In this guide · 8 sections
- 01Task matrix: which model tier for which job
- 02Prices per million tokens (October 2026)
- 03Worked costs for 1,000 runs
- 04Reasoning settings: the hidden cost multiplier
- 05How to test a cheaper model on your own examples
- 06Routing rules you can apply today
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Use a small model (GPT-6 Luna, Claude Haiku 5.5, Gemini Flash-Lite) for classifying, extracting and short drafts. Use a mid-range model (GPT-6.1 Sol, Claude Sonnet 5.5) for analysis and longer writing. Keep flagship models for hard reasoning and code. Test the cheaper model on your own examples first.
Task matrix: which model tier for which job
Vendors sell three rough tiers: small (cheap, fast), mid-range (balanced) and flagship (most capable, most expensive). The table below is a starting point, not a rule. Your own test decides.
| Task | Start with | Why |
|---|---|---|
| Classify or tag (enquiry type, sentiment, spam) | Small | Short output, clear right answer, easy to check |
| Extract fields (invoice number, date, amount) | Small | Structured output; add a validation check |
| Short drafts (FAQ replies, order updates) | Small | Template-like text; a human or rule reviews it |
| Summarise long documents or meetings | Mid-range | Needs to keep facts straight over many pages |
| Analyse (compare quotes, spot issues in a contract) | Mid-range | Multi-step reasoning; errors cost money |
| Write code or run multi-step agents | Mid-range or flagship | Hard tasks where failures waste more than tokens |
The task you think needs a big model often splits into a cheap part and an expensive part. For example, sorting incoming emails is cheap; drafting a careful reply to the one complaint a day is worth a better model.
Prices per million tokens (October 2026)
List prices in US dollars per million tokens, standard tier, from OpenAI's API pricing, Anthropic's pricing and Google's Gemini API pricing. Approximate, as of October 2026, before tax.
| Tier | Model | Input | Output |
|---|---|---|---|
| Small | GPT-6 Luna | $0.10 | $0.50 |
| Small | Claude Haiku 5.5 | $0.10 | $0.50 |
| Small | Gemini 3.1 Flash-Lite | $0.25 | $1.50 |
| Mid-range | GPT-6.1 Sol | $2.00 | $10.00 |
| Mid-range | Claude Sonnet 5.5 | $2.00 | $10.00 |
| Mid-range | Gemini 3.8 Flash | $0.75 (then $1.50) | $3.75 (then $7.50) |
| Flagship | Claude Opus 5.5 | $4.00 | $20.00 |
| Flagship | Gemini 4 Argon | $2.00 (then $4.00) | $10.00 (then $20.00) |
| Flagship | GPT-6 Astra | $10.00 | $50.00 |
Our AI API pricing comparison covers caching, batch discounts and the premium fast tiers in more detail.
Worked costs for 1,000 runs
Two common jobs, with token sizes stated so you can redo the sums for your own work. Cost = input tokens x input price + output tokens x output price.
- Classify 1,000 emails on GPT-6 Luna600 tokens in, 20 out per emailabout US$0.07
- Same job on GPT-6.1 Solabout US$1.40
- Same job on GPT-6 Astraabout US$7.00
- Draft 1,000 customer replies on Claude Haiku 5.52,000 tokens in, 300 out per replyabout US$0.35
- Same job on Claude Sonnet 5.5about US$7.00
- Same job on Claude Opus 5.5about US$14.00
List prices as of October 2026, no caching or batch discount, no extra reasoning tokens.
For a small business sending a few thousand requests a month, even the expensive option is affordable. The choice starts to matter when a workflow runs on every email, every order or every page of a long document, or when an automation tool bills you in its own units on top.
How to test a cheaper model on your own examples
Pick 20 to 30 real examples
Emails, invoices or questions from the last month, with personal data removed. Include awkward cases.
Write down the right answer for each
For classification, the correct label. For drafts, the points a good reply must cover.
Run your current model and the cheaper one
Same prompt, same settings. Record answers and the token counts each API reports.
Score and compare
Count correct answers. If the cheaper model matches on 28 of 30 and the two misses are caught by a simple check, it is good enough.
Re-test when models change
Vendors retire and reroute models often; Google routed Gemini 3.7 Flash requests to 3.8 Flash on 8 October 2026. Re-run your test set after any change.
Routing rules you can apply today
"Routing" means sending each request to the cheapest model likely to handle it. You do not need special software; an if-statement in your app or a filter step in Zapier or Make is enough.
Routing checklist
See our AI API cost control guide for those settings, and ChatGPT vs Claude vs Gemini for business if you are choosing staff chat subscriptions rather than API models.
Do it yourself or get help?
If a developer built your workflow, ask them which model and reasoning setting each step uses, and run the 30-example test above. Switching is often a one-line change.
If you would rather have it set up, our Workflow Automation (S$390) builds one workflow of up to five steps on Make, Zapier or Google Apps Script in your own account, with error alerts. Our AI Website Assistant (S$790) answers customer questions from your website and FAQ. IT Sorted is a Singapore-based team of developers with over 10 years of experience; you pay upfront, send the brief in writing and follow progress on your order page.
Frequently asked questions
For simple, checkable tasks such as tagging emails, extracting invoice fields or drafting standard replies, small models are often good enough. For analysis, long documents or code, a mid-range model usually makes fewer costly mistakes. Run 20 to 30 of your own examples through both and compare before deciding.
On October 2026 list prices, GPT-6 Luna and Claude Haiku 5.5 cost US$0.10 per million input tokens and US$0.50 per million output tokens. Mid-range GPT-6.1 Sol and Claude Sonnet 5.5 cost US$2 and US$10, so about twenty times more per token. Flagship models cost more again.
Routing means sending each request to the cheapest model that can handle it, and escalating to a bigger model only when needed. A simple version is a rule: use a small model first, check the result, and resend to a mid-range model if a required field is missing or the answer fails a validation check.
Yes. Reasoning tokens are billed as output, so the same model can cost several times more at a high setting. Artificial Analysis's October 2026 leaderboard shows Claude Haiku 5.5 at about US$0.05 per test task on medium and US$0.21 on max. Start with the lowest setting that passes your test.
Subscriptions such as ChatGPT Business or Claude Team suit staff typing questions themselves. API models suit automated workflows: chatbots, email sorting and document processing that run without a person. Many businesses use both. The model-per-task advice applies to the API side, where you pay per token.