Claude Opus 5 — the frontier stopped being the expensive one.

The pricing line worth pausing on is the one that isn't in the announcement. There is no long-context surcharge. A 900,000-token request bills at exactly the per-token rate of a 9,000-token one. GPT-5.6, by contrast, charges 2× input and 1.5× output above 272K. If your workload is "put the whole repository in the window and let it think," that is a structural pricing advantage, not a benchmark one — and benchmarks will never show it to you.

The line that quietly works against you: the tokenizer tax. Claude 4.7 and later produce roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. Per-token price is flat; token count is not. Any cost-per-task comparison you carry over from a pre-4.7 model is inflated by about 30% before you start measuring anything.

Five settings, no beta header, set via output_config: {"effort": "…"}. Anthropic's own descriptions and intended uses:

Artificial Analysis measured the whole ladder on launch day. This is the most decision-useful table in the release, because it is the only one that prices the dial:

Artificial Analysis, 2026-07-24 · captured 2026-07-27. "Cost to run the Index" is what AA paid to run the entire benchmark suite at that setting — not a per-task figure. Different denominator, kept separate on purpose.

The arithmetic that should change your defaults: 6.9× the cost from low to max, for 9.7 index points. From the default high to max it is 1.94× the spend for 1.8 points. Time-to-first-token scales 18× across the ladder while throughput stays flat at 52–56 tokens per second — you are not buying speed at any setting, only depth.

That is the genuinely counter-intuitive finding, and it is corroborated three ways. Cognition ran FrontierCode 1.1 and got Opus 5's best Main score at medium. Anthropic documents the decline itself: "a decline in FrontierCode score above high effort… Opus 5 at these effort levels [makes] more changes than the task requires (e.g., refactoring)." And on Anthropic's own FrontierBench, xhigh (44.4) beats max (43.3). On GDPval-AA v2, xhigh scores 1827 using 25% fewer output tokens than max and still beats every other model on the board.

Two mechanics that bite in production. Effort does not shorten output — Anthropic, verbatim: "on Claude Opus 5, changing effort does not reliably shorten responses, so prompt for length instead." And changing effort mid-conversation invalidates the prompt cache, because effort shapes the rendered prompt. Pick one level at the start of a long cached session and stay there.

This is the most confused question of the release, and the confusion is legitimate: the evidence genuinely splits. Every gap below falls inside the evaluators' own stated confidence intervals.

Fable 5 remains the vendor-declared ceiling and holds both the crowd crown and the official Terminal-Bench crown. Opus 5's one uncontested win is ARC-AGI-3 — 30.16% against GPT-5.6 Sol's 13.33% on the same Public Demo set — and that is the one to trust, because ARC Prize administered it, not Anthropic. Read the config even there: ARC Prize evaluated high effort only, "due to the short testing window." No max-effort ARC-AGI-3 number for Opus 5 exists.

Against Sonnet 5 ($2/$10 through August 31, then $3/$15) the split is cleaner. Per-Index-task, Sonnet 5 costs $1.53 to Opus 5's $2.03 — but scores 53.4 to 60.7, and loses badly exactly where this audience lives: FrontierBench 14.6 vs 43.5 on the public board, and Toolathlon Pass@1 74.7 vs 80.6. Sonnet 5 is the volume tier. Opus 5 is the long-horizon tier.

Secondary coverage buried these. They are in Anthropic's own summary table, and they are the fastest way to calibrate what the launch is and isn't.

• SWE-bench Pro — Fable 5 wins, 80.0 to 79.2
• DeepSWE v1.1 — GPT-5.6 Sol wins, 72.7 to 68.8
• FrontierCode Main — Fable 5 wins by 0.1 (53.5 to 53.4)
• Humanity's Last Exam, no tools — Fable 5 wins by 0.2 (56.5 to 56.3)
• HealthBench Professional, length-adjusted — Mythos 5 wins, 66.0 to 59.8
• ARC-AGI-2 — GPT-5.6 Sol wins, 92.5 to 90.42

And one row that needs its own warning

Benchmarks Anthropic published no Opus 5 number for, in either the launch post or the 194-page system card: GPQA Diamond, AIME, MMMU, MMLU-Pro, τ-bench and its variants, LiveCodeBench, and Terminal-Bench at any version. If you see an Opus 5 figure for one of these attributed to Anthropic, it is not real. Terminal-Bench appears to have been superseded internally by FrontierBench v0.1 — described as "a successor to Terminal-Bench 2.1 developed by the same team." Likewise, METR has published no time horizon for Opus 5; its page was last updated 2026-05-08, and circulating figures trace to a prediction market, not a measurement.

1. 1. Raise max_tokens. Adaptive thinking is now on by default, and max_tokens has always been a hard limit on thinking plus response text — so workloads that ran without thinking on Opus 4.8 now hit the ceiling. Anthropic recommends starting at 64K for xhigh and max.

   This is the number-one migration footgun. Code that moved over without raising the ceiling truncates mid-sentence, and the failure looks like a model regression rather than a config error.

2. 2. Delete every verification instruction you carried over. Anthropic, verbatim: "If your prompt contains explicit verification instructions ('include a final verification step for any non-trivial task,' 'use a subagent to verify'), remove them: instructions like these cause over-verification on Claude Opus 5, and removing them reduces wasted tokens with no loss in quality."

   Same for "double-check your answer," and the same for legacy harness scaffolding. The behaviour you were compensating for is gone; the compensation is now pure cost.

3. 3. Cap subagent spawning before you upgrade Claude Code past v2.1.219. Nesting depth silently went from 1 to 3 on the same day as a model Anthropic describes as delegating to subagents "more readily than prior models."

   Set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to disable nesting; CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION defaults to 200. A depth-3 tree of Opus 5 leaves at high is the documented anti-pattern — Anthropic's own effort table names low as the subagent setting, and its model matrix names Haiku 4.5. See the swarm chapter for the shape of the bill.

4. 4. Run a fresh effort sweep. Anthropic explicitly says not to inherit xhigh from Opus 4.7 or 4.8, and Cognition's data says Opus 5 peaks at medium on FrontierCode.

   This is a Ch 25 job, and it is the cheapest hour you will spend this quarter — the ladder above says the wrong default costs 1.94× for 1.8 points.