CEREBRO

Opus 5.5 and model economics

Claude Opus 5.5 is out: Anthropic claims about 40% lower cost than Opus 5 at similar capability, with external evals run before release.
Theo on why Opus 5.5 feels unlimited: Fable was capped at 50% of a Claude Code subscription's usage, and Opus 5.5 High costs over 2x less than Fable 5.1 High, so the headroom comes from both quota and per-task cost.
Arav Srinivas on Opus 5.5 vs Fable 5.1: after moving 50-100 orchestrator workflows from Fable 5.1 to Opus 5.5 he saw minimal differences, and he is asking which tasks still favour Fable.
Claude Opus 5.5 official prompting guide: the vendor's own guidance for the new default model, worth diffing against your current prompts and CLAUDE.md.
How to choose which model for Planning VS Implementing VS Review: a thread on per-phase model routing, which matters more now that the Opus/Fable cost gap is this large.
I let Opus 5.5 run a business on its own for a week: an unsupervised run that produced €0.00 and 245 dead ideas, with the AI manager ghosting the team on day 1, so it works as a failure-mode catalogue for autonomy.

openrig: defines a Claude Code plus Codex team in YAML and boots it with one command, with a lead agent coordinating specialists across teams.
Microsoft Research paper on Agensh (via omarsar0): runs 1K+ coding agents with no central orchestrator, coordinating through shared state instead, which is a direct test of flat versus hierarchical agent design.
I run up to 6 Claude Code sessions on the same codebase: an open-source control room for keeping parallel sessions from colliding.
Git worktrees solved our parallel-agent file conflicts. The test environment was harder.: worktrees isolate files but not ports, databases or fixtures, so per-agent test environments are the next bottleneck.
Open-source workspace where coding agents control the app through a CLI: exposes the app's own actions as a CLI so agents drive it directly, a pattern for making SaaS agent-operable.

Is Vibe Coding Safe? (via thesupermannx): a paper testing 12 coding agents on real tasks finds they write code that works but is often insecure, so functional pass rates overstate readiness.
Coding agents fix bugs where the error shows up, not where it starts: agents patch the symptom site instead of tracing to the root cause, so review fixes for locality.
How do you keep large test suites from overwhelming AI coding agents?: a context-budget problem, because full-suite output floods the window and buries the failing signal.
"You are a senior engineer" changes how sure the answer sounds, not how right it is: persona prompts shift tone and confidence rather than correctness, so don't rely on them as a quality lever.
Simon Willison on agents making software engineering harder: the gains are real but unlocking them demands extraordinary discipline and knowledge.
Simon Willison on dropping CLAUDE.md stubs: he can stop creating CLAUDE.md files that only contain @AGENTS.md, which suggests AGENTS.md is now read natively.

Codex 0.158.0: the fullscreen TUI gains configurable copy-on-select and right-click paste, and copied transcript selections now keep Markdown formatting.

Transformer Inference Arithmetic, part 6 (wafer_ai): kipply's approximate cost model built from per-token work, bytes moved by the GPU and inter-GPU communication, a starting point for reasoning about latency and cost.
Xiaomi MiMo-V2.6-RL-oss: an open RL dataset of 7,780 samples (12.1 GB) across software engineering, cybersecurity, web dev and knowledge work, with coding tasks verified by executable tests.
2026 in LLMs (so far): Simon Willison's annotated WeAreDevelopers keynote, a chronological pass over the year's key trends, useful as a catch-up read.

Harmony: AI agents that resolve IT and HR tickets inside Slack and Teams.
Claude-powered field operator for a solar company: an operator run through four tests on a real inverter swap, a concrete non-software agent deployment.
Vibe-coded live AI party installation: a non-coder built it with Claude Code, with a local LLM listening to the room and deciding what a laser wall shows.