9 July 2026
GPT-5.6 Sol: a new model, but no measure of its Welsh
OpenAI’s latest model family offers three options and an enormous context window. But any measure of their Welsh ability remains missing.
Today OpenAI released the GPT-5.6 family to the public, following a limited preview that began on 26 June. Sol is the most capable model, Terra is the middle option, and Luna is the fast, inexpensive option. OpenAI has not disclosed model sizes or architectures. External sources mention a 1.5-million-token context window and lower token use on coding tasks.
The data still missing
As with other large models released this year, we examined the system card for evidence about Welsh. We found no Welsh result. The GPT-5 system card’s multilingual table included around 13 or 14 languages, from Arabic to Yoruba; Welsh was not among them, and no equivalent table has been published for GPT-5.6.
OpenAI is not the only example. Here is the available evidence across the major 2026 models:
- Claude Opus 4.8 (Anthropic, 28 May): Anthropic’s documentation includes a multilingual table for 15 languages, but not Welsh. Independent reviews also identify multilingual tasks as a relative weakness.
- Claude Fable 5 (Anthropic, 9 June): a ‘Mythos’-class model. The multilingual result cited in the announcement uses MGSM, a ten-language benchmark that does not include Welsh.
- Grok 4.20 (xAI, April): the system card names English, Chinese, Arabic, Hindi and Spanish ‘plus several dozen others with varying capability’ — without naming Welsh or giving any benchmark for it.
The pattern in the material reviewed: we found no Welsh results from the major AI labs. Quantitative measurements are available from academic work in Britain — BritEval at UCL, Techiaith’s suites at Bangor and Cardiff’s CEFR work — but few published scores compare multiple models.
What next?
This is precisely the gap AIsteddfod was established to record. Each model above has a profile in our evaluations section. We identify the available evidence alongside what has not yet been measured. Those pages will be updated as new results arrive.
Sources: OpenAI announcement and GitHub Copilot changelog (9 July 2026); GPT-5 system card (arXiv 2601.03267); Anthropic multilingual support documentation; Grok 4.20 system card (7 April 2026).