Skip to content
How it works

What happens after you swap the key.

Finest serves the model you asked for. It swaps in a cheaper one only where a recorded test already proved that cheaper model does this exact job just as well. Below: the two-minute setup, a recorded head-to-head on a public SEC filing, an estimate you can run without signing up, and the things we cannot do to your traffic. What Finest is is the shorter answer.

In 2 minutes, start cutting your API spend without sacrificing quality. Free if you don’t save money.

Spend less on AI without guessing. Finest verifies an exact configuration on a frozen safe corpus before you can approve a change.

no accepted public demo record yet: numbers appear once one exists

Most AI apps are over-modeled.

Different jobs need different model capacity. Finest measures each route on evidence you approve, instead of asking your team to infer quality from a model name or a public leaderboard that never saw your workload.

No cost, latency or quality improvement is claimed before a route-specific verification record exists. Catalog prices can discover candidates; only your frozen workload can admit one.

Where the prices on this page come from

These are published provider prices read from the endpoint registry, not measurements of any model’s quality and not a claim about your traffic. They are recomputed on every render, so a price change moves the question rather than dating it.

  • 3.4×: 3.4× separates the upper-quartile endpoint (Claude Sonnet 4.6) from the lower-quartile one (Grok 4.20 (non-reasoning)) on the same 2,000-in / 400-out task, across 23 verified standard-tier endpoints at the price epochs in force now. The middle half of the catalogue rather than its extremes: dividing the most expensive frontier model by the cheapest small one is arithmetically true, describes a swap nobody makes, and reads as a typo. Published list prices, not a measurement of quality: whether the cheaper one does your job is the question the rest of this page is about.
  • : 5× is the median ratio of output-token price to input-token price across the same 23 endpoints. It is why an uncapped response is the most expensive thing a route can do by accident.
  • 50%: Every provider publishing batch terms in the registry discounts by 50% (3 providers). Batch is a migration, not a routing flip, the caller has to submit, poll and correlate, which is why Finest generates it as a reviewable diff rather than switching it on for you.

Documentation snapshots taken 2026-07-26; catalogue age 141h.

What you change

Two lines of config. That is the whole integration.

A base URL and a key, on the OpenAI- or Anthropic-shaped client you already use. Your model names stay exactly as they are.

Swap one key

2 minutes

Change a base URL and a key. Or tell your coding agent to add Finest, and it does the same thing. No annotation, no graders, no config.

  • You always get the model you asked for. A cheaper model serves a call only where a recorded test already covers that exact job: same prompt shape, same schema, same settings. If it refuses or returns malformed output, we re-run the call on the model you named and charge you only what that model would have cost you anyway. The failed cheap attempt is ours, and it never carries a share.
  • Every call comes back with a receipt. The model you asked for, the model that served it, both prices, what you saved and what we charged. Including the calls we served exactly as asked, where we earned nothing.
  • You pay the provider’s own price. Prepaid credit covers it, and we add 25% of what a swap actually saved. Your worst case is the model you asked for, at the provider’s price.
  • Leave whenever you want. One guided flow provisions your own provider keys, exports your evidence and receipts, and ends with your app not depending on us.
  • Your traffic goes through us. That is what this is, and you choose it. Set FINEST_DISABLE=1 and every call goes straight to your provider again, with no redeploy of our code.
What the 2 minutes look like →
The 2-minute motion

Three steps, and the third one is your traffic.

No annotation, no graders, no config. The models you arrive with are recorded as your baseline, so nothing we recommend later can raise the number we bill against.

  1. 01

    Create a key

    You do this yourself in the console, and the key is shown once. A coding agent can do every other step but never sees key material. The models you arrive with are recorded as your baseline, so nothing we recommend later can raise the number we bill against.

  2. 02

    Point your client at us

    A base URL and a key on the OpenAI- or Anthropic-shaped client you already use. Your model names do not change, and we never opt you into finest/auto.

  3. 03

    Send traffic

    Every call comes back with a receipt: the model you asked for, the model that served it, both prices, what you saved and what we charged. Including the calls we served exactly as asked, where we earned nothing.

The race

Same prompt. Same document. Two models.

A real filing from SEC EDGAR, cited below. Watch the clock. The interesting part is the silence in the left pane.

public SEC filing · no signup
stopwatch 0.00s
Frontier modelClaude Fable 5
ready
not started
Cheaper modelGemini 3.5 Flash-Lite
ready
not started
The document both models read, and where it came from

Apple Inc., Form 10-K for the fiscal year ended September 30, 2023: cover page. SEC EDGAR accession 0000320193-23-000106, CIK 0000320193.

Public filing, disseminated by the U.S. Securities and Exchange Commission. Retrieved 2026-07-26. sha256:bda1f34435199672c16ecdf2034c650872d2cac8399ed0d179fc25450b080b90

Open the filing on SEC EDGAR
the task

Extract the registrant facts from the filing cover page below and return JSON only, using exactly these keys: registrant, form_type, fiscal_year_end, commission_file_number, state_of_incorporation, irs_employer_id, trading_symbol, exchange, public_float_usd, public_float_as_of, shares_outstanding, shares_outstanding_as_of. Dates are ISO-8601. Money and share counts are integers with no separators. If a field is absent from the document, use null.

