Skip to main content

LLM providers and prompts

How to connect is in Connect an LLM. This page collects the differences that help you choose.

Providers​

ProviderModelsNotes
Amazon BedrockUp to 23 region-specific models across Nova, Claude, GPT, and moreOne key covers models from several providers. Round trips for short input were 2.7s for Sonnet 5 and 3.3s for Opus 5 (us-east-1).
Anthropic ClaudeHaiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1Called directly with a key from your Anthropic account (sk-ant-). Fable is not on the Bedrock list and can only be chosen here. All four have thinking on. Response times for these models have not been measured yet.
Google Gemini3.8 Flash, 3.6 Flash, 3.5 Flash-Lite (default)3.5 Flash-Lite was the most stable on the free tier. Often leaves out periods. The response time of 3.8 Flash has not been measured yet.
OpenAI (GPT)GPT-6 Luna (default), GPT-6.1 Terra, GPT-6.1 Sol, GPT-6 AstraUses the Codex CLI sign-in (codex login). Which models you can use depends on your plan. Response times for these models have not been measured yet.
Groq CloudGPT-OSS 120B (default), GPT-OSS 20BUses the same key as Groq STT. Response times for these models have not been measured yet.
Local MLXA downloaded modelText never leaves your Mac.

These timings were measured in one development environment and vary with network and region.

The Bedrock picker only shows models verified in the selected region. v0.34.0 added Claude Haiku 5.5 and GPT-6 Sol; thinking is on by default for both, so even short replies consume output tokens. v0.34.1 removed Claude Haiku 4.5 from the list. If you were using it, Sogon automatically migrates the saved setting to Claude Haiku 5.5; no manual change is needed.

The direct Anthropic models moved to newer versions in v0.34.1. A model you were using migrates automatically to the newer version of the same family: Haiku 4.5 to Haiku 5.5, Sonnet 5 to Sonnet 5.5, Opus 5 to Opus 5.5, and Fable 5 to Fable 5.1.

v0.34.4 removed 3.7 Flash and 3.5 Flash from the Gemini list. The official docs say neither model is being shut down; the service routes them to 3.8 Flash and 3.6 Flash instead. If you were using one, Sogon automatically migrates the saved setting to the same target, and the model that answers does not change.

The direct OpenAI connection signs in through ChatGPT (Codex), so its model list follows the Codex model docs, not the OpenAI API list. v0.34.4 replaced the list with four GPT-6 models. GPT-5.5 was retired from Codex with ChatGPT sign-in on October 14, 2026, GPT-5.4 and GPT-5.4 Mini on August 31, and GPT-5.3 Codex and GPT-5.2 are deprecated. A model you were using migrates automatically, following the docs' migration guidance, so no manual change is needed: GPT-5.5, GPT-5.4, and GPT-5.3 Codex move to GPT-6.1 Sol; GPT-5.4 Mini, GPT-5.3 Codex Spark, and GPT-5.2 move to GPT-6 Luna. The docs recommend GPT-6 Luna on Free and Go plans, and on Enterprise and Edu an administrator has to enable Luna first. If you choose a model your plan cannot use, correction may not apply and the original text may be inserted as is; switch to GPT-6 Luna in that case.

The Groq list shrank to GPT-OSS 120B and GPT-OSS 20B (v0.34.4). Groq shut down Qwen3 32B and Llama 4 Scout on July 17, 2026, and Llama 3.3 70B and Llama 3.1 8B on August 16, for free and developer-tier use. The old default was Qwen3 32B, so the default is now GPT-OSS 120B. A model you were using migrates automatically to the replacement Groq recommends: Qwen3 32B, Llama 4 Scout, and Llama 3.3 70B move to GPT-OSS 120B; Llama 3.1 8B moves to GPT-OSS 20B.

Correction modes​

ModeWhat it doesOriginal protection
Standard (STT Correction)Fixes misrecognitions, spacing, and English term spellingIf more than 50% of words change, the original is used.
Filler RemovalRemoves fillersSame as above.
StructuredOrganizes into lists and paragraphsNot applied, since changing the shape is expected.
CustomUses the prompt template you chooseNot applied.

After correction, English terms that were in the original but disappeared are restored, and the Dictionary corrections are applied once more.

Prompt templates​

Edit templates in Settings → LLM. There are seven built-in templates:

TemplateUse
General correctionPolishes dictation.
LLM promptTurns a request you described out loud into a structure that works well when pasted into an LLM.
Meeting notesUsed for the final pass of the minutes output.
SummaryCondenses the material into a one-line summary and key points.
Action itemsPulls out only the follow-ups that were agreed, with owner and due date when stated.
TranslationTranslates Korean to English and English to Korean, leaving code and identifiers alone.
Email draftTurns a spoken explanation into an email with a subject line and body.

The last four are built for long transcripts, so they fit History → Session → LLM request. They also work as the dictation template.

  • Set a default template for dictation and one for minutes.
  • Preview results with sample input.
  • Deleting a template asks for confirmation. Deleted built-in templates come back with Restore built-ins; user templates cannot be recovered.
  • Updating the app does not change built-in template text you already have. To get the new text, reset the template to its default in the editor.
  • Built-in templates that are new in an update are added to the end of your list automatically. A built-in you deleted earlier does not come back — that only happens through Restore built-ins.

What is sent to the LLM​

Correction sends two things:

  1. A system prompt: the mode or template text, input-handling instructions, Dictionary corrections, and the active word list
  2. The text to correct

Sogon doesn't send previous utterances or conversation history, and it doesn't send screen region images to the LLM either.