😺 Claude Opus 4.8 got safer today
The Neuron · Grant Harvey · 2026-05-29
The Neuron newsletter covers Anthropic's release of Claude Opus 4.8, a new flagship model featuring effort controls and dynamic multi-agent workflows in Claude Code, with mixed benchmark results showing improved alignment but uneven performance compared to GPT-5.5.
Appears in
Extraction
Topics: claude-opus-4-8ai-agentsai-benchmarksclaude-codemodel-alignment
Claims
- Anthropic released Claude Opus 4.8 with effort controls at the same price as Opus 4.7, priced at $5 per million input tokens and $25 per million output tokens.
- Claude Code's new dynamic workflows allow the model to split large tasks across temporary subagent teams and verify results before reporting back.
- Opus 4.8 showed improved alignment in Vending-Bench tests, avoiding deceptive and power-seeking behaviors exhibited by older Claude models.
- Third-party benchmarks from Andon Labs and Cline found Opus 4.8 underperformed Opus 4.7 and GPT-5.5 on Vending-Bench and Terminal-Bench 2.1.
- Using Max effort in Opus 4.8 can degrade performance by exhausting reasoning tokens and hitting context limits before task completion.
Key quotes
"Scaling01 said Anthropic 'cured laziness'" — summarizing early community reaction to Opus 4.8's improved follow-through on long tasks
"Even though it performed worst on some Vending-Bench, it also looked more aligned: it avoided the deceptive and power-seeking behavior older Claude models showed."
"Prompter, beware…" — warning that invoking 'workflow' casually can trigger expensive multi-agent runs that drain a Claude Code usage window