ToRun lets you use AI models from many providers, among them OpenAI, Anthropic, Google, Mistral, Groq, Together and ElevenLabs. Each provider sets its own prices and decides when its older models stop working. We read their published price lists and retirement notices and change ToRun to match, so that a chat keeps working and is charged according to what the provider charges. This update carries one round of those changes.
When a provider retires a model
When a provider switches a model off, ToRun does one of two things. If the provider names a model that replaces it, a chat set to the old model runs on the replacement and is charged the replacement's price. If no replacement is named, the model leaves the model list, and a chat still set to it is told that the model has been retired instead of failing.
In this update:
- Gemini 2.5 Pro and Gemini 2.5 Flash. Google has announced that from 1 December 2026 it answers every Gemini 2.5 Pro request with Gemini 3.8 Flash and every Gemini 2.5 Flash request with Gemini 3.5 Flash-Lite. From that day ToRun does the same, at the newer model's price, which is the same or lower. Where one of them was among a plan's default models, the newer model has already taken its place.
- Gemini 2.5 Flash Image was retired earlier. Image requests that still name it are now made with Gemini 3.1 Flash-Lite Image.
- Groq has switched off Kimi K2 Instruct, Llama 3.3 70B Versatile and Llama 3.1 8B Instant. Chats set to them run on the replacements Groq names: GPT-OSS 120B and GPT-OSS 20B on Groq. Groq has also switched off Qwen3.6 27B, which has no replacement in ToRun yet.
- Together has removed five models from its service without a replacement, among them Llama 3.3 70B Instruct Turbo and Llama 4 Maverick 17B. They are no longer in the model list.
- Mistral will retire Mistral Large 3 and Mistral Medium 3.5 on 30 April 2027. They can be chosen until then, and on that date they leave the model list.
- ElevenLabs has replaced three of its models. Speech set to Turbo v2.5 is now made with Flash v2.5, and transcription with ElevenLabs uses Scribe v2 instead of Scribe v1, both at the same price. Eleven Music v1 is being retired too, so a song request that does not name a music model now uses Eleven Music v2, at the same price per minute and with the same maximum length.
When a price goes down
As a conversation goes on, the AI model reads the earlier messages again. Providers charge less when those messages are read from their cache.
- Claude Sonnet 5.5. Anthropic charges a twentieth of the input price for cached reads on Claude Sonnet 5.5. ToRun charged a tenth, so these reads now cost half as much as before.
- GPT-5.5 and GPT-5.4 in very long conversations. Past 272,000 tokens, cached reads were charged at the full input price. They are now charged at the cached price, a tenth of it.
- Mistral. Cached reads on Mistral's chat models, when Mistral reports them, are now charged at Mistral's cached price, a tenth of the input price, instead of the full input price.
When a promotional price ends
OpenAI has told us that the promotional price of GPT-5.6 Sol ends on 30 November 2026. From that day GPT-5.6 Sol is charged at its standard price: what it reads costs 25% more and what it writes costs 50% more. If you leave the choice of model to Auto, Auto takes the new price into account by itself.
The Translator is charged like the chat
The Translator charged less than it should have. The words the AI model wrote were priced at the lower rate for the words it reads, and a translation for which the AI provider reported no usage was not charged at all. Each translation, typed or spoken, is now priced the way a chat message is, from the model's own rates for what it reads and what it writes. Translations therefore cost more than before. When the provider reports no usage, the translation is charged on an estimate of its length.
A venue's logo and floor plan image
When you set a venue's logo or a floor plan's image in the venue panel, ToRun now accepts only a file from your own Library. A logo or image that is already set stays as it is, so a colleague can still edit the venue without choosing it again. The visitor page and the venue's screens show a logo or image only when the file belongs to the venue's own account.
A workflow's video step
When a workflow's video step puts a video together, it now uses a file stored on ToRun only if the file is in your own Library. This covers the pictures, clips, music, logo and closing picture the step puts in. If one of them links to a file in another account's Library, the step stops with a message that says which input it is, instead of using the file. To use such a file, upload it yourself. Links to files elsewhere on the web are used as before.