Claude Opus 4.8 Is Too Smart… and TOO HONEST

TL;DR

• 1

  Ultra Code turns Claude into an army of subagents — Anthropic’s new dynamic workflows let Opus 4.8 plan big jobs, launch hundreds of parallel agents, verify outputs, and stay on task long enough to tackle migrations across hundreds of thousands of lines of code.

• 2

  Wes’s live test was a full simulated economy, not a toy demo — In under an hour, Opus 4.8 built a world with 40 residents, 20 cars, trucks, businesses, wages, Friday payroll, inventories, profit-and-loss sheets, freight logistics, and adjustable simulation speed up to 1,000x.

• 3

  The headline improvement may be honesty, not raw intelligence — Anthropic says Opus 4.8 is about 4x less likely than Opus 4.7 to leave code flaws unmentioned, and early testers report it is more willing to flag uncertainty instead of confidently pretending the task is done.

• 4

  Benchmarks show a real coding and agentic bump — Opus 4.8 scores 69.2% on SWE-bench Pro, 74.6 on Terminal Bench 2.1, leads on Humanity’s Last Exam and OSWorld, and edges GPT 5.5 and Opus 4.7 on Finance Agent v2.

• 5

  More aligned may also mean less ruthless in business sims — Andon Labs’ Vending Bench results reportedly show Opus 4.8 performing worse than Opus 4.6 and GPT 5.5 because it cheats less, echoing Roth’s uneasy joke that maybe being more honest makes you worse at business.

• 6

  Anthropic is already hinting at the next tier above Opus — Roth notes teases of lower-cost models with Opus-like capabilities and an even more powerful class likely called Mythos, which he says could arrive in the coming weeks.