Changelog
Model lineups and vendor guidance move fast. This is what changed here, and when.
1.2.2
2026-09-01- Qwen licence corrected. The two licence-encumbered Qwen models were described as needing a separate licence from Alibaba “above published revenue thresholds”. That is wrong on both counts. The licensor named in both texts is Qwen. The Qwen Community License 1.0 on Flash-Next puts no revenue threshold on the separate-licence requirement at all — it applies to any model-as-a-service or AI-work-assistant business, internal use excepted — while the Qwen3.8-Max License on the 2.4T applies it above US$50,000,000 of aggregate revenue over twelve months. The 100,000,000-user and US$20,000,000-monthly-revenue thresholds belong to the attribution clause, not the licensing one. Both licences are now quoted rather than summarised.
- xAI: Grok 4.6’s maximum output is now shown as unpublished, like the rest of the line. Neither xAI’s model list nor the Grok 4.6 model page states an output limit, and neither states that there is none — the site previously claimed the flagship had no text output limit.
- Mistral: the rule about where durable instructions go now follows the prompt-engineering guide, which describes the system prompt as set at the start of the conversation and managed by the developer, with concatenation into the user prompt as the fallback when you cannot control it. The guide does not say the two roles are interchangeable.
- Three specification fixes on the open-models page, each against the vendor model card: Gemma 4 E2B and E4B have a 128K context window, not 256K (the 256K figure applies to the 12B, 26B-A4B and 31B); Qwen3.8-27B states 27B parameters, not 28B; Mistral Small 4 activates 6.5B parameters per token.
1.2.1
2026-09-01- Anthropic’s flagship updated: Claude Fable 5.1 released on 1 September 2026 and replaces Claude Fable 5 at the top of the lineup. The context window (1M) and output limit (128K) are unchanged, and Claude Fable 5 is still listed as available, as a legacy model.
1.2
2026-09-01- Five more frameworks: PICSE from the UK Government, TELeR and ReAct from peer-reviewed research, plus CREATE and PREPARED. Two of them are not fill-in forms — TELeR is a rubric for grading how specified a prompt is, and ReAct is a loop — so they have pages rather than builder entries.
- A new provenance tier: frameworks published by a government body, which PICSE is the first of.
- Three more providers: xAI, Mistral and DeepSeek, each with its own builder adapter reflecting what that vendor documents — including two that document almost nothing, which the pages say plainly.
- Two new sections: open-weight models you can self-host, and a self-hosting guide covering runtimes, hardware, quantisation, how prompting differs locally, and the GDPR and AI Act position.
- Glossary expanded from 15 to 34 terms.
- Meta’s page reframed: Llama 4 has not been updated since May 2025, so it is presented as a known quantity rather than a current recommendation.
- Legal page updated for the July 2026 amendment to the AI Act, which moved several deadlines, and for the Article 4 AI-literacy duty.
1.0
2026-09-01- First release: 10 frameworks, 5 providers with 14 models, the prompt builder with five model adapters, the prompt improver, and the framework and model pickers.
- All framework and model pages verified against vendor documentation on 1 September 2026.
- Full English and Dutch interface, with prompt output language selectable independently.