About
What Anvat is, and what it isn't.
A transparent account from the team — why we built this, how the money flows, and what we refuse to do.
Why Anvat exists
Frontier LLMs are powerful. They are also genuinely expensive when you pay full list price, and a pain to switch between when each provider wants its own auth, billing, and SDK.
We were running Claude Code, Cursor, custom agents, and long-running batch jobs ourselves. The bill kept arriving and kept being avoidable — the same tokens going through a gateway with sane defaults cost roughly half. So we built the gateway we wanted to use, then opened it.
That's the whole story. There is no enterprise sales motion behind a marketing-friendly cover. The product is the honest path to lower-cost frontier inference, and the business model is whatever surplus is left after we pass the wholesale rate through to you.
How we make money (the actual mechanics)
Nobody likes opaque pricing. Here is the literal flow of a dollar through Anvat:
- You pay $100 (prepaid credit pack). Your account balance shows $200 — we 2× the prepaid amount up to the published cap.
- You spend tokens. Each request is metered at the upstream provider's published list price × 0.70 (we apply a 30% discount). The cost is deducted from your $200 balance.
- We pay the upstream provider list price. Anthropic, OpenAI, Google, DeepSeek — whatever your request hits. We have negotiated wholesale rates that make this economical at our volume.
- The spread between wholesale and what we charge you is our revenue. It funds engineering, infrastructure, support, and the blog/leaderboard/benchmark research you're reading right now.
When prepaid 2× match is in effect, your effective cost per dollar of usage lands around $0.50. We're profitable on the long-term LTV because wholesale + steady usage rate clears the upfront credit match within a typical user lifetime. We're explicit about this because the math is the product.
What we will not do
- No prompt or completion logging. We store billing metadata only (model id, token counts, latency, status, cost). Your message bodies are not on our disks. Verifiable: turn on full request inspection in Cloudflare and watch the upstream POST stream by.
- No silent model substitution. When you ask for
claude-opus-4-8we forward toclaude-opus-4-8. Bit-identical responses (modulo upstream non-determinism) vs going direct. - No hidden per-request markup beyond the published discount. Every response carries an
x-anvat-cost-usdheader. Add it up — that's your bill. We do not round, we do not pad. - No fake testimonials, no fake user counts. We are early. We refuse to invent social proof. If the product is good (it is), the math (which is also good) will earn the customers honestly. The day we have real logos to show, we'll ask permission and show them.
- No lock-in. You moved one env var to use us. Point it back at
api.anthropic.comany time. There is no clawback, no exit fee, no contract. We earn keep on the math, not the contract.
Who's behind it
Anvat is independently founded and built by a small team of engineers who use the product daily. We don't have a hundred-person sales org. We don't have outside investors yet. We have a working gateway, a profitable wholesale relationship, and a slowly-growing customer base that pays for inference at a discount.
You can reach the team directly:
- Email: hello@anvat.app
- Telegram: @anvat_app — we answer
- Status / uptime: /status
- Press / brand assets: /press
Roadmap (what we're working on)
The full public roadmap lives at /roadmap — Now / Next / Later with rationale for each item. We move entries to a release note the day they ship.
The math is the product. Try the math.
Free $2 signup credit, no card required. The SDK doesn't change — only the bill does.