/aicg/ meta

Comparison between the different services/models and frontends used by /aicg/.
These ratings aren't gospel. They're the opinion of one anon who tried to incorporate suggestions and push no agenda.

Services/models

If you're curious about a service, the OP should have the information you need.

🏆 - Best
🥈 - Great
🥉 - Good
🌱 - Usable
⛔ - Bad
🔀 - Depends
NSFW - general ERP-readiness (taking into account the need to jailbreak)
SFW - general RP-readiness
MEMORY - more size than retention, more memory = higher price for both cloud and local
CREATIVITY - low only for dumb assistants, can be increased with a CoT prompt
COMPLEX INSTRUCTIONS - stats tracking, logical conditions, multiple characters, etc.
PRICE - per token, per month, or for running local models one time purchase or per hour rental of the hardware
EASE OF USE - OAI keys or all proxies are easy, but still require JBs; Horde and subscription based local/NAI offer inbuilt nobrain presets; running local is not that hard because there's enough guides; anything can require shuffling prompt parts and finding the best temperature and other settings
/lmg/ - local models in general, the ones you can download and run offline, the best ones in the rankings

Notes:

• Gemini API: free access is rate limited.
• NovelAI models:
  • You must use "Generate more" functionalities to get longer replies
  • Prompting properly may require some tweaking, there's not enough guides on chat usage
• Local models:
  • Check out /lmg/
  • Small brain/short memory models can be run on very basic hardware, but size+memory+speed vs. requirements curve is very steep
  • The settings and presets are offered as a single file to be easily loaded
• OpenAI o1: not enough conclusive data.
• Grok API: free credits until the end of 2024.

Quick notes on how to access each:

• GPT:
  • Official API, revendors.
  • Or use a reverse proxy that has it. (agnai video guides / tavern guide)
• Claude: Official API, revendors.
• Claude (AWS): Pay for it https://aws.amazon.com/bedrock/pricing/
• /lmg/
  • Use a reverse proxy or Horde.
  • Pay for a subscription.
  • Some services offer free limited API access to small models.
  • Or learn to run it locally, on Colab or via GPU rental services.
• Gemini API
  • Use your own API key (https://makersuite.google.com/app/apikey)
  • Reverse proxy if available (they support OpenAI-formatted input)
• Mistral
  • Official API, revendors, via native support on ST/Risu, or just using it like a reverse proxy since it supports a subset of OpenAI input format.
  • https://docs.mistral.ai/platform/pricing
• Cohere Command R+: Official API (https://dashboard.cohere.com).
• Grok: Official API, revendors.

Frontend (service support) (outdated)

• Character Card V3 Implementation Tracking: https://rentry.org/zdwfrgqq

Three frontends are currently recommended, Silly, Agnai, and Risu. All have a lot of different features and are improving fast.

Miku is noteworthy, but difficult to compare. It uses sentiment classification to show different sprites for each message, and text-to-speech, supporting OAI, Kobold, Whisper, AzureTTS, ElevenLabs, and NovelAITTS.

Gemini works on all frontends if using it via reverse proxy. Mistral API supports OpenAI input (using the official API like a reverse proxy).

Frontend (features) (outdated)

• Only agnai has support for choosing which character speaks at any time in multi-user rooms, although on the flipside it doesn't support fully automatic mode
• (ext) means you have to install the Silly extras. They can be a little complicated to setup for non-technical users.

Outdated uninteresting information

• Poe: Create a bot over at https://poe.com/. Example: https://poe.com/coolreisen
• Slaude: it's still viable for sfw purposes for anons with actual enterprise accounts
• Bing: there are ways to use this version of GPT-4 for free, but it's "filtered"
• Clewd: there are ways to use claude.ai version of Claude for free, but it's "filtered"
• text-davinci-003 has been removed as OpenAI is discontinuing it.
• Davinci-3: Use an OpenAI key directly in SillyTavern/agnai. You can use your own or buy one (search "buy openai key" on yandex).
• Turbo: Same as davinci-3 but you can also use a reverse proxy.
• Gemini Advanced: the rating is provided for Gemini 1.0 Ultra, which was served to subscribers in the beginning. Pay for a subscription.
• Consensus on Gemini 1.5 Pro is: better than 3.5 Turbo, worse than 4 Turbo.