Three tiers, in plain language
Each major provider sells a family of models at different sizes: a large one that is slower and pricier, and smaller ones that are faster and cheaper. Prices are per million tokens, the unit models read and bill by; a page of English is roughly 500 to 700 tokens. You pay one rate for what you send and a higher one for what the model writes.
The thinkhard.si catalog groups models into three tiers. The “good for” column is a starting point for which models to test, not a verdict.
| Tier | Examples and prices | Good first candidate for |
|---|---|---|
| Top | Claude Fable 5.1 and GPT-6 Astra, $10 / $50. Claude Opus 5.5, $4 / $20. Gemini 4 Argon, $2 / $10 until November 30, 2026, then $4 / $20. | Long documents that need judgment, multi-step reasoning with numbers, the final version of something important. |
| Mid | Claude Sonnet 5.5 and GPT-6.1 Sol, $2 / $10. Gemini 3.8 Flash, $0.75 / $3.75 (rising to $1.50 / $7.50 on January 1, 2027). | Everyday drafting and analysis, where tone matters and a person edits the result. |
| Fast | Claude Haiku 5.5 and GPT-6 Luna, $0.10 / $0.50. Gemini 3.5 Flash-Lite, $0.30 / $2.50. gpt-oss-20b on Groq, $0.075 / $0.30. | Sorting, tagging, extracting fields, short replies, anything at high volume with a checkable answer. |
What 1,000 tasks cost
Here are five common office tasks priced per 1,000 runs. The token counts are illustrative assumptions for a typical case, not measurements; your own prompts will differ, and the AI Cost Calculator will price them exactly as they are, with no key.
| Task | Claude Haiku 5.5 or GPT-6 Luna | Gemini 3.8 Flash | Claude Sonnet 5.5 or GPT-6.1 Sol | Claude Opus 5.5 | GPT-6 Astra or Claude Fable 5.1 |
|---|---|---|---|---|---|
| Tag a support ticket (600 / 30) | $0.075 | $0.56 | $1.50 | $3.00 | $7.50 |
| Pull fields from an invoice (1,500 / 200) | $0.25 | $1.88 | $5.00 | $10.00 | $25.00 |
| Draft a customer reply (2,000 / 250) | $0.33 | $2.44 | $6.50 | $13.00 | $32.50 |
| First draft of a web page (1,500 / 1,500) | $0.90 | $6.75 | $18.00 | $36.00 | $90.00 |
| Summarize a 20-page report (12,000 / 600) | $1.50 | $11.25 | $30.00 | $60.00 | $150.00 |
- Input
- Output
| Item | Input | Output |
|---|---|---|
| Tag a support ticket | $1.20 | $0.30 |
| Pull fields from an invoice | $3.00 | $2.00 |
| Draft a customer reply | $4.00 | $2.50 |
| First draft of a web page | $3.00 | $15.00 |
| Summarize a 20-page report | $24.00 | $6.00 |
Cost per 1,000 tasks, from the token counts in the table above. GPT-6.1 Sol has the same prices, so the same split. Source: thinkhard.si model catalog, list prices checked October 20, 2026.
Read the table down a column as well as across. A small business drafting 300 replies a month pays about $2 on Claude Sonnet 5.5, so the choice there is about quality, not money. A company tagging 50,000 tickets a month pays $3.75 on the fast tier and $375 on the top one, every month, and that difference only makes sense if the fast model fails.
Monthly cost of 50,000 calls with about 600 tokens in and 30 out each. List prices checked October 20, 2026; your bill may differ.
Three price details catch people out:
- Scheduled changes. Gemini 4 Argon’s introductory price ends on November 30, 2026, and Gemini 3.8 Flash doubles on January 1, 2027. Budget for the price you will pay in three months.
- Long-prompt surcharges. Some models charge more above a prompt length. Claude Haiku 5.5, for example, rises to $0.50 / $2.50 above 100,000 tokens.
- Discounts for waiting. Many providers offer batch pricing at around half price for jobs that can wait up to a day, such as overnight tagging or monthly summaries.
The full arithmetic, including caching, is in what an LLM call really costs.
Where expensive models earn it
These are starting points for your test, not rules. Plenty of fast models surprise on hard tasks, and plenty of top models fumble a simple one.
- Start with a fast model when the answer is short and checkable: a category, a yes or no, a date pulled from a document, a list reformatted. If it is wrong, a simple check catches it.
- Start with a mid model for everyday writing that a person will edit: replies, summaries, first drafts.
