~50% lower API line on your grant
30% off published list rates + 2× prepaid credit match. A $5,000 API budget effectively becomes $10,000 of inference. Same Anthropic / OpenAI / Google / DeepSeek models — we forward, not modify.
Anvat for researchers
Frontier models at half the cost. Reproducible, citable, batchable.
AI research is grant-funded. Anvat cuts the API line of your budget by ~50% without sacrificing model fidelity. Same Anthropic / OpenAI endpoints behind your existing SDK — every run reproducible against the published model id.
- OpenAI-compatible
- Anthropic-compatible
- MCP-ready
- One API key
- 30% off list
- 2× credit on prepaid
- No card to start
- 5 languages
Why researchers use Anvat
Reproducibility built in
We pin to the same model IDs the provider publishes (claude-opus-4-8, gpt-5, gemini-3-5-flash). Cite the model id in your paper — anyone can re-run the experiment against api.anvat.app or api.anthropic.com and get bit-identical responses.
Per-request cost telemetry for clean grant accounting
Every response carries x-anvat-cost-usd. Tag each batch with your grant code in the dashboard. Export CSV monthly for the finance office — no more guessing where the spend went.
Built for batch + long-running jobs
No per-seat licence, no minimum monthly. Pay only for what you call. Background eval runs, multi-day batch labelling, longitudinal benchmarks — they're all cheaper at Anvat than going direct.
Where the math actually changes
| What hurts today | What changes on Anvat |
|---|---|
| API budget runs out 2 months before the grant cycle ends | 30% off every token + 2× prepaid match — typical grant covers 2× more inference |
| Same eval re-run on the new model version — costs double the original budget | Pin model id in your paper. We forward to the same upstream — bit-identical responses. |
| Finance office wants line-item cost attribution per experiment | Per-request cost header + CSV export. Tag runs with grant codes for clean attribution. |
| Anthropic Pro plan rate limits cap throughput during paper crunch | Subscription plans grant higher RPM than Anthropic consumer Pro at similar price. |
| Reproducibility paper-trail messy when responses came from a third-party wrapper | No prompt/completion logging on our side. Only billing metadata. Reproducibility intact. |
Recommended plan
Pay-as-you-go (Free) + prepaid pack
Most research workflows are bursty (batch eval runs separated by analysis weeks). Pay-as-you-go avoids paying for idle months. Buy a prepaid pack right before a big run to get 2× credit match — a $500 pack becomes $1,000 of inference. Move to Pro X5 / Ultra only if you have sustained > $500/month for 3+ months running.
See plan details →Questions researchers ask before signing up
→Are the responses bit-identical to going direct to the provider?
Yes for non-streaming responses, modulo upstream non-determinism (temperature > 0, sampling, etc. — same caveats apply going direct). We forward the request unchanged to the provider's published endpoint and return what comes back. For reproducibility runs, set temperature=0 and pin the model id — you'll get the same answer from us as from api.anthropic.com.
→Can we cite Anvat in a paper?
You can. The cleanest citation pattern: cite the model id (e.g. 'claude-opus-4-8, accessed via the Anvat gateway, May 2026') so readers can re-run against either Anvat or the upstream provider. Anvat is a transport layer — the inference results are the provider's.
→Do you store prompts or completions for any purpose?
No. We log billing metadata only — model id, token counts, latency, status, cost. Prompt and completion bodies are not stored or logged. This matters for IRB / institutional review boards that require no third-party retention of human subject data.
→Can we get an institutional invoice / W-9 / EIN?
Yes. Stripe issues a real invoice with your institution as the bill-to. We can supply a W-9, EIN, and vendor-setup forms — email hello@anvat.app with the form and we'll turn it around in a day or two.
→Do you support batched inference / async jobs?
Anvat is a real-time gateway — every call is forwarded immediately. For latency-tolerant batch work, use Anthropic's native /v1/messages/batch (forwarded by Anvat) which gets the upstream 50% batch discount on top of our 30% off. Net effective: ~65% off list. Excellent for eval runs.
Ready to cut your AI bill?
Free signup, $2 credit, no card required. The SDK doesn't change — only the bill does.