06-reference

innermost loop fable5 share loss thread lines autonomous weapons

2026-08-25·reference·source: The Innermost Loop·by Alex Wissner-Gross (curator)
anthropic-model-mixcompute-politicsagent-autonomymemory-cyclefrontier-digest

"Welcome to August 25, 2026" — Alex Wissner-Gross

Why this is in the vault

Daily curated digest threading internal Anthropic model-mix economics (Fable vs. Opus spend share), compute-politics backlash ("thread lines" rationing intelligence), and autonomous-weapons/agent-autonomy escalation into one issue; filed as a fresh dated anchor on the model-mix and compute-politics threads this source tracks daily.

Issue contents

Mapping against Ray Data Co

The concrete connection is the Ramp data point that Fable 5 is stuck at 11% of Anthropic's internal spend share and losing ground to Opus 5 on price and retention terms — a live signal against [[feedback_delegation_model_effort_pairing]]'s standing rule that Fable delegations get paired with high/xhigh effort and reserved for meaningfully large tasks. If Anthropic's own usage data shows Opus displacing Fable at the margin, the SOP's implicit bet that Fable earns its premium on big-task delegation is worth re-checking against real cost/quality tradeoffs the next time delegation defaults come up, rather than treating the pairing rule as fixed. Secondarily, the "thread lines" compute-rationing frame (public backlash pushing toward self-inflicted compute queues) is a fresh macro data point for the RDCO capacity assumptions underneath [[project_l5_north_star_strategic_direction]] — L4→L5 progress is downstream of agent capability, and agent capability is downstream of available compute; state-level demand-side backlash (70-30 against local data centers) is a headwind on that supply line RDCO doesn't currently track anywhere.

Related