$ cat /about.md

UPROUTER.ONLINE is an independent, developer-facing directory of AI API providers: model aggregators, official first-party APIs, free relays, subscription routers and cloud platforms. For every entry we track normalized pricing (USD per 1M tokens), free-tier value, risk notes, data confidence, live status from our own probes, and a full change history. Uprouter also operates UPROUTER CONNECT, an OpenAI- and Claude-compatible gateway that lets you route requests to many providers through one API and one compute balance.

The directory exists because AI API pricing is scattered, unit-inconsistent (per token, per 1K, per 1M, per minute of audio), and routinely buried behind marketing. We do the normalization once, cite our sources, timestamp every verification, and show you when data goes stale — so you can compare providers on facts instead of vibes.

Pricing normalization

All token prices are normalized to USD per 1M tokens (1,000,000 input or output tokens). When a provider quotes per 1K tokens we multiply by 1,000; when they quote per token we multiply by 1,000,000; non-USD list prices are converted at the time of verification and noted in the plan row. Batch, off-peak and commitment discounts are recorded in plan notes rather than in the headline figure. Canonical model prices come from the model operator’s own published rates; each provider entry stores what that provider actually charges, which may include routing fees.

Free-credit value

When a provider grants free credits with a stated dollar value, we list the provider’s own figure. When the allowance is metered in something else (requests/day, tokens/month, audio minutes), we convert it to a dollar value using the provider’s pay-as-you-go rate for the cheapest capable model a typical developer would use. Any figure that required assumptions carries the est. flag on the entry — treat estimated values as an order of magnitude, not a quote. When a free tier exists but cannot be quantified honestly (capacity-dependent pools, request-count-only limits), we say “free tier present — value not quantified” and we never print a misleading zero. When research could not find any public free-tier information for a provider, it gets an honest “Unknown” state — unverified is not the same as absent. Providers verified to offer no free tier get an explicit “No free tier” state.

Deduplication

One entry per provider product, keyed by normalized domain (lowercased, www. stripped). A company that runs two distinct products (e.g. an upstream API and a separate aggregator) gets two entries only if the products have distinct domains and pricing. Community duplicates submitted through /submit are merged into the existing entry rather than published twice.

Uptime probes

Uprouter runs its own lightweight probes against provider endpoints and status pages. Latency is a best-effort round-trip measured from our infrastructure; probes fail for reasons that may not indicate a provider outage (network blips, geo-restrictions, our own cadence gaps). Status history on entry pages shows the raw probe log. Probes are best-effort, not a guarantee — for production decisions, always cross-check the provider’s own status page.

Staleness & data confidence

Every entry carries verification timestamps and a confidence level: verified (re-checked against official sources by our editors recently), community-reported (sourced or confirmed via community evidence, awaiting editorial re-check) and stale (not re-verified within our staleness window — currently 45 days — and shown with an amber outline). Staleness is computed from the verification dates printed on every entry, never hidden.

Model manifest & capabilities

Models are cataloged against an OpenRouter-parity manifest: creator, modalities, context window, max output tokens, tokenizer, knowledge cutoff, license, and capability flags (streaming, tool calling, structured outputs, image/audio input, reasoning, image output). Capability flags are documented-only: a flag is shown exclusively when the vendor documents it; undocumented capabilities render as absent, never assumed. Per-endpoint rows add each provider’s own price, supported context, and disclosed quantization. Throughput figures appear only where measured from real gateway traffic — never estimated.

Coding benchmarks

Benchmark results (SWE-bench Verified, LiveCodeBench, HumanEval, MBPP+, Aider Polyglot, BigCodeBench, …) are stored per model with the score, unit, provenance (independent / vendor / community), evaluation-conditions note, source URL, and dates. Scores are never fabricated or interpolated — an absent benchmark means we have no sourced result, never a zero. Because scaffolds and agent setups differ, scores are comparable only within matching evaluation conditions; wherever conditions differ we label the comparison as not directly comparable.

Some entries include an affiliate link (/go/…). When you use one, we may earn a commission from the provider. The rules: commissions never affect what data we record, how we score risk or confidence, or where an entry sorts — every sort key in the directory is a neutral data field. Affiliate links are disclosed inline, at the link itself. Entries without an affiliate relationship link directly to the provider and say so. We do not accept pay-to-rank, sponsored listings or paid placement of any kind.

Listing submission is free — there is no fee, no payment gate and no payment processor anywhere in the pipeline. Inclusion is an editorial decision on the merits, never sold: editors review each submission for data quality, verifiability and fit. Published submissions meet exactly the same data standards, carry the same source URLs and verification dates, sort by the same neutral keys, and can be rated high-risk like any other entry. Rejection — with a stated reason — is the only gate.

Corrections: if you find an error in any entry, send it to the address on our contact page. Corrections that check out are applied and journaled publicly.