411 models behind one OpenAI-compatible endpoint, with spend caps, tenant isolation and an auditable ledger in the request path — at a median 68% below OpenRouter pricing.
api.openai.com/v1→llm.b33r.us/v1
Models from
Pricing
388 models priced line by line. Set the output share to your traffic mix and every figure recalculates.
| Model | OpenRouter in | OpenRouter out | b33r flat | Saving |
|---|
USD per million tokens. OpenRouter figures from their public models API, 15 August 2026.
Adoption
We implement the OpenAI wire format, so your SDKs keep working — and leaving is as cheap as arriving.
One key per team or environment, each with its own spend cap and expiry.
Set base_url to https://llm.b33r.us/v1. No SDK swap, no rewrite.
One string selects the model. Streaming, tool calls and vision pass straight through.
from openai import OpenAI client = OpenAI( base_url="https://llm.b33r.us/v1", api_key="sk_live_...", ) stream = client.chat.completions.create( model="anthropic/claude-opus-4.5", messages=[{"role": "user", "content": "Hello"}], stream=True, ) for chunk in stream: print(chunk.choices[0].delta.content or "", end="")
Platform
Budget enforcement, metering and isolation are properties of the gateway — not dashboards layered on after the fact.
Spend caps and expiry are enforced at the gateway, not at invoice time. A runaway agent loop hits a ceiling.
Each billable call writes one row — model, tokens, latency, cost. Catalog reads and status polls are never metered.
Jobs are bound to the key that created them. Another tenant's job id returns 404, never a redacted record.
In-flight work is written to a durable job store. A 250-second render survives a deploy, a restart or a scale-out.
Text, image, video and audio share one format and one key — a single surface for security review, not one per vendor SDK.
Generated media is re-hosted and served from llm.b33r.us — never a vendor CDN link that expires under your product.
Security & operations
Every item below is live on llm.b33r.us/v1 today, verifiable against the running API — not a roadmap entry.
Keys are stored as SHA-256 hashes, never plain text. A database disclosure yields no usable credential.
We do not train on your prompts or completions. What we keep is metering metadata: model, tokens, latency, cost.
Every job is bound to the key that created it. An id from another tenant returns 404, not a redacted record.
The endpoint is HTTPS-only; plain HTTP is redirected, not served. Every upstream call is made over TLS.
Upstream failures surface with the provider's reason attached. A rate limit stays a 429, never a blind 500.
In-flight work is written to a durable store. A restart or deploy mid-render does not drop the job or its result.
What we do not claim. b33r holds no third-party audit certification — no SOC 2, no ISO 27001 — and will not display a badge for one until it exists. If a formal attestation gates your procurement, raise it before you build on this.
Due diligence
The questions a review board raises — pricing, model integrity, data handling, compliance posture.
We charge one flat rate per model, applied equally to input and output. Most gateways price output four to five times higher than input, and production traffic is output-heavy — that multiplier is where the money goes.
The rest is structural: bulk capacity and a thin gateway rather than a marketplace with per-provider margin. It also makes spend forecastable, since your cost moves with total tokens rather than with the input-to-output ratio of a given month.
The real ones, at the provider's own precision. anthropic/claude-opus-4.5 is that model, not a smaller variant wearing its name. We do not silently downgrade to protect margin — if a served model changed, the model id would change with it.
Two values: the base URL and the key. We implement the OpenAI wire format, so /v1/chat/completions, streaming, tool calls and /v1/models behave exactly as your SDK expects.
We do not train on your traffic. What we keep is metering metadata — model, tokens, latency, cost — because you need it to audit the bill. Traffic is HTTPS-only inbound and TLS outbound.
One boundary worth stating plainly: your prompts are still processed by whichever provider serves the model you select, under that provider's own terms. We will not imply otherwise.
No, and we will not display a badge suggesting otherwise. There is no third-party audit report for b33r today.
What we can put in front of a reviewer is the behaviour itself: SHA-256 key storage, per-key spend caps and expiry, tenant-bound job lookups, HTTPS-only transport, and a per-call usage ledger — each verifiable against the running API. If a formal attestation gates your procurement, raise it early.
Issue a separate key per team, environment or customer, each with its own spend cap and expiry. The cap is enforced at the gateway, so a runaway agent loop hits a ceiling instead of an invoice.
Every billable call writes one ledger row, so chargeback and per-team attribution come off a record you can query. Catalog reads and status polls are not metered.
Because the gap moves with the model and with your input-to-output ratio — a single headline number would be tidier and less true. We publish the median across 388 comparable models and give you the table to check the ones you actually run.
You get the real status code with the upstream reason attached — a rate limit stays a 429, never a generic 500. And because every model shares one request shape, failing over is a one-string change in your retry path.
Send one prompt to four models side by side and read the real cost of each response — before you commit a line of code.