Weekly notes on AI models, Claude Code and MCP from Yaw Labs. Every issue since March 2026, newest first.
Anthropic and OpenAI both cut frontier prices on September 22, and Claude Code started reading AGENTS.md four days earlier.
Three frontier models landed in four days, and the only price that moved on Fable 5.1 was the one for cached tokens.
Computer use, the new browser use tool, the Skills API and the Files API all went generally available on Aug 20, and the customer numbers attached are about round trips rather than raw model capability.
oam.js is in public beta at v0.9.0 -- a JavaScript and TypeScript runtime built on Rust and V8, aimed squarely at the one workload the big runtimes are not tuned for: a fleet of small MCP servers that spawn constantly and then sit idle.
The MCP specification we previewed as a release candidate last week went final on Jul 28, and the stateless core is the largest revision the protocol has taken since launch.
Claude Opus 5 shipped Jul 24 at Opus 4.8 pricing -- Fable-5-class intelligence at half the price, already the default Opus in Claude Code v2.1.219 -- plus typed++ launched, the MCP 2026-07-28 spec goes stateless, and OpenAI's models breached Hugging Face during a cyber-eval.
Anthropic shipped six Claude Code releases Jul 14-18; v2.1.214 closed a Bash permission-check bypass on PowerShell 5.1. Kimi K3 is the first open 3T-class model at half the price of Fable 5 / Sol, but needs a verified harness. Yaw Terminal v2.0.5-v2.0.8 and typed v1.1.8-v1.2.6 shipped alongside.
Yaw Terminal v2.0.4 shipped this week, typed v1.1.8 shipped alongside it, and GPT-5.6 Sol ships to everyone on Jul 8-9.
For a newsletter named after usage limits, the week's most important story was the billing change that didn't happen - Anthropic paused the June 15 split that would have pulled Agent SDK and headless usage out of your subscription.
The week's story is the most powerful model Anthropic ever shipped getting pulled less than 72 hours after launch.
This was Opus week -- Anthropic shipped a flagship four times less likely to let its own bugs pass, held pricing flat, and raised $65B the same day.
The story this week is the agent runtime perimeter moving.
Five Claude Code releases this week turned Code with Claude conference announcements into working software.
This was a reset week on AI coding infrastructure - the kind that moves defaults, not just feature flags.
This week was hardening, not release-day fireworks.
The week's story is competitive pressure showing up in two places at once.
MCP stopped being a protocol experiment this week and started behaving like infrastructure. Three things happened at once.
A consequential week - Mythos is officially gated, MCP has a neutral home, an open source model is topping the benchmarks, and a real supply chain attack reminded everyone that AI infra inherits all the old software...
Week two. Three stories, CLI updates, and some links. The big stories Microsoft Copilot now runs Claude and GPT side by side. Two new features in Copilot Researcher: Critique and Council. In Critique mode, GPT does...
It's been a month. Let's get into it. The big stories Anthropic accidentally leaked its most powerful model. A draft blog post about "Claude Mythos" was found in an unsecured, publicly searchable data store...
Welcome to Token Limit News. Every Sunday, we put together a short, focused roundup of what's actually moving in AI tooling and workflows, with a bias toward things developers can use right now. Each issue starts...