Anthropic made the million-token window generally available for both Opus 4.6 and Sonnet 4.6 (blog), extending what Opus 4.6 launched with in February to the cheaper tier. The newsletter pairs it with Sam Altman’s bet that today’s frontier models can discover the architecture that follows the transformer (post), which reads against the same issue’s WebAssembly interpreter hard-coded into transformer weights. Book-length native context is the alternative to the recursive-call approach that let models handle contexts 100x their window in January.
Development
Million-token context ships across a model family
Book-length context becomes a default across a model family rather than a flagship premium, recorded in the same breath as a lab chief betting those models will find their own successors' architecture.
Claim
Anthropic is shipping million-token context windows for Opus 4.6 and Sonnet 4.6, while Sam Altman bet that today's frontier models can discover the architecture that follows the transformer.
Reported metric: 1M tokens
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to March 16, 2026 (theinnermostloop.substack.com)
- 1M context is now generally available for Opus 4.6 and Sonnet 4.6 (claude.com)
- Sam Altman bets frontier models can discover the architecture after transformers (x.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
Related to
Reported in
Referenced by
- Test-time training gives continual learning constant latency
- A model takes over its own context via a Python REPL
- Recursive models handle contexts 100x their window
- A distilled model approaches the full one at twice the speed
- A 4B model dominates long context by rewriting its own weights in flight
- Welcome to March 16, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.