What matters in AI, and why.
The one that matters
Anthropic's new default on Claude Max claims performance close to Fable 5 at half the price - $5 in / $25 out per million tokens - its fourth model release in under two months.
Why it matters
The frontier is commoditizing itself: near-flagship capability at mid-tier prices resets what every team should expect to pay for top-model tokens.
Learn one thing
View full-size diagram ↗What it is
A base model is what pretraining produces: a network trained only to predict the next token across trillions of words. An instruction-tuned model is the same network after post-training teaches it to answer requests and behave like an assistant.
How it works
Pretraining yields a completion engine: ask a raw model a question and it may reply with more questions, because a quiz sheet is a plausible continuation. Post-training reshapes that: supervised fine-tuning on instruction-response pairs, then preference optimization (RLHF or DPO), plus a chat template marking user and assistant turns. Capability comes from the base run; post-training shapes it.
A concrete example
Prompt a base model with 'What is the capital of France?' and a typical continuation is 'What is the capital of Germany?' - it imitates a quiz sheet. The instruct version replies 'Paris.' Compare any open '-base' checkpoint with its '-instruct' sibling on Hugging Face to see it live.
Why it matters
Open releases ship both variants, and the choice changes results: fine-tune the base for a clean slate on narrow formats, fine-tune the instruct to keep chat ability. It also explains daily failure modes - an instruct model run without its chat template quietly degrades.
What people get wrong: That instruction tuning adds knowledge or intelligence. Nearly all capability comes from pretraining; post-training mostly changes behavior. Nor are base models useless - they are often the better start for custom fine-tunes and tasks that want diverse continuations.
Today's top 5
The price-performance frontier moved again - retest your routing assumptions.
The open-vs-closed lobbying fight went fully public, with APEC economies piling on.
An ad-subsidized lab attacking API margins pulls agent running costs down for everyone.
The largest open-weights release to date is days away, with the architecture already published.
A major training-data precedent from India, pointing the opposite way to US lawsuits.
A concise daily briefing on the AI developments worth understanding. Subscribe on the homepage.