UNITED STATES SECURITIES AND EXCHANGE COMMISSION
Washington, D.C. 20549

FORM 10-K

☒ ANNUAL REPORT PURSUANT TO SECTION 13 OR 15(d) OF THE SECURITIES EXCHANGE ACT OF 1934
For the fiscal year ended September 30, 2023

Commission File Number: 001-36743

Apple Inc.
(Exact name of Registrant as specified in its charter)

California                       94-2404110
(State or other jurisdiction     (I.R.S. Employer
of incorporation)                Identification No.)

One Apple Park Way
Cupertino, California 95014
(Address of principal executive offices) (Zip Code)

(408) 996-1010
(Registrant's telephone number, including area code)

Securities registered pursuant to Section 12(b) of the Act:
Title of each class                          Trading symbol(s)   Name of each exchange on which registered
Common Stock, $0.00001 par value per share   AAPL                The Nasdaq Stock Market LLC
1.375% Notes due 2024. The Nasdaq Stock Market LLC
0.000% Notes due 2025. The Nasdaq Stock Market LLC

Securities registered pursuant to Section 12(g) of the Act: None

The aggregate market value of the voting and non-voting stock held by non-affiliates of the Registrant, as of March 31, 2023, the last business day of the Registrant's most recently completed second fiscal quarter, was approximately $2,591,165,000,000.

15,552,752,000 shares of common stock were issued and outstanding as of October 20, 2023.

Racing your own prompt is not open yet. This demo runs on a public filing only.

Your own numbers

What you are probably overpaying, itemised.

No signup, no key, nothing uploaded. We multiply the dollars you type by published provider prices. It is an estimate of where to look, not a saving we have proven.

Your numbers
$

We start you in the middle of our catalogue, not at the priciest model. Change it to what you actually run and every line below recomputes.

What the model is asked to do
25%
25%
20%
20%
10%
The overspend estimateevery line an estimate from published prices
JobYou pay nowCheaper model we would testPossible saving
Ticket classification
Route a support ticket to a queue, or tag it with an intent.
25% of traffic · 300,000 requests · assumes 1,200 in / 20 out
$2,000
DeepSeek V4 Flash
91.2% cheaper per request
$1,824
per month
Document extraction
Pull fields out of an invoice, form or filing into a schema.
25% of traffic · 300,000 requests · assumes 4,000 in / 400 out
$2,000
DeepSeek V4 Flash
93.0% cheaper per request
$1,860
per month
Summarisation
Condense a thread, a call transcript or a long document.
20% of traffic · 240,000 requests · assumes 6,000 in / 600 out
$1,600
DeepSeek V4 Flash
93.0% cheaper per request
$1,488
per month
Customer-facing generation
Draft a reply, a description or a piece of copy a person reads.
20% of traffic · 240,000 requests · assumes 1,500 in / 900 out
$1,600
DeepSeek V4 Flash
95.5% cheaper per request
$1,529
per month
Agent steps
One step of a tool-using loop, run five to thirty times per task.
10% of traffic · 120,000 requests · assumes 8,000 in / 700 out
$800.00
DeepSeek V4 Flash
92.8% cheaper per request
$742.47
per month
Possible monthly saving$7,443Estimated
  • Estimated price opportunity = your stated spend for the line × (1 − lowest compatible catalog price ÷ current catalog price) at the token profile shown. Only the ratio comes from the registry; route-specific quality is not established here.
  • Prices are the epochs in force in the registry, read from official provider documentation.
  • Every figure here is an estimate. Provider invoices can reconcile actual spend; the counterfactual and savings remain estimates.
  • Registry documentation snapshots taken 2026-07-26 (141.5h ago).
  • Published price records used: pe_FY7T82ZRGJYXYXJ7BPAWSNRT4R, pe_VQT6FE8Y5H3VTFGD34BKNFVNZM.

Cheaper models are usually faster too. We have measured nothing about your app, so this estimate stays in dollars.

Limits

What we cannot do to your traffic.

Each of these is a property of how Finest is built, not a policy we could quietly change.

If anything in Finest fails before your call reaches a provider, your own model serves it.

A missing policy, an expired signature or a routing error all fall back to the configuration you started with. Once a provider may already have received the request, we return a clear error instead of risking a second billable call.

Kill switch: FINEST_DISABLE=1 and it’s as if we were never there.

One environment variable. No redeploy of our code, no support ticket, no negotiation. FINEST_DISABLE=1 hands your client the direct provider configuration, and we are out of the path.

Policies expire back to your original configuration on their own.

Every policy we sign carries an expiry. Stop paying us and they lapse back to where you started. Nobody has to switch anything off by hand, and we cannot hold a route hostage to get you to renew.

We cannot make leaving hard, because we build and advertise the exit.

Eject provisions your own provider accounts, exports your evidence and receipts, converts the policies we serve into policies you hold, and verifies your app no longer depends on our uptime.

We cannot charge you for a saving we did not measure.

A fee needs two things: a recorded before and after that you approved, and your provider’s own invoice to reconcile it against. We bill against the lower of what you were spending and the option we recommended, so saying yes to a recommendation can never raise the number we charge against.

The parts that touch your traffic, the client library and the usage collector, are open source. The security page tracks exactly where they are.

25% of the savings we can prove, and nothing if we prove none. The worked example and the FAQ are on the pricing page.

Pricing →