Skip to content

AI Providers & Models

There are no API keys to manage. You pick a provider and model for each AI capability in Settings, and FeynmanLM bills the usage to your prepaid balance at that provider's own API list price plus a 15% margin — metered on the exact cost the provider reports, not an estimate. New accounts start with $5 of credit.

You can also skip the built-in AI entirely and drive FeynmanLM from a Claude, ChatGPT, Gemini, or Grok subscription you already pay for, connected via MCP at 1¢ per tool call. Sources, Schedule, and semantic search work either way. Most people mix the two.

Chat and reviews (text models)

In-app Chat, Feynman reviews, and background tasks (like paper metadata extraction) use the text provider and model selected in Settings → AI Provider. The default is Anthropic with Claude Sonnet 5.

ProviderModelsDefault
AnthropicClaude Sonnet 5, Claude Fable 5, Claude Opus 4.8, Claude Haiku 4.5Claude Sonnet 5
OpenAIGPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 LunaGPT-5.6 Sol
Google (Gemini)Gemini 3.5 Flash, Gemini 3 Pro Preview, Gemini 2.5 Flash LiteGemini 3.5 Flash
xAI (Grok)Grok 4.3, Grok 4.5Grok 4.3
OpenRouterDeepSeek V4 Flash, DeepSeek V4 Pro, Kimi K3, GPT OSS 120B, GPT OSS 20B, Qwen3 32B, Qwen3 235B, GLM 5.2, Muse Spark 1.1, Nemotron 3 Ultra, Nemotron 3 Super, Nemotron 3 NanoDeepSeek V4 Flash
Together AIInklingInkling

Notes:

  • All open-weight models — DeepSeek, Kimi, Qwen, GLM, GPT OSS, Muse Spark, and Nemotron — are served through OpenRouter. The picker groups them by the lab that trained them, and you can add any other model from OpenRouter's directory. Together AI remains a chat provider solely for Thinking Machines' Inkling, which isn't listed on OpenRouter.
  • DeepSeek and Kimi are offered only through OpenRouter's US-routed inference; the labs' own APIs are deliberately not integrated.
  • The model picker shows a curated list; you can hide models you never use in Settings → Chat Models.
  • Each provider remembers its own model choice, so switching providers doesn't reset your selection.

Read Aloud (text-to-speech)

Read Aloud generates audio with the provider selected in Settings → Text-to-Speech. The default is OpenAI.

ProviderModelsVoices
OpenAI (default)GPT-4o mini TTS (default), TTS-1, TTS-1 HD13 voices (default: Sage)
Google (Gemini)Gemini 3.1 Flash TTS (default), Gemini 2.5 Flash TTS, Gemini 2.5 Pro TTS30 voices (default: Kore)
GroqOrpheus6 English voices
Together AIOrpheus (default), Kokoro, Cartesia Sonic 3.5/3/2Orpheus, Kokoro, Cartesia, Rime, and Minimax voice banks
ElevenLabsMultilingual v2 (default), Flash v2.5, Turbo v2.5, v310 preset voices
xAIGrok TTS5 voices
LocalQwen3-TTS on your own local serverFree — runs on your own machine

Gemini also supports overnight batch audio: queue a source in the evening and FeynmanLM generates the audio through Gemini's Batch API at half the live-request price.

Voice input (speech-to-text)

Voice dictation in Chat uses the provider selected in Settings → Speech-to-Text:

  • Groq (default): Whisper Large v3 Turbo
  • OpenAI: GPT-4o mini Transcribe

Audio is sent only to your configured provider and is not persisted after transcription.

Image generation

  • Apple Image Playground (default): on-device, free, nothing billed
  • Google (Gemini): Gemini 3.1 Flash Image
  • OpenAI: GPT Image 2

Semantic source search runs entirely on-device using Apple's NaturalLanguage contextual embeddings. It costs nothing and sends nothing over the network.

Costs

  • Every provider above is available immediately — nothing is behind a tier, and there is no key to obtain. Signing in with your Apple account is the only setup.
  • Settings → AI Usage tracks every request against the provider's list price plus the 15% margin, so you can see exactly what your learning costs, to the cent. Models without a published rate are shown as unpriced rather than guessed.
  • Your balance carries over — credits don't expire and don't reset monthly. Top up when you run out.
  • Background tasks (paper metadata extraction, transcript cleanup, search embeddings) draw on the same balance and are itemised per task in the ledger. Each one's model is configurable in Settings → AI Models → Background Tasks, so you can point the noisy ones at something cheap.
  • If a provider or model can't complete a request, FeynmanLM surfaces the error — it never silently falls back to a different model or to heuristic grading.