Anthropic dropped Opus 4.8 yesterday. The benchmark gains are real but not the story. The story is one paragraph in the release notes that most coverage skipped over.
Claude Code now has a primitive called dynamic workflows. Here is what it does.
One Opus session reads your task, decides how to decompose it, spawns hundreds of subagents in parallel, audits each one's output before accepting it, and merges everything into one delivered artifact. For code, that artifact is a PR ready for human review.
The example Anthropic flagged is a codebase migration spanning hundreds of thousands of lines, taken from kickoff to merge inside a single Opus session. That is the test prior subagent patterns could not pass without you hand-stitching the orchestration code yourself.
Why this matters more than the benchmark bumps:
The model is now the orchestrator. You stop being the project manager for your own subagents. The session still runs inside Claude Code's permission model with hooks and gates intact, so safety boundaries hold. What disappears is the spec-each-subagent step that ate the most setup time.
A few other things worth holding in your head:
Pricing is unchanged from 4.7. Same dollar per million tokens.
Fast mode (the 2.5x speed variant) is now roughly three times cheaper than equivalent settings on prior generations.
Agentic coding score moved from 64.3 to 69.2 percent.
Honesty improved measurably. Anthropic says 4.8 is about 4x less likely than 4.7 to let a code flaw pass unflagged. That number is the lever that makes the parallel subagent pattern safe to actually use.
Two Opus releases in 41 days at fixed pricing. If you've been waiting for the moment to move serious work onto Opus, this is the cleanest one in months.
I put together a 3-page briefing PDF if you want the full numbers and the cadence read. It's in the resources folder.