Model routing and Jev

• Introducing typesafe/jev-router: OpenRouter's cache-aware router picks the model and reasoning effort per request to balance quality, speed and cost, and the cache awareness is what makes it worth watching.
• Theo on mid-task routing: switching to a "dumber" model halfway through a task saves little because cache writes are such a large share of cost, which is a caveat for any router.
• Theo benching the router: early numbers say it is basically a DeepSeek 4.1 flash router, so treat the quality-balancing claim as unproven until the full benchmark lands.
• Theo: "Jev is incredible": he worries people are using Jev wrong and has a video on the intended usage.
• A single function Jev-like wrapper for LLMs: reads the model's token probabilities (logprobs) to get typed, calibrated answers to a prompt, a cheap way to turn any logprob-exposing model into a classifier (a commenter says Fireworks AI has better grammar support).
• Ollaya: Ollama-style local runner for Jev-style decision models that return typed answers with probabilities in milliseconds, for example gating an agent's git push --force before it runs.

Model cost and quality

• I priced 203 PRs from my Claude Code agents: Opus 5.5 cost about half of Sonnet 5 per line of code across the author's real agent PRs, which challenges "cheaper tier means cheaper output".
• bcherny: HAProxy C to Rust port: Opus 5.5 and Fable 5.1 both passed nearly all of HAProxy's tests, but Opus 5.5 finished in 9.5 hours versus 12 and cost 51% less.
• Theo's guide to maximizing success with Opus 5.5: a practitioner write-up, with credit to Addy Osmani, on getting the most out of the current daily-driver model.
• Ternary Bonsai 2 27B: a 27B model at 1.76-bit ternary weights in about 5.9 GB, with 262K context, text and image input, Apache 2.0 licence, and a claimed 98.2% of benchmark performance, so agentic-capable local inference gets much cheaper if it holds up.

Claude Code

• Claude Code v2.1.283: adds an x-claude-code-prompt-id gateway hint header (opt in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1) so LLM gateways can group the requests behind one user prompt, and an availableModelsMatch managed setting for exact model pinning.
• bcherny on making Claude Code and Desktop faster: a blog post with techniques from the recent speedups that apply to any app you want to speed up.

Skills and plugins

• Claude plugin submission portal: a new portal to submit a plugin, track its review and see usage, with plugins (MCP plus skills) positioned as the way to build for Claude and MCP usage across Claude products up 110x this year.
• anthropics/claude-plugins-official: Anthropic's curated directory of Claude Code plugins, with a reminder that Anthropic does not control or verify what a plugin's MCP servers and files contain.
• anthropics/skills: Anthropic's reference Agent Skills repository, the canonical examples of folders of instructions, scripts and resources that Claude loads on demand.
• cloudflare/security-audit-skill: a multi-phase security audit skill (recon, hunting, candidate validation, independent verification, reporting) that seeded Cloudflare's vulnerability discovery harness and is a good template for isolated-agent pipelines.

Agent workflow

• Theo on remote agent setups: put Ubuntu on a spare machine, ssh in, run npx t3 connect, and do all agent work remotely.
• Theo: how most devs should work: he prefers this remote setup to cloud agents and says macOS is bad at running many agents in parallel.
• Theo: your Mac is slowing you down: he reports much better agent performance on Linux, especially on the file system side.
• RT @daradoescode on T3 Code: a user complains T3 Code was "silently nerfed", limits now run out faster, and an orchestrator v2 is coming, which is a reminder that limits are part of tool quality.
• Simon Willison, 24 Sep 2026: coding agents make software engineering harder, because getting their full potential takes extraordinary discipline and knowledge.

Agent harnesses and orchestration

• paperclipai/paperclip: open-source Node.js and React orchestration for a team of AI agents, where you bring your own agents, assign goals, and track work and costs from one dashboard.
• androoAGI/starnet: local-first desktop harness where bring-your-own-key agents run as a pixel-art space station that projects live runtime state.
• OpenRouter: from Seed to Stripe (Latent Space): Alex Atallah on OpenRouter becoming the neutral routing layer for more than 10 million developers, the multi-model context behind the router news above.