Opus 5.5 in Plain English

The price math

Prices are per million tokens. Opus 5.5’s headline numbers are $4 per million input tokens and $20 per million output tokens — 20% below Opus 5’s $5/$25. But the line item that matters most for agent workloads is cache reads, which dropped to $0.20 per million: a 60% cut.

Copy-paste: estimate your own bill

Here’s a small snippet that prices one agent run. Plug in your run’s token mix and it compares Opus 5 to Opus 5.5 at list prices:

def estimate(input_m, output_m, cache_read_m):
    # prices per million tokens; cache read price uses 0.50 for Opus 5 (60% cut -> 0.20)
    opus5 = input_m * 5.0 + output_m * 25.0 + cache_read_m * 0.50
    opus55 = input_m * 4.0 + output_m * 20.0 + cache_read_m * 0.20
    return opus5, opus55

# example: a heavy agent session — 2M input, 0.5M output, 8M cache reads
old, new = estimate(2, 0.5, 8)
print(f"Opus 5:   ${old:.2f}")   # $26.50
print(f"Opus 5.5: ${new:.2f}")   # $19.60
print(f"saving:   {(1 - new/old)*100:.0f}%")  # 26% on this mix

Note what the example shows: with 8M cache reads in the mix, the cache-read cut does most of the work. That’s the realistic shape of long coding sessions — the model re-reading your codebase and its own plan — which is why Anthropic’s ~40% figure for typical workloads is bigger than the 20% list-price cut.

Speed + context

Cheaper is only half the upgrade. Output tokens arrive 30% faster, the context window is 1 million tokens, and a single response can run up to 128,000 tokens. That combination is what makes long agent runs practical instead of just possible.

Speed matters differently for agents than for chat. In a chat, faster tokens are nicer. In an agent loop, faster tokens are cheaper wall-clock time per iteration — the planner-writer-reviewer cycle that agent-harness pipelines run gets through more iterations before you context-switch. And the 1M-token context window means the harness can keep the repo map, the session history, and the memory vault in context at once, while 128K of max output lets the model write the long artifacts — full diffs, test files, migration scripts — without being cut off mid-thought.

The benchmark scores

Marketing claims are free; third-party evaluation isn’t. Opus 5.5 was evaluated externally — by Frontier Design and METR — and the numbers below are the reported results.

The pattern across all three: the gains are concentrated in doing things — terminals, computers, long agentic tasks — not in trivia. That’s the workload profile this model was built for.

What it changes in your day

Benchmarks are abstract; the API bill and the terminal window are not. Here’s the concrete translation for someone who uses Claude Code every day.

FAQ

Key takeaways

1. Opus 5.5 (claude-opus-5-5, launched Sep 22, 2026) is Anthropic’s most capable model — and the first flagship upgrade that’s cheaper and faster, not just smarter.
2. The price story is the cache-read story: $4/$20 per million tokens (20% below Opus 5), but cache reads at $0.20/M — a 60% cut on the tokens long agent runs spend most on.
3. 30% faster output plus a 1M-token context window and 128K max output makes long, unsupervised agent runs practical.
4. Externally evaluated by Frontier Design and METR: 66.4% Terminal-Bench (up from 52.3), 81.8% OSWorld 2.0, 1846 Elo on GDPval-AA.
5. For daily work: agent loops fit the budget, caching becomes the obvious default, and routine terminal work stops needing a babysitter.
6. Sonnet 5.5 and Haiku 5.5 arrive within weeks — the everyday and fast tiers get the same upgrade.