- Start with a top model when the input is long or tangled and the answer depends on many details at once (a 60-page contract, a thread with twelve people in it), when it has to reason through several steps with numbers, or when one mistake would cost more than a year of model bills.
Signs a model is too small for the job: it follows four of your five instructions and quietly drops one; it loses details from the middle of long documents; its format drifts from one run to the next; it passes your easy test cases and fails the hard ones. Signs you are overpaying: a cheaper model passes the same tests, and in a blind comparison your team cannot reliably tell which answer came from which.
Speed, and when it matters
Speed is two numbers: how soon the first words appear, and how long the whole answer takes. Both count as latency.
- It matters for live chat, anything a customer is waiting on, and tools staff use dozens of times a day. A ten-second wait, repeated, is a reason people stop using a tool.
- It hardly matters for overnight reports, batch tagging and weekly summaries. Use the time to save money with batch pricing instead.
- Bigger usually means slower, and so does a higher reasoning setting. Some providers, such as Groq, specialize in running open models very fast.
- Measure it on your own prompt. Speed changes with prompt length, answer length, time of day and load. A single published figure tells you little.
The cheapest model that passes
There is no general answer to which model is best. There is a clear answer to which model is the cheapest one that passes your tests, and you can find it in an afternoon:
- Write down what a good result must do, as five to ten rules a reader could check. “Uses one of our eight categories.” “Mentions the order number.” “Under 150 words.”
- Collect 20 to 30 real examples, anonymized, including the hard and odd ones.
- Pick three candidates: one fast, one mid, one top, ideally from at least two providers.
- Run the same prompt on all three. Score them with checks and an AI judge, and read the failures yourself.
- Set the bar before you look at prices. For example: every must-pass rule holds on every example, and the judge is satisfied on at least 9 in 10.
- Price the models that clear the bar at your real volume.
- Choose the cheapest one that clears it, and keep the runner-up as a fallback in case of an outage or a price change.
- Re-run when a new model comes out, a price changes or you change the prompt.
A hypothetical to show the shape of the decision. Say a 30-person property management firm wants to tag 4,000 tenant emails a month into eight categories, and tests Claude Haiku 5.5, Gemini 3.8 Flash and Claude Sonnet 5.5 on 40 emails its staff already tagged. Suppose all three agree with the staff on at least 37 of 40, and the misses are the same few ambiguous emails. At that volume, tagging costs about $0.30 a month on Haiku and $6 on Sonnet. Both are trivial, so the firm can choose on speed or on how the misses look. At 400,000 emails a month the same choice is $30 against $600, and the cheap model wins clearly.
- 60-page contract: Volume 1/10, How hard the task is 9/10
- 300 replies a month: Volume 2/10, How hard the task is 4/10
- 4,000 emails tagged: Volume 4/10, How hard the task is 1/10
- 50,000 tickets tagged: Volume 9/10, How hard the task is 2/10
Illustrative: examples from this guide, placed by judgment, not measured. At low volume the bill is small on any model; at high volume the gap between tiers is about a hundredfold.
Mixing models
Sometimes the answer is two models. A fast model handles most requests and flags the ones it is unsure about, and a stronger model or a person takes those. The triage prompt in the customer support guide sets a “needs_human” flag for exactly this kind of routing.
Only mix when your test results show the failures cluster in a way you can detect. Every extra model is another price to watch, another set of terms to read and another thing to re-test when something changes.
What else decides it
- Data terms. A model is only an option if its provider’s terms fit the data you will send it. The risk checklist for AI at work covers what to look for.
- How you buy it. Directly from the provider with a key, through a reseller such as OpenRouter, or through a cloud platform your company already has a contract with. The guide to API keys explains the first two.
- Document length. Every model has a context window, the most text it can take in at once. Most current models take well over 100,000 tokens, but check before you plan to send whole contracts.
- Retirement. Providers retire models on a schedule. Anthropic, for example, commits to at least 60 days’ notice. Keep your test set so moving takes an hour, not a project.
- Portability. A prompt written as a clear brief moves between providers with small changes. The guide to Claude, GPT and Gemini covers the layout differences that remain.
Key takeaways
- Pick the cheapest model that passes a test built from your own real work.
- Fast models suit short, checkable answers. Top models earn their price on long, tangled inputs and multi-step reasoning.
- The price gap between tiers is about a hundredfold. At low volume, choose on quality; at high volume, test hard before paying for a top model.
- Prices here were checked October 20, 2026. Watch for scheduled price changes and long-prompt surcharges.
- Speed matters when someone is waiting. Otherwise use batch pricing and save.
- Keep the test set, and re-run it whenever a model, a price or the prompt changes.