Before you begin
No previous experience required unless you choose a coding exercise. Use an account you are allowed to access; features vary by product and region.
What you will learn
- Separate a release claim from a result you have verified.
- Distinguish API token charges from a subscription.
- Calculate a small illustrative request cost.
- Define a pass condition before comparing outputs.
What changed on September 22?
Anthropic announced Claude Opus 5.5 on September 22, 2026. Its announcement describes lower operating costs and improved performance compared with Opus 5. These are provider-reported results, not measurements made by this publication. This article was checked on September 28; it does not describe a release happening today.
The announcement lists standard API input at $4 and output at $20 per million tokens, with cache reads priced separately. Anthropic's reported 40% reduction on typical workloads is a broader task-cost claim, not a promise that every bill or subscription becomes 40% cheaper. No paid request or hands-on comparison was run for this guide.
Before comparing prices, define the job
Suppose you want an assistant to turn club meeting notes into a checklist. A polished paragraph is not enough: each task needs the right owner, a supported deadline and no invented commitments. You can judge that job without advanced mathematics or an expensive experiment.
Here is our fictional source: 'Mina will test the light sensor on Tuesday. Omar will bring jumper wires; no date is agreed. The robot demonstration is proposed for Friday, subject to teacher approval.' A correct checklist preserves all three qualifications. It must not assign Tuesday to Omar or describe Friday as confirmed.
Write those conditions down before opening an AI tool. Otherwise a persuasive answer can change what you consider acceptable. This is a small evaluation: a repeatable check against a defined goal. It evaluates this narrow task, not the model's overall intelligence. Prerequisites: basic multiplication and the ability to compare an answer with its source; no programming required.
Work through one price example
A token is a unit of encoded content, not a fixed number of words. Input is the material sent for processing; output is what the service generates. An API lets an application request model work. A consumer subscription is a different way of accessing a product and should not be calculated as if it were an API bill.
For an illustrative request with 2,000 uncached input tokens and 500 billable output tokens, using the standard rates above: input costs 2,000 ÷ 1,000,000 × $4 = $0.008. Output costs 500 ÷ 1,000,000 × $20 = $0.010. The combined token charge is $0.018, or 1.8 US cents.
Those counts are assumptions, not a measurement of the meeting-note prompt. The illustration excludes other charges and pricing modes. Actual usage records and current billing documentation determine the bill. Do not estimate exact token counts by counting English words.
A cheap attempt can still make an expensive workflow
Now imagine two fictional systems. System A needs three attempts costing two cents each before its checklist passes. System B passes once at five cents. Their successful outcomes cost six cents and five cents respectively, even though A's individual attempt is cheaper. This is invented arithmetic for teaching, not a Claude benchmark.
Review time matters too. If you spend ten minutes repairing an output, record those minutes separately from money. Avoid converting time into an invented dollar figure. A useful comparison sheet has columns for attempts, actual charges where available, review minutes, unsupported claims and pass/fail.
Keep prompts and source notes identical between tools. Log the model label and date, because a later update can change behavior. Repeat with several fresh examples and retain the failures. Three successes on a familiar paragraph do not establish reliability for private business documents or physical robot control.
Ask for evidence, then inspect it yourself
Weak prompt: 'Make these notes better.' It leaves both the output and the definition of better unclear. The improved prompt below asks for a specific checklist, preserves missing information and requires source evidence.
Follow-up: 'For each deadline, quote the supporting words from the notes. Mark proposals separately from confirmed commitments.' This helps reveal hidden assumptions. It is not proof by itself: a model can quote the wrong passage or misread a condition. Compare each quotation with the original.
Do not upload classmates' personal details or confidential meeting notes just to compare models. The fictional paragraph here is enough. If your available plan does not expose Opus 5.5, complete the reasoning exercise with paper; no upgrade is necessary to learn the method.
Terms to keep and what to read next
Token: an encoded unit used in model processing. API: an interface through which software requests a service. Evaluation: a test against defined criteria. Retry: another attempt after an unsatisfactory result. Cost per checked result: the accumulated expense of obtaining an output that passes your checks.
A cache reuses supported previously processed material; whether a request qualifies is a product-specific billing question. Do not assume repeated text is automatically free. Continue with the existing 'Learn Claude: Better Briefs, Better Reviews' guide and the Understand AI course to practise clear instructions before attempting larger integrations.
Try this prompt
This is a suggested exercise, not a tested guarantee of any model’s output.
Using only the fictional notes below, produce a checklist with task, owner, deadline and status. Write 'not specified' for missing information. Preserve proposals and conditions; do not create calendar events or messages. Notes: Mina will test the light sensor on Tuesday. Omar will bring jumper wires; no date is agreed. The robot demonstration is proposed for Friday, subject to teacher approval.
Mini project & challenge
- Set aside 15–25 minutes. Write the three expected checklist entries before using a model.
- Try the prompt once in a tool you are already allowed to use, or manually draft the output. Do not buy access for this exercise.
- Check the two uncertain items: Omar's deadline and the demonstration's approval. Record each error and any retries.
- Finish when every entry matches the notes and you can explain why the lowest per-request price need not win.
- Challenge: replace one confirmed deadline with a conditional deadline and predict the necessary change before testing again.
Common mistakes
- Confusing the advertised workload saving with a guaranteed subscription discount: identify the exact access route and unit being priced.
- Counting words as exact tokens: use actual usage records for real bills.
- Keeping only the strongest answer: retain failures and retries in the comparison.
- Accepting an invented date because the checklist looks tidy: mark missing information explicitly.
Check your understanding
- When was Opus 5.5 announced in the checked source?
- Are the release's performance claims independent tests by this Learning Lab?
- Why is a subscription fee different from the example API calculation?
- What is the illustrative charge for 2,000 input tokens at $4 per million?
- What is the example's combined token charge including 500 output tokens at $20 per million?
- Which fictional workflow is cheaper: three two-cent attempts or one five-cent attempt?
- What deadline should Omar's checklist entry contain?
- Should the Friday demonstration be labeled confirmed?
- What should stay fixed when comparing the same task across models?
- Why keep review minutes separate from request charges?
Show answers
- September 22, 2026; the source check for this draft is September 28.
- No. They are attributed provider claims; no hands-on comparison was performed here.
- They are different access and billing arrangements; one does not directly determine the other.
- $0.008.
- $0.018, excluding any other charges or pricing modes.
- One five-cent attempt; the other totals six cents.
- Not specified; the notes explicitly say no date is agreed.
- No. It is proposed and depends on teacher approval.
- The source notes, prompt and pass criteria; also record settings and date.
- They measure different costs and reveal manual effort without inventing a monetary value.
In short
A model release offers a reason to test, not a reason to skip checking. Define success, preserve uncertainty, and compare total attempts and review effort before deciding which tool fits.
Sources & review notes
- Anthropic: Introducing Claude Opus 5.5
Announcement dated September 22, 2026; checked September 28. Prices and performance are provider statements. Examples and comparison arithmetic here are illustrative, not observed product results.
Product information was checked on 2026-09-28. This is a selected beginner guide, not an exhaustive archive of every announcement. Recheck access, pricing and compatibility before relying on a current product detail. Supplied screenshots remain unreplicated claims.