Anthropic ships Opus 4.8. How much better is it?

Hook

Open Claude today and you're running on Opus 4.8. Anthropic shipped it this morning. Same per-token price as the model you used yesterday. Fast mode runs 2.5x faster at one-third the cost. And the new model is roughly four times less likely than Opus 4.7 to let flaws in its own work pass without flagging them.

The news

Anthropic released Claude Opus 4.8 on May 28. Headline numbers from the announcement:

• 84% on Online-Mind2Web (the agentic browsing benchmark). Beats Opus 4.7 and GPT-5.5.
• First model to clear 10% on the all-pass standard of the Legal Agent Benchmark.
• ~4x less likely than Opus 4.7 to let coding mistakes slip past without comment.
• Standard pricing unchanged: $5 per million input tokens, $25 per million output tokens.
• Fast mode runs 2.5x faster at $10 / $50 per million tokens (three times cheaper than the prior fast mode).
• Effort control ships on claude.ai and Cowork. Pick how hard Claude thinks per task.
• Dynamic workflows in Claude Code (Max, Team, Enterprise) let one session run hundreds of parallel subagents.
• Mythos preview coming "in the coming weeks." A higher-intelligence model class currently locked to cybersecurity testing.

What it means for non-technical CEOs

You didn't have to do anything. The Claude your finance lead, your hiring manager, and your sales ops person opened this morning is materially better than the one they used yesterday. No upgrade, no migration, no procurement call.

Three things actually changed for your week:

The output your team trusts gets less subtly wrong. The "four times less likely to lie about its own code" line is being read as a developer thing. It isn't. It applies to anything Claude produces. Financial models. Hire / no-hire memos. Vendor diligence summaries. The honesty improvement means fewer answers that look right and aren't.

Fast mode is a cost story, not a speed story. Per-token price on fast mode dropped 3x today. If you have an FP&A analyst pushing thousands of cells through Claude, or a hiring ops person ranking hundreds of candidates, that's a line item that just got cheaper.

Effort control is the new dial on the dashboard. On claude.ai you can decide whether Claude thinks for ten seconds or two minutes on a task. The default is fine for most things. The dial matters when the answer matters more than the speed: a board memo, a hire, a deal post-mortem.

The bigger story is the Mythos tease. Anthropic is openly saying a more capable model class is weeks away. The cadence from 4.7 to 4.8 to Mythos is the real pattern to read.

What to do next

Open claude.ai yourself and run today's hardest task through Opus 4.8 twice. Once on fast mode. Once on default with effort turned up. Read the difference yourself before you decide which mode goes in your team's playbook.