New Models, and Some Honest Retirements

We went through every provider's own documentation in a single sitting and reconciled the catalogue against it, line by line. Not an aggregator's list: aggregators put their own margin in the middle and have their own reasons for what they show. The pages we read were the providers'.

Joining

From OpenAI: gpt-6-astra, a reasoning model; gpt-5.3-codex for code; gpt-realtime-2.1-mini, a smaller realtime voice model; and gpt-transcribe for speech-to-text. From Anthropic: claude-fable-5-1. From Google: gemini-3.8-flash, gemini-3.5-flash-lite and lyria-3.5 for music. From xAI: grok-4.6. DeepSeek's V4.1 Flash tier joined the same evening under its own name.

Every one of them that has a runtime path was called for real with a development key before its row was written. A catalogue entry that has never been exercised is a guess with a price on it.

Retiring, and repricing

Models whose makers have published an end date now carry that date, taken from the deprecation page rather than from memory: several Claude generations, the older Mistral line, two image models, an OpenAI voice model, a family of search models. A model that has genuinely shut down is marked as shut down rather than left sitting there with an active price.

Prices are never edited in place. When a provider changes a rate, the old row is given an end date and a new row starts — so a bill from last month still reads at the rate it was actually charged, and always will. That is why this pass produced far more rows than it produced models.

The shutdown date that did not exist

The find that mattered. Four Gemini 2.5 models in our catalogue carried a retirement date in October. Google's deprecation page lists no shutdown date for them at all. Left alone, three models people use today would have switched themselves off on a date we had invented, and the failure would have looked like the provider's.

It has been removed. Its siblings — dates that are published — stayed, and two that were wrong in the other direction were corrected.

Capability-first means a wrong row breaks everything

ToRun does not pick a model by name. It works out what a turn actually needs — can it see, can it call tools, does it need a long context, is it going to speak — keeps only the models that can do all of it, and ranks what is left. Which means a capability recorded wrongly does not degrade one feature. It quietly misroutes everything that needs that capability.

So this pass also corrected the joins. One example: our top Anthropic tier had no intelligence score at all, which meant it scored as unknown and ranked below the cheapest model in its own family. Another: a music model whose identifier fell through a catch-all and silently ran a different model than the one named on the row.

Choosing a model by hand is still honoured exactly as before, which is also why it can still be refused for a job it cannot do.