Daily AI Builder Brief - July 2, 2026

The Operator's Take

The shift that matters today isn't a single feature, it's the new default loop. Sonnet 5 is now the out-of-the-box model in Claude Code with a native 1M-token context at promo pricing through August 31, and background agents now commit, push, and open draft PRs on their own. That means cheaper, longer-context agents are shipping code unattended by default. Re-baseline your harness on Sonnet 5 now while the pricing is halved, but put a human gate on auto-PR background agents before one of them pushes something you didn't read.

Executive Summary

Claude Code 2.1.198 makes Claude in Chrome generally available.
Background agents now auto-commit, push, and open draft PRs when they finish code work.
Sonnet 5 (2.1.197) is the new default model: native 1M context, $2/$10 per Mtok through Aug 31.
New /dataviz skill ships with a runnable color-palette validator.
The /agents wizard was removed: edit .claude/agents/ directly.
Anthropic redeployed Fable 5 globally after export restrictions lifted July 1.
Anthropic proposed an industry jailbreak-severity scoring framework with Amazon, Microsoft, and Google.
Beta SDKs landed for the MCP 2026-07-28 spec RC, which drops the session handshake for a stateless core.

Claude Code

2.1.198: Claude in Chrome GA, auto-shipping background agents

Claude in Chrome is now GA, and background agents automatically commit, push, and open draft PRs on finishing code work. The Explore agent inherits the session model (capped at opus), subagents inherit extended-thinking config, and transient network drops now retry with backoff instead of aborting the turn. The /agents wizard is gone. Changelog

2.1.197: Sonnet 5 is the default

Sonnet 5 replaces the prior default with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31. Sonnet 5

Anthropic

Claude Science and Fable 5 redeploy

Claude Science, a research workbench with auditable artifacts, launched June 30. Fable 5 returned globally July 1 after US export restrictions were lifted, priced at $10/$50 per Mtok. Alongside it, Anthropic proposed an industry-wide jailbreak-severity scoring framework with Amazon, Microsoft, and Google, which matters for anyone tracking prompt injection risk. Newsroom

MCP Ecosystem

Beta SDKs for the 2026-07-28 spec RC

Beta SDKs for the release candidate landed June 29. The RC removes the initialize handshake and the protocol-level session, so a stateless server can sit behind a plain load balancer. That simplifies scaling but moves auth and MCP security fully onto OAuth/OIDC, so re-check your gateway assumptions. MCP Blog

Broader AI

Nothing today beyond the Anthropic items above.

What This Means For Builders

Test your agent workflows on Sonnet 5 now: the 1M context and half-price window close August 31.
Gate auto-PR background agents behind review before they run on any repo you care about.
If you run remote MCP servers, plan the stateless migration now: sticky sessions stop being a protocol guarantee.
Swap the removed /agents wizard for direct .claude/agents/ edits in your setup docs.