Changelog
What changed, and when.
By Daniel Reeve · figures read at source 2026-09-10 · page updated 2026-09-18
AI pricing, free-tier limits and model specifications move monthly. This is the dated record of what moved, what we re-verified, and what we still could not confirm at the vendor’s own source.
Dated entries
Everything that changed.
Prices, limits and model specifications move constantly in this category. Rather than restamp a date across the site, we log what actually changed and when we checked it. The current figures live here →
-
Dataset restructured: 11 attributes per tool
Every tool moved into a single structured record covering price, free tier, context window, web search, file upload, image generation, citations, EU hosting and open weights. Pages are now generated from that record, so a corrected figure propagates everywhere at once instead of being retyped page by page.
-
Five vendors publish no per-model context window we could verify
Google, Mistral, DeepSeek, Perplexity and Meta do not publish a per-model context window in a form readable at their own source. Rather than repeat a figure from another comparison article, those cells now say Not verified. Only ChatGPT (1.05M), Claude (1M) and Grok (500K) publish a number we could check.
-
Full verification pass against vendor sources
Prices, free-tier terms, context windows and API rates re-read from each vendor's own documentation. This is the date stamped on every figure currently published on the site.
ChatGPTClaudeGeminiPerplexityMicrosoft CopilotGrokMistral Le ChatDeepSeekMeta AIMultipleChat
-
DeepSeek's current token pricing is not readable at source
We could not read a current per-token API rate on DeepSeek's own pages, so no figure is published for it. The free chat tier is confirmed and requires no card. Anyone budgeting around DeepSeek's API should take the rate from DeepSeek directly.
-
Grok's published training cutoff is 1 February 2026
Confirmed in xAI's own model documentation. With search switched off, anything after that date is invisible to the model — which matters most for API use, where search is not on by default.
-
Mistral's open weights split across three licence tiers
Only the Apache 2.0 models (Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, Mistral Nemo) permit commercial use. Ministral 8B, Mistral Large and Pixtral Large ship under the Mistral Research License and Codestral under a Non-Production License. Mistral's own documentation also notes the open weights are not safety-tuned.
How this works
Why a changelog and not a date stamp.
A comparison page that says “updated this month” without saying what changed is asking for trust it has not earned. Every figure on this site carries the date it was read at the vendor’s own source, and every movement in that data lands here with a date against it. When we cannot verify something, that is an entry too — the gaps are part of the record. The full method is here.