Strong models land across the coding agents

New default models

The week's biggest shift was under the hood of OpenAI's Codex. GPT-6.1 Sol launched in Codex and ChatGPT Work as a near-Astra model aimed at complex, long-running work across code, apps, and documents at lower cost, and it became the default in the bundled catalog — plus the Amazon Bedrock Mantle and Runtime catalogs. Anthropic's Claude Code countered with Claude Sonnet 5.5 as its default Sonnet model, bringing a 1M context window and $2/$10 per Mtok pricing, with the same release repairing damaged response streams and letting MCP tool calls in resumed sessions wait for a still-connecting server.

Devin moves into the tools around it

Devin's week was about meeting developers where they already work. It is now a native agent inside Jira, starting sessions when a work item is assigned or mentioned, with automations able to run a preflight script after each trigger before the agent starts. That release also made messages sent right before a connection drop survive on the server, sent Devin Review findings on PRs back to the originating session for auto-fix, and introduced Devin Mono, a new monospace typeface for code blocks and terminals, alongside a read-only live shell preview. Separately, Devin made view preferences follow you across browsers and devices — search mode, changed-files tree layout, and sidebar filters now sync — and added Option+J/Option+K session navigation, child sessions that carry a parent summary, and pipeline log support for Bitbucket Data Center and Azure DevOps.

Tightening the sandboxes and the terminal

Claude Code shipped six updates this week, and the throughline was safety. Sandbox deny-rule bypasses got closed off so a mod's approval can no longer override deny or ask rules on nested parts of compound commands, Bash deny rules hold under auto-allow even when a variable assignment precedes the command, and a terminal freeze on pathological code blocks is gone. An earlier update caught a dangerous rm slipping through unpropmaned in a bash -c script and prompted re-authentication when an MCP server requests more OAuth scope. Credential leaks were a theme too: secret-redaction gaps closed across MCP errors, percent-encoded tokens, and URL passwords, while --resume and --continue no longer lose turns after a batch of parallel tool calls. Claude Code also opened the desktop app straight from the CLI with --desktop, added an allowedProviders managed setting, and introduced Claude Mods with a "You should know" side agent that watches for things you or Claude might miss.

Codex's CLI churn was quieter but practical: instant interrupt lets new input steer a response in progress, with prompt suggestions removed and Windows console flashing suppressed for background processes. Terminal settings now persist correctly across sessions, and queued messages resume after reconnection without duplicate sends.

Smaller fixes, same direction

Grok Build let custom agents be chosen directly with spawn_subagent and added MCP token files that are re-read on every request so rotating credentials stay fresh. A follow-up made permission rules with relative paths apply through symlinked working directories and fixed video cards getting stuck extracting posters without ffmpeg.