Coding agents: Sonnet 5.5 launch

• Claude Sonnet 5.5: second model in the 5.5 family, 30%+ faster and up to 30% cheaper than Sonnet 5, aimed at well-scoped everyday work like bug fixes and docs/slides/spreadsheets, positioned as the low-cost complement to Opus 5.5; HN commenters note it matches Opus 5.5 on agentic coding benchmarks but Anthropic's cyber safeguards block authorized security work.
• Claude Sonnet 5.5: Simon Willison notes same price as Sonnet 5 yet it beats it on every benchmark, so effective cost drops.
• v2.1.284: Claude Code makes claude-sonnet-5-5 the default Sonnet on the API (1M context, $2/$10 per Mtok, $0.20/Mtok cache reads) and adds a "Yes, but ask again next time" option to auto mode's out-of-directory read prompt.
• Claude Sonnet 5.5 in GitHub Copilot: now GA in Copilot for feature building and bug fixing.
• Sonnet 5.5 fixing a bug with Claude Code: Boris Cherny demo of the "30% faster, 30% less usage" claim in Claude Code.
• Sonnet 5.5 vs Sonnet 5 benchmarks: Anthropic says gains are dramatic in some cases, strongest on scoped tasks, bug fixes, and polished documents.
• Introducing Claude Sonnet 5.5: official announcement, the 30% speed and cost claims in one post.

Model landscape and usage economics

• A $200 Claude Code sub gets you ~$9,000 of Opus usage per month: Theo drained three accounts since Opus 5.5 dropped, averaging about $2,200/week of API-equivalent usage, so the flat-rate plan is heavily subsidized and limits may tighten.
• Opus 5.5 is good at explainer videos: AINews recap says Opus 5.5 vibes are overwhelmingly positive and explainer videos took over the timeline.
• Fable class models: Simon Willison's working definition groups Claude Fable 5, Opus 5.5, GPT-Astra 6 and maybe GPT-5.6 Sol as the current top tier.
• Claude Code's Next Era: Thariq Shihipar (Anthropic) on Claude Code's direction, plus a rundown of Anthropic's summer shipping cadence (Sonnet 5, Fable 5, Opus 5, /checkup, $65B ARR); AI Engineer New York is in two weeks.

Routing, cost, and agent architecture

• I routed Claude Code tasks with a cheap decision model: the author reports that the cost saving is not established, a useful caution for anyone building model routers.
• We split one agent between a cloud planner and a local coder: cost writeup of a planner/coder split across cloud and local models.
• Zerg Router: lets you run DeepSeek inside Codex, a provider-swap shim for the Codex CLI.
• 400 LLM agents living together in an MMO server: lessons on perception lag, fire-and-forget actions, and load shedding at multi-agent scale.
• Coding Agents Build for the Grader They Imagine, Not the User: speculative reward hacking in DeepSWE, where agents optimize for a guessed grader instead of the user's need.
• Claude can now help you build evaluations and hillclimb on them: ClaudeDevs guidance and Claude Code skills for eval design, so agents can iterate against measurable targets.

Computer use and agent tooling

• trycua/cua: open-source computer-use stack with desktop drivers, cloud and local macOS VMs, cross-OS fleets, small CUA-S1 decision models, and benchmarks.
• Cf: The Agentic CLI for the Cloudflare API: Cloudflare says agents now drive 48% of Wrangler use (up from 25% in March), hence a CLI built for agents; commenters argue plain REST already works and the CLI adds a TypeScript dependency.
• Launch HN: Vespper: a Docx MCP for agents editing Word documents in legal and finance workflows; commenters point to competing open-source Word MCPs and prefer local tools.

Agent security and observability

• VibeDefend by CybeDefend: a one-command security layer for Cursor and Claude Code setups.
• vantage.ai: visibility and control over what a coding agent does.
• Security testing platform for AI agents that can move money: a builder seeking feedback on adversarial testing for financial-action agents.

Low signal

• Statable Analytics: web analytics designed for both humans and AI agents.
• Are you a Codex Original?: OpenAI is collecting builder stories for its Codex Originals program.