Releases Index

BetaWeekly digests are a beta — we're trying something new. Feedback welcome.

Claude Code and Codex tune models while Devin rebuilds its session view

October 26 – November 1, 2026

AnthropicOpenAICognition
Anthropic, OpenAI, Cognition

Claude Code began retiring Opus 4.8's fast mode and handed subagent effort its own dial, Codex CLI 0.166.0 brought GPT-6.1 Terra to every model catalog along with a Vertex AI provider, and Devin merged its session tabs into one live Activity stream.

The week at a glance

Replay the week
FeaturesEnhancementsFixesMajorLandmarkTile size reflects impact

Model choices shift underfoot

The week's biggest story is a deprecation with a deadline. Claude Code deprecated fast mode for Opus 4.8, which will be removed on November 30, 2026 — after that, requests that ask for it return an error rather than quietly falling back to standard Opus 4.8. Migrating to Opus 5.5 keeps the speedup, so teams that lean on fast mode should plan the switch rather than discover it. The same work gave subagents their own effort setting via CLAUDE_CODE_SUBAGENT_EFFORT, so a session's effort no longer has to be inherited wholesale by everything it spawns, and added pooled group budgets and layered policy mode for the Claude apps gateway.

Over in Codex, the model story is additive: GPT-6.1 Terra is now in the bundled, Vertex AI, and Amazon Bedrock catalogs, following its first appearance a couple of days earlier. Codex also added a built-in Google Cloud Vertex AI provider configured with model_provider = "gcp", turned on memory v2 for fresh profiles, and let reasoning effort change mid-conversation by default — plus closed a marketplace URL validation gap that could have let Git connect to a different host than the one checked.

Claude Code tightens the guardrails

Beyond the model news, Claude Code spent the week hardening how the Write tool behaves. It now refuses writes outside the working directories in every permission mode via permissions.blockWritesOutsideWorkingDirectories, and it stops Write from clobbering large CLAUDE.md files — previously a very large rules file or @-included file could be overwritten after only its beginning was read, and Claude is now made to read the whole file or edit in place instead. Sessions also got an end to turn-after-turn "Invalid effort level" failures when an admin caps a model's effort, and idle compaction for long 200K+ token conversations parked in VS Code and Claude Desktop.

Reliability fixes rounded out the week: startup and post-edit freezes traced to very long CLAUDE.md list items are gone, standard-context sessions that crossed 200K tokens now compact and continue instead of stopping on a 1M-context billing prompt, and interrupt latency during stalled MCP reads, out-of-order transcript messages in claude -p and SDK sessions, and a Write/Edit stall on huge files were all addressed.

Devin folds its session view into one stream

Devin reworked how you watch a session work. It merged Progress and Shell into a single Activity stream with search, type filters, and a pinnable details drawer; rows now fold under the user message they respond to, and shell cards list the files each command edited. HTML artifacts render inline from a new component kit matching the app's design system, and the Terraform provider gained an on-call responder resource and schemas data source. A couple of days later it retired the separate Tasks tab in favor of a Progress tab that shows the task list as a live drawer — click a task and the Activity stream filters to that work — alongside one-press jump to the end of the transcript and Bitbucket Data Center PR decline support.

Codex polishes the terminal

Codex CLI's second release of the week was mostly about navigating long sessions: keyboard-driven transcript selection with word, line, and page motions, integrated search, and configurable shortcuts, plus a repository filter in the session picker and directory grouping in the agent command center for people juggling linked worktrees. Approval settings now survive resume and fork, and local image attachments come back with restored composer text.

AI-generated digests may contain mistakes.
Releases covered8