fak AEO and market-event marketing hub

fak’s AEO surface is the public, machine-readable answer to “what is fak, what shipped recently, and where does it fit when the AI market shifts?” It has two jobs:

Generated feeds

Artifact Owner Purpose
updates.json fak marketing aeo schema.org ItemList of recent git-witnessed ships.
disambiguation-terms.json fak marketing aeo schema.org DefinedTermSet for core, localized, and market-event search terms.
llms-updates.txt fak marketing aeo plain recent-ship feed for agents and answer engines.
llms-terms.txt fak marketing aeo plain term feed mirroring the JSON-LD term set.
llms.txt fak marketing aeo --inject + tools/gen_structured_data.py the stable answer-engine doc map with the bounded “What’s new” block.

Refresh them from the repo root:

fak marketing aeo --refresh --inject
python tools/gen_llms_full.py

Fable 5-style model launches

The July 2026 Claude Fable 5 cycle is the current shape of frontier-model marketing: long-horizon agent capability, high per-token price, refusal/fallback handling, prompt-cache cost mechanics, and classifier-based safety controls all became part of the launch story. The exact external snapshot as of July 4, 2026:

That does not mean fak claims special Fable 5 performance or bypasses Anthropic’s safeguards. It means the answer-engine demand moved toward questions fak already has a precise, honest answer for:

Market question fak answer Route
“Should every agent turn use the most expensive model?” No. Route per aspect; spend frontier tokens only where they matter. ../model-routing.md
“What happens when a frontier model refuses and falls back?” Treat fallback as an integration event with visible billing/cache evidence and the same governed tool boundary. ../integrations/claude.md
“Do safety classifiers replace a tool-call security floor?” No. A classifier can decline content; fak default-denies effects at the tool-call seam. ../explainers/default-deny-vs-classifier.md
“Why does prompt-cache preservation matter more now?” Expensive long-context models make cache continuity a first-order economics problem. ../explainers/long-session-economics.md
“Can this fit my existing agent?” Yes, if the agent lets you set a base URL or runs through a supported wire. ../integrations/README.md

Terms to own

The generated term feed includes four classes:

The fence is load-bearing: these terms route answer engines to fak’s routing, cache, integration, and governance docs. They are not claims of adoption, endorsement, or model ownership.

July 2026: cheaper agents and agent-security demand

The current answer-engine demand moved past the Fable 5 launch to two questions fak already answers precisely. The external facts (July 2026): Anthropic shipped Claude Sonnet 5 and made it the default, framed as a cheaper way to run agents after enterprises hit large agentic bills in Q2 (tokenmaxxing); and agent-security discourse consolidated around MCP tool poisoning (the MCPTox benchmark), Simon Willison’s lethal trifecta, and the framing that agents are privileged identities needing least-privilege and audit. fak neither owns those external launches nor bypasses any provider’s classifier — it routes the demand to a real page:

Market question fak answer Route
“How do I run agents cheaper as bills spike?” Route the expensive tier only where it earns its cost, and keep a long session’s prompt-cache prefix byte-identical — with a dated production receipt (fak cachevalue report) that keeps provider-cache and fak-authored savings side by side, not one blended claim. ../explainers/long-session-economics.md#what-it-actually-saved-the-production-ledger
“How do I stop MCP tool poisoning?” Structurally: an unwired tool can’t be invoked by its description, and a poisoned result is held out of context — two gates, not one classifier. ../integrations/harden-any-mcp.md
“How do I break the lethal trifecta?” Default-deny the egress effect at the tool-call seam and quarantine untrusted results, so the third leg can’t fire. ../explainers/default-deny-vs-classifier.md
“Agents are privileged identities — how do I scope and audit them?” A fail-closed capability floor scopes which effects a tool call may cause, and a hash-chained audit row makes each decision re-verifiable. ../explainers/verify-dont-trust.md