Compare · Alternative · Checked on 27 October 2026
Hopper and Petrova.
Hopper hosts fast, cheap open models and is live. Petrova is one balance with spend rules across suppliers, and is not live. This is Hopper Inference, not the travel app.
Hopper is live. Petrova is a preview.Petrova runs on test data only. It has no real supplier, and you cannot buy it today.
How Hopper works
Fast, cheap open models for real-time voice agents.
- You make a key in its console, or your coding agent gets a trial key itself and you keep it with a short code.
- Your agent measures speed on your own prompts and shows you a table before changing any code.
- You switch your LLM to its OpenAI-format API. It gives tested code for voice frameworks.
- It hosts open models, plus speech-to-text and text-to-speech.
- You pay from prepaid credits. Each answer shows the cost and the balance in its headers. At zero balance, calls stop.
How Petrova works
The account your agent calls. Today this all runs on test data.
- Your agent keeps running where it runs today. Petrova does not run it.
- You change two settings: the base URL and the key. Anthropic and OpenAI formats both work.
- Before each call, Petrova holds the most that call could cost.
- A supplier answers. If one fails before the first word, the call moves to another.
- You are charged what the call used, plus a 5.5% fee. The rest comes back. Never above the hold.
Side by sideChecked on 27 October 2026
| Hopper | Petrova | |
|---|---|---|
| Live today | Yes | No. Preview on test data |
| Runs its own models | Yes | No. Petrova owns no GPUs |
| Open-model price, per million in / out | Gemma 4 31B: $0.30 / $0.40. DeepSeek V4.1 Flash: $0.09 / $0.85 | No open models. Cheapest: Haiku 5.5 at $0.1055 / $0.5275 (not on sale) |
| Speech | Text-to-speech, speech-to-text, voice cloning | No |
| Agent gets a trial key by itself | Yes, live | A key at $0 in one call, test mode only. No trial money |
| Cost and balance on every answer | Yes, in headers | Planned. Today, read the receipt |
| MCP connector with tools that need no account | Yes, live | MCP server on test data; cost tools need no key |
| API formats | OpenAI | Anthropic and OpenAI (test data) |
| Claude models | Not found | Yes (test data) |
| Spend limits | Stops at zero balance. Caps: not found | Account, project, key, agent, session (test data) |
| Charge never above a hold set before the call | Not found | Yes (test data) |
| Failover to another company | Not found | Yes (test data) |
| Tools and memory on one balance | Not found | Yes (test data) |
Green marks the stronger answer in a row, red the weaker. "Test data" means Petrova's code does it in tests, not as a service. "Not found" means we did not find it on their public pages on the check date. It may still exist.
Where Hopper is ahead
- Live today
- Runs its own models
- Open-model price, per million in / out
- Speech
- Agent gets a trial key by itself
- Cost and balance on every answer
- MCP connector with tools that need no account
What we could not check
- Its speed claims under real load.
- Whether every model runs on its own GPUs.
- Its customers. None are named.
Read on 27 October 2026, on their public pages: withhopper.com · withhopper.com/skill.md · docs.withhopper.com · its MCP model list. Things may have changed since. Each claim of theirs is their own statement.