One big change, two smaller ones inside Claude Code, and a tool at the end you can go try right now without touching the terminal.
Opus 5 becomes the default Opus model
On July 24, Opus 5 shipped and immediately replaced Opus 4.8 as Claude Code’s default Opus model (version 2.1.219). Same price as 4.8 too: $5 per million input tokens, $25 per million output, with fast mode still at double that for roughly 2.5x the speed. On the benchmark side, Anthropic is pointing at new highs on its coding benchmarks and, more striking, a 3x score jump over competitors on ARC-AGI 3, which is built specifically to test novel problem-solving rather than memorized patterns.
Take the benchmark numbers with the usual grain of salt, they’re picked by the people releasing the model. What’s actually useful here is the pattern underneath the announcement: this is the second default-model swap in under three weeks. Early July, Sonnet 5 took over as Claude’s overall default model. Now Opus 5 takes a similar role, just scoped to the Opus tier specifically, a different “default” than Sonnet’s, but the same underlying habit. If you’ve pinned a specific model in a script, a CI job, or a custom skill’s frontmatter (model: claude-opus-4-8), it’s worth checking whether you actually meant “whatever Opus is best right now” or specifically 4.8. Defaults like this can shift without warning, a few weeks apart.
What’s new in Claude Code
Two changes that make most sense read against what happened a week earlier: one gets /code-review out of your way, the other hands nested subagents back without removing the guardrails.
/code-review moves to the background
Claude Code 2.1.218 changed how /code-review runs: instead of taking over your conversation while it works, it now runs as a background subagent. It reviews whatever slash commands you’ve stacked up, rather than whatever happens to be on screen at that moment. You can kick it off and keep talking to Claude about something else in the same session instead of watching a review scroll past for a few minutes.
Worth reading together with last week’s change: /code-review, along with /verify, already stopped invoking itself automatically back in 2.1.215. So the two changes stack. It won’t run unless you ask, and once you do ask, it no longer blocks you. If you’ve been skipping /code-review on a big diff because it used to eat your whole session for five minutes, that reason is gone now.
More room to nest subagents
Claude Code 2.1.217 capped concurrently-running subagents at 20 by default and, in the same release, turned off nested subagent spawning entirely: a subagent could no longer launch its own subagents unless you explicitly raised CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH. A few days later, 2.1.219 partially walked that back: nesting is on again by default, now up to depth 3.
If you build orchestrator-style skills, where one agent delegates to a few specialists that might themselves delegate further, you get that structure back without extra config. But the safety net from 2.1.217 didn’t disappear: the 20-agent concurrency cap is still there, and --max-budget-usd now actually halts running background agents once you hit the limit, instead of just refusing new ones. If you let Claude Code run semi-unattended overnight on a big refactor, that budget cap is the setting worth double-checking before you do, not the depth limit.
To end on: ask Claude how AI is actually being used
On July 22, Anthropic shipped a connector for the Anthropic Economic Index. It’s built from millions of anonymized Claude.ai conversations. Turn it on from the connectors menu in claude.ai and ask plain questions like “which occupations use AI the most” or “what do teachers actually use Claude for,” and it answers from the real dataset instead of guessing.
Worth trying even if you have no research reason to. It’s free on every plan, including the free tier. What you get back is a straight, sourced answer about how people are actually using Claude, occupation by occupation and task by task, grounded in real usage data rather than a hot take from either extreme of the AI-and-jobs debate. The catch, which Anthropic states upfront, is that it only reflects Claude usage and not the whole labor market. We intend to come back to this connector in a few months and put exactly the same questions to it again. The difference between the two readings will say something a single snapshot can’t – while still only ever showing us Claude users.