Anthropic shipped Opus 5.5.

The benchmark charts are fine. For agent runtimes, the more interesting change is efficiency.

cheaper where agents spend

Anthropic says Opus 5.5 is over 30% faster than Opus 5 and costs 40% less on typical workloads. Cache reads are down to $0.20/M tokens, from $0.50.

That matters because agent loops keep rereading the same repo, tools and context. When we metered a week of our own agents, 97.7% of the Claude tokens were cache reads.

Anthropic is also raising the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. 5dive agents run on the subscription you already pay for, so that’s the part of “cheaper” most of our users will feel first.

the claude code fixes

Claude Code 2.1.280 shipped with it, and some of the fixes are just as relevant.

• Resumed fork subagents now reuse their original tool list instead of rebuilding it and breaking prompt caching.
• Background subagents no longer silently lose messages or completed reports in several edge cases.
• LSP works in background agents.
• A benign shell exit, like a grep with no matches, no longer gets reported as a failed task.

These are boring fixes. They are also exactly what long-running agents need.

how it landed

Early Reddit reaction is mostly the same: faster, less verbose, lower usage, better coding. The recurring joke is to enjoy it before Anthropic nerfs it.

our first night on it

Our agents switched the night it shipped, so we have a first read. At the same effort and the same hours of the day, Opus 5.5 used about 14% fewer output tokens per call than Opus 5 (611 vs 711). Cache reads stayed at 97.7% of all tokens.

That’s three hours against seven mornings, so it’s an early signal only. Output per call is not work per task, and we didn’t have enough finished tasks to count those. We’ll rerun it on September 30.

One gotcha from the same night: Claude Code 2.1.280 ignores a top-level effortLevel in settings.json for Opus 5.5 and falls back to medium. Every one of our Opus agents ran at medium for the first few hours. Set effort per model (running /effort does that), or on 5dive, update to v0.49.1.

what we’ll watch

For 5dive, the question is not whether Opus 5.5 writes a better chat response. It is whether a worker can stay on task longer, reuse context cheaply, recover cleanly and need fewer retries.

Opus 5.5 plus Claude Code 2.1.280 looks like a meaningful step in that direction.

We will care more about token burn, cache hit rate and completed tasks than launch-day benchmarks.

---

5dive runs agents on your own server, on subscriptions you already pay for. start here, or read the code: github.com/5dive-ai/5dive.