About Our AI
Fight Health Insurance uses AI to draft your appeal letter. You read it, change what you want, and choose what to send. AI is a tool to help you, not a replacement for medical or legal advice. We also search PubMed for medical research your appeal can cite.
The models we use
Several models can write drafts for the same appeal, and you choose the one to send. Outside services only get your case if you allow outside models when you start your appeal or a chat. That box starts ticked, and you can untick it.
This list changes. We add and retire models as needed, and update this page when we do.
| Model | From | What it does | Where it runs |
|---|---|---|---|
| FHI model | Fight Health Insurance | Writes your appeal draft. Our main writer. | Our own server |
| Nemotron 3.5 Super | NVIDIA, through DeepInfra | Writes a backup draft. One of the first outside models we ask. | Outside service, only if you allow outside models |
| DeepSeek V4-Pro (0813 release) | DeepSeek, through DeepInfra | Writes a backup draft. | Outside service, only if you allow outside models |
| Gemma 4 26B | Google, through DeepInfra | Writes a backup draft. Fills in for our model on summaries and questions. | Outside service, only if you allow outside models |
| Mistral Small 3.2 | Mistral AI, through DeepInfra | Answers in the chat alongside our model. Can also write a cover letter to a regulator or your plan's medical director. | Outside service, only if you allow outside models |
| GLM-5.3 Flash | Z.ai, through DeepInfra | Answers in the chat alongside our model. Can also write a cover letter to a regulator or your plan's medical director. | Outside service, only if you allow outside models |
| DeepSeek V4.1 Flash (1017 release) | DeepSeek, through DeepInfra | Answers in the chat alongside our model. Can also write a cover letter to a regulator or your plan's medical director. | Outside service, only if you allow outside models |
| Qwen3.8 Max | Alibaba, through DeepInfra | Answers in the chat alongside our model. Can also write a cover letter to a regulator or your plan's medical director. | Outside service, only if you allow outside models |
| Kimi K3 | Moonshot AI, through DeepInfra | At important moments in a chat, gives a second answer you can compare with ours. | Outside service, only if you allow outside models |
| Perplexity Sonar | Perplexity | Finds sources to cite, and suggests questions about your case. Doesn't write letters. | Outside service, only if you allow outside models |
| Jev | TypeSafe | Doesn't write letters. Reads your denial letter for key details, ranks the drafts, checks research before we cite it, and checks chat answers. | Outside service, only if you allow outside models |
Changes
- 2026-10-31: Moved to DeepInfra's newer releases of DeepSeek V4-Pro and DeepSeek V4.1 Flash, and replaced Qwen3.8, which DeepInfra stopped serving, with the cheaper Qwen3.8 Max.
- 2026-11-01: Added NVIDIA's Nemotron 3.5 Super to the outside models we ask for a backup draft, near the top of the list.
- 2026-10-31: Stopped using DeepSeek V4-Pro. This was the preview. DeepInfra published the finished DeepSeek V4-Pro (0813 release), at the same price, so we moved to it.
- 2026-10-31: Stopped using DeepSeek V4.1 Flash. DeepInfra published an updated release (1017) at the same price, with shorter answers, so we moved to it.
- 2026-10-31: Stopped using Qwen3.8. DeepInfra stopped serving it on 2026-10-12. Replaced by Qwen3.8 Max, the same Qwen flagship at a lower price.
Which model writes the best appeal?
We do not take the models' word for it, or only our own tests' either. In the Appeal Arena, visitors read several AI drafts for the same made-up case side by side, without knowing which model wrote which, pick the one they would send, and flag any draft that invents a study, a law or a test result. Those votes help train the selector that picks drafts in the real appeal flow.
Sustainability
We care about the environmental impact of running AI. A few things we do:
- Low-power inference: Our own models run on a single GPU, using a fraction of the energy that large cloud-hosted models need.
- Solar powered when possible: During the day, our inference servers run on solar power. We aim to shift compute to daylight hours when solar generation is available.
- Efficient serving: We use vLLM, a high-throughput inference engine that maximizes the work done per watt.
Open Source
Fight Health Insurance is mostly open source. You can find our code on GitHub. We believe transparency is especially important when AI is involved in healthcare.
For how we handle your data, see our Terms of Service. You can also request deletion of your data at any time.
Limitations
AI-generated appeals are a starting point, not a finished product. Our models can make mistakes, hallucinate citations, or miss important details about your specific situation. Always:
- Review the generated appeal carefully before sending it.
- Verify any cited medical literature or regulations.
- Add personal details about your medical history that the AI does not know.
- Consult a healthcare professional or attorney if your situation is complex.