AI-generated cover; illustrates the topic of this article: condensing the choice of how to pay for Claude into a single decision tree.
Summary. There are already too many reviews of Claude with contradictory conclusions: some say the subscription is worth it, others claim only the API bill makes sense, and some advise against Opus outright. The conclusions don’t add up because each one starts from different assumptions. This analysis doesn’t add another conclusion: it condenses verified rates and publicly available tests as of October 2026 into a five-question decision tree. You answer in order and get one of four answers: subscription, Sonnet, Opus, or Haiku. All prices are cross-checked with Anthropic’s official pricing page (claude.com/pricing), as of October 23, 2026. The amounts in the examples are arithmetic estimates based on those rates, not invoices: no payment test was executed for this article.
First, the name; then, the tree
Anthropicnow names by tier: Opus 5.5 (flagship), Sonnet 5.5 (default model) and Haiku 5.5 (cheap and fast), with Fable 5.1 above in price. “Claude 5.5” is not a concrete model: it’s what this entire generation from September 2026 is usually called. The rates that support the tree: Sonnet costs $2 per million input tokens and $10 for output; Opus 5.5, $4/$20. The difference is 2 times — in the previous generation it was as high as 5 — and that data is precisely what makes the tree valid: when the gap was 5 times, discarding Opus was rational; with 2 times, you have to calculate case by case.
The tree: five questions, in this order
Q1 Is your task invoked by a program or a person at the chat?
├─ Person → A1: subscription. A usage-based invoice adds
│ nothing to interactive use.
└─ Program → Q2
Q2 How many times per day is that same task completed?
├─ Fewer than 2 per day → A2: go with API, but at default rates.
│ Optimizing rates saves few dollars per month;
│ not worth the time.
└─ 2 or more per day → Q3 (from here on, Sonnet's monthly
bill is around $29 and each choice moves money)
Q3 Does the task usually work the first time or does it need retries?
├─ Frequent retries → A3: try Opus. Narrow example:
│ Sonnet costs $0.48 per pass; three passes, $1.44;
│ Opus once, $0.96. The most expensive tier comes out cheaper.
└─ Works the first time → Q4
Q4 Does the prompt exceed 100,000 tokens?
├─ Yes → A4: trim first, or switch to a model with different
│ tier rules. In Haiku 5.5, going over 100K triggers the
│ surcharge on **the entire** prompt: with 2 % more
│ content, the bill goes up 5.10 times.
└─ No → Q5
Q5 Will the prefix repeat within 5 minutes? Can the task wait?
├─ Prefix repeats → A5a: enable cache. With a 200K prefix
│ sent twice: no cache $0.80; with cache write and
│ read, around $0.52.
└─ Can wait and doesn't repeat → A5b: Batch API, everything at half price.
Screenshot of the official pricing page, taken October 23, 2026. It’s a snapshot from that day, not a permanent guarantee.
Where each number comes from
The retry math (Q3) is the only step in the tree that can reverse the tier choice. Its premise needs to be stated in full: Opus 5.5 costs twice as much as Sonnet on input and output, and with a 2x difference, the old intuition of “one retry and I’ve already lost money” stops working. That said, the example doesn’t prove Opus is more accurate: it only proves it’s worth verifying once with the logs of your own tasks. And there’s a documented limit: Opus 5.5’s thinking cannot be disabled, and at its maximum level there’s a failed public test. Simon Willison, on September 22, 2026, ran Opus 5.5 in max thinking mode twice and neither returned a result: the reasoning consumed the 128,000 token output limit, about $2.56 and nearly 20 minutes per attempt (his exact phrase: “max is effectively useless”). That’s why the correct action in A3 is to try Opus at its default level, without enabling max.
The 100K cliff (Q4) enters the tree because it’s the costliest and most overlooked mistake: the tier surcharge applies to the entire prompt, not just the overage. That is, a 101K token prompt is billed whole at the high rate, and 101K versus 99K is 2% more content but 5.10 times more expensive in that tier. On top of that, since generation 4.7 the tokenizer changed: the same text takes 1.0 to 1.35 times more tokens (official figure; Willison measured a system prompt that went from 5,039 to 7,335 tokens, 1.46 times; a 30-page PDF only went up 8%). The rate table didn’t change; your bill did.
The cache (Q5) has two hard conditions that the tree decides for you: the default TTL is 5 minutes (verifiable in the type definition of Anthropic’s official Python SDK) and the prefix must be identical byte-for-byte. If the cache expires, the $0.28 savings from the example vanishes; for workloads that can wait, Batch at half price is the most stable saving.
Screenshot of the ephemeral cache parameter definition from the official SDK. It shows where the TTL is documented, not the behavior of an actual account.
What the tree can’t answer
-
The quality difference. The tree only treats quality insofar as retries move the cost; it doesn’t assess whether one answer is better than another.
-
Your actual hit rate. The condition that reverses branch A3 is in the logs of your tasks, not in any review — including this one.
-
Invoice precision. All amounts are arithmetic estimates based on public rates; no payment test was executed and no figure should be read as an invoice.
Conclusion: after walking the tree
For whom the subscription is: if you work at the chat dialog, with few interactions per day and no desire to manage tokens, the tree exits you at Q1; the other four questions don’t affect you.
For whom the API bill is: if a program invokes the model at least twice a day, your answer is a combination of leaves from the tree, and most fall into “Sonnet by default + trim before 100K + cache or Batch at Q5”. Opus only enters if the retry math truly reverses with your data; Haiku only saves if your prompt clearly stays below the 100K line. A quick reference to get your bearings: the same task programmed and completed 2 times per day with Sonnet runs around $29 per month (0.48 × 2 × 30, no retries); below that magnitude, optimizing rates is noise.
The rates for each tier, contrasted with competitors in a single table, are in Crazyrouter’s API pricing comparison table; it’s worth checking against the official page, because each leaf of the tree must correspond to a concrete figure.
Sources note. Prices, tiers, and the 100,000 token rule cross-checked with Anthropic’s official pricing page (October 23, 2026; clickable link at the start of the article; other sources cited by name). The Opus 5.5 max thinking test and the tokenizer figure (1.46 times) come from Simon Willison’s public log from September 2026. The cache TTL comes from the type definition of the official Python SDK. The Batch discount comes from official documentation. All amounts in this article are arithmetic estimates based on public rates, not invoices; no payment test was executed. Cover generated with AI.