Naderu Weekly
Weekly reports
What actually changed in foundation models, one week at a time — what shipped, what it costs, what it licences under, and what that means if you have to build on it. Every figure is linked to the page it came from and dated. Where we could only find a claim second-hand, we say so — and when we get something wrong, the correction opens the next report.
-
Episode 12 · latest
Why web search busts your prompt cache
Per-search grounding fees and dynamic snippets wipe out 95% token caching savings — why live web retrieval and enterprise RAG require separate architecture.
-
Episode 11
Stop paying for words you throw away
OpenAI and Perplexity both ship POST /v1/decisions, Claude Haiku 5.5 lands at a tenth of Haiku 4.5's price, and Sonnet 5.5's cache read halves to match GPT-6.1 Sol.
-
Episode 10
Same price, new rules
GPT-6.1 Sol and Claude Sonnet 5.5 keep their predecessors' prices and change what your code may send, Perplexity's Sonar now bills search per call, and Alibaba ships a Jev-class model.
-
Episode 9
Three launches, no price rises
GPT-6 Sol and Luna at half the GPT-5.6 price, Claude Opus 5.5 below Opus 5, Grok 4.7 at Grok 4.6's rate — and Jev, a model that answers in types.
-
Episode 8
The retirement that didn't happen
DeepSeek walked back V4-Pro. Sonar Chat Completions has ten days. And Gemini shipped a Live model we left off the board on purpose.
-
Episode 7
The name stayed. The model didn't.
DeepSeek retired a model and kept its name answering. GPT-6 Astra ties the top of the board and caches at four times the price. And three of this week's corrections are ours.
-
Episode 6
Nobody changed a price this week
And the board still moved. A licence that isn't MIT, a rate that became promotional without moving, and a date that isn't on the model you think it's on.
-
Episode 5
The price has a clock on it
GPT-5.6 Sol fell a third with no announcement. Four cheap rows carry expiry dates — and one correction is ours.
-
Episode 4
The weights aren't the whole model
Qwen3.8-Max ships open weights that leave out what the API can do. Meta's Muse Glimmer doesn't cut anything.
-
Episode 3
The clock is now part of the bill
DeepSeek's price changes on Sunday — and then again at one in the morning, every morning.
-
Episode 2
The same weights, twice
One copy refuses. One copy doesn't. Only one of them is generally available.
-
Episode 1
The open-weights inversion
The largest open model in history shipped this week. Meta went closed the same month.
Get it by email — one a week, one click to leave. The numbers behind every report live on the model board.