A faster mid-tier model from Anthropic, a conclusion at the Supreme Court about ChatGPT as support for a damages claim, and a change to the Gemini app that affects your work instructions. This is what you need to know.
Claude Sonnet 5.5: faster, same token price
On 28 September, Anthropic released Claude Sonnet 5.5. According to Anthropic, the model generates well over 30% faster than Sonnet 5, and because it uses tokens more efficiently, it costs up to 30% less for most tasks. The API price per token stays the same: $2 per million input tokens and $10 per million output tokens. That is half the price of Opus 5.5 ($4 and $20). The model is available via Anthropic's own API, AWS, Google Cloud and Microsoft Azure, including with zero data retention.
Those figures come from Anthropic itself, not from independent tests. If you use a tool that runs on Sonnet, test the new model on your own documents before switching. Nothing changes for your data: business accounts and the API do not train on your input, and EU hosting is available via the European regions of AWS or Google Cloud. Details are on our Claude page.
ChatGPT in the courtroom: allowed, but it must be verifiable
In a criminal case, the victim estimated the value of stolen jewellery at € 49,874. In doing so, ChatGPT converted old guilder prices and valuations to today's value, using the gold price and inflation of roughly 2.5% per year. The court adopted that amount. Advocate General Paridaens writes in his opinion of 22 September, published on 24 September, that "in principle no rule of law prevents" you from using an AI tool such as ChatGPT for such a justification. However, the calculation must be "sufficiently transparent and verifiable" for the other parties and the judge.
That was not the case here. The example calculation from ChatGPT cited no source at all, and nobody had confirmed the percentages used. The Advocate General advises annulling the ruling on this point. The Supreme Court has yet to decide.
Are you working as a lawyer or at an insurer? Treat AI output in a case file as your own calculation. Record every percentage and amount with a verifiable source, such as CBS inflation figures or a valuation report. That ChatGPT worked it out is not a source.
Gemini: Gems become skills from 17 November
Google will automatically convert Gems in the Gemini app into "skills" from 17 November 2026, reports 9to5Google. Gems are custom assistants with fixed instructions. You call skills with a "/" in the input field, and you can use several in a single conversation. Creating skills is currently only possible with Google AI Pro or AI Ultra. It is not yet clear what will happen for users with a free account and for shared Gems.
Have you built fixed workflows on Gems, such as a house style check or a standard email? Then save your instructions separately now, so you can check them after the transition. You can read more about the model itself on Gemini.
Background: open models now process most tokens
According to Vercel AI Gateway's production index, open-weight models handled 56% of all tokens in August. In April that was still 13%, writes Emerce. They accounted for only 14% of spending: most of the money still goes to closed models. Cheap says nothing about where your data ends up. Check /veiligheid/ for where each model runs.
What you can do now
- Test Sonnet 5.5 alongside your current model on a few real tasks before you switch.
- Always back up AI calculations in case files or claims with sources you can verify.
- Export the instructions for your Gems before 17 November.
- Compare models on price and data storage via /modellen/.