Opus 5.5 cuts cache reads 60%, reroutes some calls to older models

Anthropic's new flagship is 20% cheaper across the board and stretches Pro/Max limits ~25% further — but safety classifiers can swap your model mid-run.

Nowline SEP 25 9:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Cache reads drop 60% — agentic work gets cheaper

    Opus 5.5 lists at $4/M input and $20/M output (both 20% under Opus 5), but the real move is cache reads at $0.20/M — a 60% cut, with cache writes at $5/M. Since replayed context dominates the bill for long coding agents, sustained sessions get materially cheaper.

  • Your agent calls can silently reroute to an older model

    Anthropic's safety classifiers "fall back transparently" mid-request: most cybersecurity prompts route to Opus 4.8, and biology work runs behind a Life Sciences Verification Program. If you pipe security or bio tooling through agents, a step may quietly execute on a model you didn't pick.

  • Pro and Max limits stretch about 25% further

    Anthropic raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise, and says the lower per-token cost makes five-hour and weekly caps go ~25% further. Reporting cites 3.3x longer work per prompt and 40%+ more calls each. Benchmarks: Terminal-Bench 4.0 66.4%, CursorBench 4.0 57.8%.