Sonnet 5 and Claude Code reshape the frontier

A new Sonnet changes the calculus

The biggest story of the week is Claude Sonnet 5, Anthropic's latest model that closes the gap with Opus 4.8 while pricing at $2 per million input tokens through August (standard $3 thereafter). The model ships with a native 1M-token context window and adaptive thinking enabled by default, making it the most agentic Sonnet yet. Claude Code immediately adopted it as the default model, and the Platform API got the full treatment — 128k max output tokens, full tool support, and introductory pricing. The launch is paired with a detailed post on Fable 5's cyber safeguards and a proposed Cyber Jailbreak Severity framework, plus a HackerOne program for security researchers.

Claude Code: from tool to platform

The defining release of the week is Claude Code itself, Anthropic's new agentic coding assistant that can autonomously execute code, understand deep codebases, edit multiple files, and recover from errors. The proof of concept arrived fast: the Government of Alberta used Claude Code to scan 466 million lines of code across 1,280 applications in 20 hours — work estimated at 6.5 years with traditional methods — identifying and fixing security vulnerabilities across government infrastructure.

Alongside the announcement, the tooling itself matured significantly. Version 2.1.200 changed the default permission mode to Manual across CLI, VS Code, and JetBrains, and fixed a critical bug where background sessions silently stopped after sleep/wake. Version 2.1.201 removed the mid-conversation system role for Sonnet 5 sessions. Earlier in the week, v2.1.199 fixed a daemon self-kill loop on Linux and improved stacked skill invocations, while v2.1.198 made Claude in Chrome generally available and added retry logic for transient network drops. v2.1.196 fixed a critical bug where waking a background job deleted the conversation and added security hardening around MCP server spawning.

Science and safety

In a parallel track, Anthropic launched Claude Science, an AI workbench for reproducible research integrating over 60 curated skills for genomics, proteomics, structural biology, and cheminformatics. It runs locally or on HPC clusters and integrates with NVIDIA's BioNeMo Agent Toolkit. The Platform API also added a new agent-memory-2026-07-22 beta header with stricter memory listing behavior, and Managed Agents gained event deltas, backward pagination, and session overrides.

On the safety front, Claude Fable 5 and Mythos 5 were redeployed globally after export controls were lifted, with an improved safety classifier blocking a reported bypass technique in over 99% of cases. Anthropic also published model entitlements for Enterprise plans in beta, giving admins granular control over model access and effort-level settings.

OpenAI rounds out the week

OpenAI shipped less headline news but meaningful updates. GPT-5.5 Instant Mini became the new fallback model for rate-limited users, with improved intent tracking and reduced repetitive responses. The iOS Codex client gained task management from conversations with branch filters and SSH support. ChatGPT for PowerPoint went GA for Business workspaces, followed by Enterprise and Edu the same day, while token-based pricing took effect for Excel/Sheets and Workspace Agent runs. The Codex CLI patched a WebSocket trace-log exposure, and the .NET SDK bumped to 2.12.0.