An inference platform for building voice AI agents. It lets you find and try speech recognition (ASR), speech synthesis (TTS), and the large language models (LLMs) that compose the replies — all in one place — and then deploy them as production endpoints. Hathora, a US company known for game server hosting infrastructure, launched it on November 12, 2025 to address a new use case: real-time voice processing. The goal is to let you assemble low-latency voice pipelines without running your own GPU servers or Kubernetes clusters.
That said, its availability needs a caveat. In March 2026 Hathora announced that it was joining Fireworks AI; the game server business shut down on May 5, 2026, and existing customers were pointed to GameFabric by Nitrado for migration. As of August 2026 models.hathora.dev is unreachable, so whether the service is still operating cannot be confirmed on the official site. Read the following as what was known at launch.
Key Features
- All the models a voice agent needs in one catalog: ASR, TTS, and LLMs — the three model types essential to a voice agent — are listed side by side so you can search across them for the right fit
- Try models in the browser: Pick a model, send it audio or text, and check response speed and quality on the spot. You can catch mismatches like “slower than expected” or “the voice doesn’t fit” before you commit
- Design voice pipelines by combining models: Build and evaluate the full flow of listen (ASR) → think (LLM) → speak (TTS) with a different model at each stage. Because a voice agent’s perceived speed is the sum of every stage, this combination testing matters
- Production endpoints without infrastructure work: Publish the model you tested as an inference endpoint directly. Securing GPUs, orchestrating containers, and tuning scale are not yours to run
- Multi-region delivery for lower latency: Traffic is served from several regions so processing happens near the user, shortening round-trip time. In spoken conversation, a few hundred milliseconds decides whether the exchange feels natural
- Learning documentation: The official Learn page collects the basics — the building blocks of voice AI and how to think about latency
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Pay-as-you-go | Billed by usage | You pay for tokens and audio processed, with no charge for idle GPUs |
| Dedicated endpoint | Billed by GPU running time | A dedicated instance for stable, predictable throughput |
| Enterprise | Contact sales | For large-scale use and custom requirements |
Pricing information is current as of August 2026. Hourly rates for dedicated GPUs were reportedly published at launch, but because the official pricing page cannot be reached as of August 2026, the exact figures are unconfirmed. Please check the official site for the latest pricing.
Pros & Cons
✅ Pros
- Compare all three model types a voice agent needs on one screen, cutting the time spent on selection
- Trial and production deployment are continuous, so you avoid maintaining separate test and production setups
- You can build a latency-conscious architecture without operational knowledge of GPUs or container platforms
- The multi-region operations know-how built up on game server infrastructure fits real-time voice well
⚠️ Cons
- As of August 2026 the official site is unreachable, so the service’s continued availability cannot be confirmed
- Hathora announced in March 2026 that it was joining Fireworks AI, making the future of the offering hard to read
- Current pricing and the model lineup cannot be verified against primary sources, leaving little basis for a production adoption decision
- Launched only in November 2025, it has almost no long-term operational track record and very little Japanese-language coverage
Comparison with Similar Services
| Criteria | Hathora Models | Fireworks AI | Together AI | Deepgram |
|---|---|---|---|---|
| Provider | Hathora (USA) | Fireworks AI (USA) | Together AI (USA) | Deepgram (USA) |
| Main role | Trying and deploying voice models | Fast inference for open models | Inference and training for open models | Speech recognition and synthesis API |
| Voice focus | Dedicated to voice AI | General purpose (voice included) | General purpose | Voice focused |
| Model breadth | ASR, TTS, LLM | LLMs, images, and more | LLMs, images, and more | Mainly its own voice models |
| Availability | Official site unreachable as of August 2026 | Available | Available | Available |
Who Is It For
- Developers building services with real-time spoken dialogue, such as phone support or voice assistants
- Developers who want to compare ASR, TTS, and LLM combinations without spending time procuring and operating GPUs
- Teams isolating the sources of latency in voice AI to tighten perceived response speed
- That said, because the official site cannot currently be confirmed, verifying that the service is still offered is a prerequisite before considering it for production
Summary
Hathora Models arrived in November 2025 as a developer platform for trying the ASR, TTS, and LLM pieces a voice agent needs and shipping them straight to production. Bringing multi-region operations honed on game server infrastructure to voice — a domain where latency is experience quality — is a sound idea. On the other hand, Hathora announced in March 2026 that it was joining Fireworks AI, and as of August 2026 the official site is unreachable. If you are considering it, start by confirming that the service is still being offered.