Changelog

Google's newest Gemini is live as gemini-4-argon at $0.60 input and $3.00 output per million tokens, against Google's list price of $2.00 and $10.00. MiMo V2.6 Pro is back in the catalog.

The customer API is described in an OpenAPI 3.1 document at api.uplinkswitch.com/openapi.json: every route, its request and response shapes, the authentication headers and the x-uplink-* headers. Model ids are enumerated by GET /v1/catalog, not in the document.

This page, grouped by day, with an RSS feed at /changelog/feed.xml.

Your organization's image, video, speech and 3D generations in one list, with a panel to try a model as one of your keys.

A key can be narrowed to a list of models or given an expiry after it is created. The account page takes email change and account deletion requests.

The usage page in the console filters by key, model and outcome, and exports the rows as CSV.

The form at /contact reaches us directly. An expired console session shows the sign-in instead of an error.

One failed probe stops marking a service degraded after six hours.

Every listed chat model with a developer list price, our price next to it, at /pricing-index, also as JSON and CSV. Two-model comparisons at /compare, price hubs per family at /pricing, and media models by task at /models/media/tasks.

What Claude Code costs on API billing, the October LLM API price comparison, and what the daily real-model check found about hidden prompt tokens.

OpenCode, Cline, Roo Code, Kilo Code and OpenClaw join Claude Code and Cursor. Guides for Nano Banana Pro and Veo 3.1, and a Vercel AI SDK provider: npm install @uplinkswitch/ai-sdk-provider.

Every model and price on one line each for AI agents, at /llms-full.txt. robots.txt points to both files.

Side by side with sources, at /alternatives/openrouter.

When a route's own hidden prompt would outweigh a short request, the router prefers a cleaner route for that model. The price does not change. Six routes that no longer answered are off sale.

The developers' published prices are on record for both, so the catalog, receipts and the savings headers show the saving.

Both image models show Google's own price next to ours.

No media setting costs more than a larger one of the same model, and audio off never costs more than audio on. Veo 3.1 shows Google's price next to ours.

Both are priced and callable again.

The media catalog shows the developer's list price wherever the developer publishes one.

POST /v1/images/generations takes OpenAI's image request for the image models. POST /v1/generations takes any media model and answers with a generation to poll at GET /v1/generations/{id}. Files are served at /v1/files for 30 days, and GET /v1/catalog/media lists the models, their settings and prices.

Register up to five golden prompts, each against up to five models. They run once a day through your key, and a change in the answer or in the billed input count shows in the console and arrives by email.

Share your link and earn 10% of a referred account's top-ups for 12 months, as balance. Gift cards of $10 to $500 redeem as prepaid credit.

Once all customers together pass a volume tier on a model in a month, its price drops for everyone. The catalog shows the tiers and the current discount.

A large system prompt that many accounts send gets a cache breakpoint on Anthropic-format requests. Cached reads bill at the cached rate, and the response says x-uplink-crowd-cache: on. Send x-uplink-crowd-cache: off to leave a request out.

A key can name a cheaper model to answer with as it nears its monthly limit. The response names the model that was asked for in x-uplink-autopilot.

x-uplink-waste on a response says what share of the request repeated earlier text or was a pasted blob, when it is 10% or more.

Every non-streamed answer carries x-uplink-receipt: a signed statement of which model answered, what was billed, and the hashes of the request and the response. Any request's receipt is at GET /v1/requests/{id}/receipt, and the public key at GET /v1/statements/public-key.

An answer the model refuses is not billed. A request that asks for JSON or a schema, or sends x-uplink-require: json, pays only for an answer that matches. A miss is retried once and the missed answer is free.

A host's input count can never exceed what the request could hold. The tokens on a receipt are the tokens billed.

x-uplink-charge, x-uplink-list-cost and x-uplink-saved on non-streamed responses. A stream has the same figures at GET /v1/requests/{id} once it finishes.

Each top-up locks the rates you pay for 30 days. Inside the window a price can only go down.

A key set to fastest tries the route that answered quickest lately. An alias pins a name of your choosing to a model. POST /v1/estimate prices a request on every model before you send it. Replay sends a kept request to another model.

Each month's statement is signed with Ed25519 and can be checked at /statements/verify without an account.

Every day one real request runs on each format's fallback route, and the status page shows the result.

A model costs the same in every wire format and on every route behind it. Metered rates are held inside a band under the developer's list price.

Try models side by side on the home page. A fixed prompt goes to every route of every model each day, and /proof shows whether each model's fingerprint held.

A chat playground with per-request timing marks, keys shown as env rows, and a status page.

GET /v1/status probes the first route per format and its fallback every hour.

GET /v1/catalog carries the developer's published price next to ours.

New accounts have a request-rate ceiling for 24 hours and a spend cap until the first payment settles. Each key can carry a monthly spend limit.

A page per model, and guides for Claude Code, Cursor and the SDKs.

The site and the console open dark. The toggle still switches.