consolidates tool management in one place. Neither matches Cowork's reach into the live page, but together they show both labs treating the browser as the next agentic battleground.

Hardening the CLI and migrating the SDK

Claude Code shipped several releases this week, with v2.1.232 the most consequential: subagent forking is now on by default, letting subagent_type: "fork" inherit the full conversation and prompt cache, alongside fixes for PowerShell and Git Bash permission bypasses that could overwrite settings and redirect file access. The follow-up v2.1.233 patches an NTLM credential-leak vector in Windows device-prefix paths and removes todo tools on newer models unless explicitly re-enabled. Meanwhile, the OpenAI Python SDK v3.0.0 is the breaking release of the week: HTTPX2 is now the default HTTP client, httpx is no longer auto-installed, and anyone using custom transports must migrate or drop to a runtime-only legacy escape hatch. If you run a custom HTTPX stack against OpenAI's API, this is the week to read the migration guide.

Security, enterprise controls, and platform reach

The rest of the week was steady expansion. Daybreak cyber models landed on AWS Bedrock, giving defenders frontier security models inside existing AWS environments. The ChatGPT desktop app for Linux hit preview on Ubuntu, Debian, and Fedora, with the notable bonus of importing setup, skills, and recent work from Claude Code and Cursor. Enterprise admins got a wave of controls — audit logs in the Global Admin Console, thread-level Codex cost data, service accounts for automation, and a new ChatGPT for Intune iOS app — while an optional Computer History feature for macOS brings cross-app context to Pro, Business, and Enterprise users. The Go SDK also added model-picker discovery and plugin URL flags, and a new Codex app for macOS targets teams juggling multiple parallel coding agents. A quieter but meaningful change: project memory settings in ChatGPT are now editable after creation, letting teams scope context per project.