Pricing
You pay for tokens, not for time.
Every run is priced by the tokens its models read and write, at each model's list price plus a 30% fee. No subscription and no seats: a month with no runs costs nothing.
- Tokens usedcounted per model, per step
- List pricethe model maker's own rate
- 30%our fee, nothing else
- Your chargeitemised on every report
So every estimate on this site gives the tokens first and the dollars after: an invalidity search at High depth is about 27 million tokens (roughly $27).
What a token is
A piece of a word.
A token is about three quarters of an English word. Models are priced per million of them, in three kinds: text they read (input), text they have already seen and read again from cache (cached input, a tenth of the price or less), and text they write (output).
Reading one reference in full against the claims takes about 60,000 tokens, measured over runs between 7 and 24 September 2026. That is why the size of a search is set in references read, and why its price follows the tokens.
What you see, and when
- Before a run: the estimate beside the controls, in time, tokens and dollars.
- During a run: the tokens used so far and what they cost, live.
- After it: the exact tokens per model and step, and the charge, on the report and your balance page.
- Estimates are the middle of measured runs. The bill is what was actually used.
You choose the size
How long it runs, how much it reads.
The search page sets the depth with one slider, or exactly in Advanced settings: how many references are read in full, whether outside patent databases are asked, and how hard the builder checks its draft. Each setting moves the estimate before you start.
Invalidity search
| Depth | References read | Time | Estimate |
|---|---|---|---|
| Fast | 90 | 9 min | about 5.4 million tokens (roughly $2.91) |
| Medium | 216 | 10 min | about 13 million tokens (roughly $9.01) |
| High | 450 | 12 min | about 27 million tokens (roughly $27) |
| Extra high | 900 | 16 min | about 54 million tokens (roughly $54) |
| Max | 1,800 | 23 min | about 108 million tokens (roughly $108) |
Advanced settings take any depth from 50 to 1,000. Reading more from a finished report later charges only the new references. Building the papers of a third-party build after its search adds about 2.5 million tokens (roughly $6).
Prior-art search
| Setting | References read | Estimate, with an account |
|---|---|---|
| Fast | none | about 40,000 tokens (roughly under 1 cent) |
| Medium | 20 | about 520,000 tokens (roughly $0.20) |
| Medium, with outside sources | 20 | about 790,000 tokens (roughly $0.31) |
Without an account the prior-art search is free: it runs on our own index and nobody is charged. With an account, every run is charged by the token, this one included.
Observations & Oppositions builder
| Strength | Time limit | Steps | Estimate |
|---|---|---|---|
| Low | 1.8 hours | 5 steps, the final QA repeated until ready | about 10 million tokens (roughly $10 to $30) |
| Medium | 3 hours | 7 steps, the final QA repeated until ready | about 20 million tokens (roughly $20 to $60) |
| High | 6 to 7 hours | 11 steps, the final QA repeated until ready | about 60 million tokens (roughly $50 to $200) |
You pay for the tokens the sessions use, not the time allowed: a build that finishes early costs less.
The newest models, used sparingly
Strong where it counts, efficient everywhere else.
Each step runs on the least expensive model that does it well. A fast model reads the thousands of candidates; stronger ones only split claims, check findings, draft and review. Text a model has seen is read again from cache: about 97% of the builder's tokens are cached input, at a tenth of the input price or less.
| Model | What it does | List price, per million tokens | You pay, with 30% | ||||
|---|---|---|---|---|---|---|---|
| Input | Cached | Output | Input | Cached | Output | ||
| Gemini 3.1 Flash-LiteGoogle | Screens candidates and reads references in full | $0.25 | $0.03 | $1.50 | $0.33 | $0.03 | $1.95 |
| Gemini 3.8 FlashGoogle | Splits claims into limitations, checks what was read | $1.50 | $0.15 | $7.50 | $1.95 | $0.20 | $9.75 |
| voyage-4-liteVoyage AI | Embeds passages for meaning-based search | $0.02 | none | none | $0.03 | none | none |
| Claude Opus 5.5Anthropic | Drafts and assembles filings | $4.00 | $0.20 | $20.00 | $5.20 | $0.26 | $26.00 |
| Sol 6.1OpenAI, house rate | Reviews filings and builds the forms | $10.00 | $1.00 | $50.00 | $13.00 | $1.30 | $65.00 |
| Claude Fable 5.1Anthropic | Final QA and fix of every package | $10.00 | $1.00 | $50.00 | $13.00 | $1.30 | $65.00 |
| Astra 6OpenAI, house rate | Final QA and fix of every package | $10.00 | $1.00 | $50.00 | $13.00 | $1.30 | $65.00 |
| Gemini 3.1 ProGoogle | Rules on objections (High builder strength) | $2.00 | $0.20 | $12.00 | $2.60 | $0.26 | $15.60 |
Rates as the ledger bills them today. Sol 6.1 and Astra 6 run on subscription plans with no per-token price, so they are charged at a house rate set to the dearest frontier model with a published rate, so that they can never undercharge.
What we add
30% on top of the token cost.
- No fee per run, no subscription, no seats.
- It pays for the full-text corpus, the servers and the checks between models.
- Your balance page shows each run's model cost and what was charged, side by side.
Credit
$30 to start.
- Registration is free and every new account starts with $30 of credit.
- Top up from $10 by card; credit does not expire. Automatic top-up refills the balance when it runs out.
- A run that has started is never stopped halfway for lack of credit.
- Granted credit has no cash value; credit you paid for is refundable while unspent.
Official fees
The office's fees are separate.
Fees at the patent offices and courts are yours to pay when you file, and each filing page states them. IPtorch files nothing and pays no fee for you.
Try it on your patent.
Without an account the prior-art search is free. An account starts with $30 of credit.
Create a free accountToken figures are medians of runs between 7 and 24 September 2026; dollar figures come from the same estimator the search page uses. The charge for a run is what its models actually used, priced as above. Not legal advice.