Get cheap frontier model tokens for China model prices
Claude and GPT, through the official APIs, at up to 50% off list. Not a Chinese model. Not a proxy to one. The same models you already use — we just get better prices, and pass them on without the quality going down.
- 50%
- off list price
- 99.9%
- uptime target
- Zero
- prompt retention
- 60s
- to migrate
client = Anthropic(- base_url="https://api.anthropic.com"+ base_url="https://api.lowcostllm.com" )Same SDK. Same request format. One line changed.
The obvious question
Why we can charge less
If a discount does not come with a mechanism, it is a trick. Here is ours, in full.
We forward, we don't substitute.
Your request arrives at our edge and goes straight to the frontier model you named, through that provider's official API. No swapping in a cheaper model, no silent downgrade under load, no distilled stand-in. The model string you send is the model that answers.
We buy committed capacity up front.
We contract for throughput in bulk and pay for it whether or not it gets used. Providers discount that commitment, because it is worth more to them than pay-as-you-go traffic. The discount is what we resell — we are not subsidising your tokens out of venture funding.
Volume is the whole business.
The more traffic we route, the better our terms get. That makes our incentive to keep you, not to squeeze you. Whether you are running an agentic harness that burns tokens all day or a single product feature, it should be affordable enough that you stop rationing it.
Side by side
The same tokens, priced two ways
| Model | List | Your price | You save |
|---|---|---|---|
Claude Opus 4.1 Anthropic · 200K | $15.00/$75.00 | $7.50/$37.50 | 50% |
Claude Sonnet 4.5 Anthropic · 200K | $3.00/$15.00 | $1.50/$7.50 | 50% |
Claude Haiku 4.5 Anthropic · 200K | $1.00/$5.00 | $0.600/$3.00 | 40% |
GPT-5 OpenAI · 400K | $1.25/$10.00 | $0.688/$5.50 | 45% |
GPT-5 mini OpenAI · 400K | $0.250/$2.00 | $0.163/$1.30 | 35% |
GPT-4.1 OpenAI · 1M | $2.00/$8.00 | $1.10/$4.40 | 45% |
USD per 1M tokens, shown as input / output.
See all models →Assumes a 70/30 input-to-output split and no prompt caching. Cached reads make the gap wider, not narrower.
You’d save
$120/mo
- Direct from provider
- $239.63
- Through lowcostllm
- $119.82
- Discount
- 50% off
For people who live in a terminal
Or use our agent and skip the API entirely
Everything you’d want from a terminal coding agent, running on whichever model you’re paying half price for.
- Reads and edits your repo, runs commands, and shows you the diff before it touches anything.
- Switch models mid-session — start on a cheap one, escalate only when the task earns it.
- Per-project config, permission rules, and an audit trail of every command it ran.
- Open source, and a fork of pi.dev — so nothing about it is a black box.
$ npm i -g @lowcostllm/agent
$ lcllm auth login
$ cd ~/code/your-project
$ lcllm
ready · claude-sonnet-4-5 · 50% off list
> refactor the billing module to use the new
Invoice type and update the tests
read src/billing/*.ts (7 files)
edit src/billing/invoice.ts +48 -12
edit src/billing/charge.ts +9 -3
run npm test 42 passed
done in 1m 12s · 84k tokens · $0.21Data handling
What happens to your prompts
We don't train on anything.
Not your prompts, not your completions, not your metadata.
There is no model of ours to train — we do not have weights.
Logging is off by default.
We keep token counts and timestamps for billing, and nothing else.
You can turn on full request logs per-key if you want them.
US-hosted infrastructure.
Served from [US-EAST-1 / US-WEST-2], with upstream calls to US provider endpoints.
No request or response data routes through China.
Full detail in docs / data handling.
Questions
The things you’d ask a reseller
We put the uncomfortable ones at the top on purpose. Ducking them is what makes discount resellers look like scams.
Still unsure? Ask us directly.
What happens when your provider contract expires?
Rates are fixed through [YYYY-MM-DD] under our current agreement. If our costs change after that, you get 60 days’ written notice before any price change, and your existing keys keep working at the old rate for the whole notice period. We will not reprice you mid-month or mid-project.
If we ever cannot renew on terms that keep the discount real, we will tell you that instead of quietly narrowing the gap until it is gone.
What happens if you go out of business mid-project?
This is a new kind of business and we are not going to pretend the risk is zero. Here is how we make it small: you buy credit in advance and spend it down, so the most you can ever lose is the balance you chose to top up — not a year of prepaid commitment.
Keep your balance near your actual burn rate and your exposure stays a few days wide. Migration back is a one-line change, so the switching cost if we disappear is measured in minutes.
Is this the same model, or a quantized version?
The same model. We do not host, quantize, distill, fine-tune, or otherwise touch the weights — we do not have them. Your request is forwarded to the provider’s own API and their response is returned to you unmodified.
If you want to verify that rather than take our word for it, run your own evals against both endpoints. We will give you the credit to do it.
What are the rate limits, and do I share them with other customers?
Limits are per-key and tiered by how much credit you have added — the full table is on the pricing page.
Be aware that below the Dedicated tier, your throughput is drawn from a shared pool of contracted capacity. In practice that is invisible, but under a global spike it means queueing rather than a private lane. If you need capacity that is genuinely yours and isolated, that is what the Dedicated tier is, and it is a contract rather than a checkbox.
Do you support streaming, tool use, vision, batch, and prompt caching?
Streaming, tool use, vision, structured outputs, extended thinking, and prompt caching all work exactly as they do upstream, because we pass the request through untouched. Batch works too, but at list price with no discount — we say so on the pricing page rather than letting you find out later.
The full support matrix lists every feature per provider.
How do I migrate back if I want to leave?
You change the base URL back. That is the entire process. We do not have a proprietary SDK, a bespoke request format, or stored state that belongs to us — nothing about your integration is specific to us except one string. A provider you cannot leave in under a minute has leverage over you, and we would rather not have that kind of leverage.
Which models are covered by the discount?
Everything listed on the pricing page, with the per-model discount shown against each row. Discounts vary by model because our contracted rates do. Any model not in that table is not available through us yet — we would rather show a short honest list than a long one with asterisks.
Do you see my prompts?
Your requests pass through our infrastructure in memory in order to be forwarded — that is unavoidable for a proxy, and anyone claiming otherwise is misleading you. What we do not do is persist them. Request and response bodies are not written to disk or retained after the response is delivered, logging is off by default, and we never train on anything. You can enable request logging per-key if you want an audit trail, and that is your choice to make.
Frontier models. Half the invoice.
Change one line, keep everything else, and check the bill at the end of the month.