Elasticity Elasticity

Open-weight inference,
90% off.

90%
off retail
on every open-weight model below
45
open-weight models
subsidized right now, priced live below
3
redaction passes
pattern match, entity recognition, model audit
01 / Drop-in

Change the base URL. Keep your code.

Point any OpenAI-compatible client at the gateway, swap in your pilot key, and pick an open-weight model. The 90% discount is applied on every call.

OpenAI-compatible API LiteLLM
from openai import OpenAI

client = OpenAI(
    base_url="https://api.elasticity.dev/v1",
    api_key="YOUR_ELASTICITY_KEY",
)

resp = client.chat.completions.create(
    model="qwen3-235b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)
qwen3-235b input: retail $0.090/M → $0.009/M at 90% off
02 / What we subsidize

45 open-weight models, priced live.

Retail and 90%-off prices come straight from the gateway, in $ per 1M tokens (input / output) — the number that lands on an invoice. Shaded “Closed” rows are reference prices from OpenRouter; Elasticity never bills for them.

Open-weight models with retail and 90% off input and output prices, plus benchmark scores
Model Input Output Discount AA Index AA Coding Context

Swipe sideways for every column.

Open-weight models carry their makers' licenses; some require attribution or a separate agreement at scale. See Model licenses.

03 / How redaction works

Sensitive detail is removed before anything is kept.

Each trace passes a three-stage pipeline that replaces names, emails, credentials and other identifying detail while keeping the conversation readable.

Exhibit A · demo view, de-identified From Sarah ChenJordan Reyes at Northbridge CapitalBellwether Capital — the board wants confirmation before Friday on Project Nightingale[Project Codename 1]. Wire instructions are with j.reyes@fenwickcounsel.comm.alvarez@example.com; routing number 071000013[Aba Routing Number 1] was pasted in plain.

This is real output from the pipeline’s demo view, not a mockup — paste your own text in the live demo to see it run. Every mention of the same person gets the same substitute throughout the request, so a model can still follow who did what.

01

Pattern match

Deterministic regex takes everything with a shape — emails, API keys, tokens, phone and account numbers — and applies the rule below.

02

Entity recognition

A local NER model finds what has no fixed shape: people, organizations, places, project codenames, regulated ID formats.

03

Model audit

A model re-reads the already-redacted text hunting for whatever the first two missed. It flags what it finds; it does not quietly rewrite anything.

What “de-identified” actually means

One rule, however a detail gets caught: the real value is removed and a numbered placeholder such as <PERSON_1> takes its place — never the real value, and never a fake that could pass for a real one. The same person, company or account gets the same number throughout a request, so a model can still follow who did what. The demo below shows realistic stand-in names so it reads naturally; the stored trace uses the numbered placeholders.

Numbered, so who’s who stays clear

Names · Organizations · Emails · Phone numbers · Places · Street addresses

Redacted outright

Credentials & secretsAPI keys · Access & OAuth tokens · Passwords · Private keys · Cloud & vendor keys · Webhook URLs · Database connection passwords

Payment & bankingCard numbers, expiry & CVC · Bank accounts · IBAN · Routing & SWIFT/BIC codes · Crypto wallets

Government & national IDsSSN · Tax IDs · Passport · Driver’s license · National ID numbers from the US, UK, India, Australia, Singapore, South Korea, Italy, Spain, Poland, Finland, Nigeria and Thailand

Health & insurance IDsMedical license · NHS · Medicare · Provider & beneficiary numbers

Personal & account identifiersDates of birth · Usernames & @handles · Employee & account IDs · Home-folder usernames · Vehicle plates · Postcodes

Devices & infrastructureIP & MAC addresses · Device IMEI · Internal URLs

Raw content is never sold. The de-identified trace is what makes the discount possible — nothing else leaves this pipeline.

What Elasticity controls

Every trace is de-identified within minutes of your call completing, and the raw version is then deleted.

Your model provider still receives the prompt as you sent it, as with any direct API call — a gateway can’t change that.

04 / Try the demo

Run it yourself.

The real pipeline, all three passes, on whatever you paste or upload — the same audit model the production pipeline runs, not one you pick. No login required.

Exhibit C · your document

A de-identified copy of what you run here may be kept to improve the product.

How to read the result

Three passes, run once. Nothing here is scored, flagged, or stored. Click any highlighted span to see why the pipeline treated it that way.

Synthetic — a fake name/email/org, consistent for the same entity
Anonymized — same entity always gets the same label
Preserved — kept verbatim, by design

Start with a pilot key.

Pilot keys are self-serve — get one in minutes, no sales call. Every call through it is protected by the same 3-pass de-identification pipeline described above, so you get 90% off without worrying about what’s in your prompts.

Get an API key
Self-serve. No sales call.