CEREBRO

Claude Opus 5.5

Introducing Claude Opus 5.5: first model in the Claude 5.5 family, roughly Fable 5.1 level on most tasks at 40% lower run cost than Opus 5.
Opus 5.5 cache reads 60% cheaper: on top of the 40% overall price cut, cache reads drop 60%, which matters most for long agent loops that re-read big contexts.
Opus 5.5 is the result of your feedback: cheaper per token than Opus 5 with Fable 5.1 intelligence, works across every effort level, and 5h rate limits go up with a banked reset.
Box tests Opus 5.5 on enterprise knowledge work: against Opus 5, Box saw 63% fewer tokens, 42% less verbosity and 30% faster runs on unstructured-data tasks.
Opus 5.5 vs GPT-6 Sol side by side: claims 66.4% vs 55.8% agentic coding against Fable 5.1 at $4/$20 per million tokens, with terser answers but slower on complex tasks (35 min and ~200K output tokens on one run).

AGENTS.md support in Claude Code 2.1.277: if a folder has no CLAUDE.md, Claude falls back to AGENTS.md, and you can toggle this in /config, which removes the duplicate-instruction-file workaround for cross-agent repos.
AGENTS.md is a built-in mod: it's built on Claude Code mods, the upcoming way to customize the harness, and you'll be able to build your own project-instruction variants.
Simon Willison on the mods system: points to more example mods and calls the system interesting.
Projects run from one conversation: Claude directs parallel threads that keep working after you close the laptop, in beta for select Pro and Max users in cloud sessions.
Ethan Mollick on Claude Projects orchestration: a main orchestrator spins up specialist agents and mixes expensive and cheap ones to your preference, which is the same pattern as model-tiered subagent routing.
Claude models a program, finds counterexamples, fixes bugs: it builds a model of a tricky state machine or race-prone area, finds counterexamples, reproduces them and fixes the code, without formally verifying the whole codebase.
skill-creator from anthropics/skills: drafts a skill from a described job, tests it on real prompts and compares results, installed with npx skills add anthropics/skills --skill skill-creator.

Cursor's prompt for improving harness token efficiency: a reusable prompt covering system prompt, tool definitions, request assembly, caching, compaction, retrieval and agent splitting, aimed at cheaper runs without worse results.
Context compaction measured: FutureOS vs Codex vs OpenCode: a head-to-head measurement of how three harnesses compact context, useful data since compaction quality is usually anecdotal.
Claude Code orchestration benchmarked against a solo session: the author's configured workers ran 0 times, a warning that defining subagents doesn't mean they get invoked.
How Anthropic made claude.ai 3x faster in two weeks: a write-up of using Claude to measure, debug and improve performance, with prompts and methods included.

Swipe: Agents that work while you sleep: a roundup of tools that keep coding agents running after the laptop closes, keep API keys out of chat, and let parallel sessions hand off work.
AgentsView: imports session histories from Codex, Claude Code, Cursor and 20+ other agents into local SQLite, with web and CLI search across messages, tool calls and results plus combined token and cost reports.
Graft: builds a symbol and call graph of a repo as local files (tree-sitter, no model or API key on the structural path) so agents can query callers, references and ranked files instead of using a vector DB.
Windows workspace where Claude Code sessions remember earlier attempts: open-source cross-session memory of what previous sessions tried.
codex-resume: continues OpenAI Codex chats in Claude Code as native /resume sessions.

Voice conversation with a Datasette blog backup via datasette-mcp: the ChatGPT iPhone app can now talk to an MCP server directly, which makes personal data MCPs usable by voice.
Claude with web search vs 18.5k NASA aviation safety reports as an MCP: compares open web search against a curated domain corpus exposed over MCP.
Floot MCP: builds and ships web and mobile apps from inside Claude or ChatGPT.

TradingAgents: an open-source multi-agent trading team of analysts, bull and bear researchers, a trader and risk managers who debate evidence, with inspectable reasoning.