06-reference

every model shelf life anthropic s1

2026-10-28·reference·source: Every·by Grant Harvey
anthropics-1model-lifecycleagent-reliabilitycontext-windowevery

Every - Model Has a Six-month Shelf Life Now

Why this is in the vault

Every's own analysts turn one number buried in Anthropic's S-1 — models older than six months earn less than 20% of its revenue — into a live argument about model-switching behavior and the economics of frontier-model dependence, with a secondary read (Thorsten Ball's "you pay for the gap in retries") that lands directly on RDCO's verification-gate thesis.

The core argument

Anthropic filed to go public on Monday, and its S-1 quantifies something Every's staff had only felt anecdotally: when a new model ships, heavy users switch almost immediately. The filing's $16.3B first-half-2026 revenue predates Opus 5.5 and Fable 5.1 — "this is all Opus 5 money," per Every's head of evals Mike Taylor — and customers from a year ago now pay roughly nine times as much. Writer Evan Armstrong (whose full S-1 teardown is already filed separately, see Related) frames this as Anthropic behaving like a movie studio: only as good as its next release. The risk: Anthropic has committed to ~$155B/year in compute by 2028, mostly non-cancelable, while customer spend is not — if model quality gains ever stall ("the discernment horizon," per Steve Yegge, where most tasks aren't hard enough to show a difference between generations), that compute commitment flips from inventory to debt.

Every CEO Dan Shipper extends this into an "inchworm" model of AI adoption: engineers (leading edge) stretch forward into frontier capability first, knowledge workers follow, the rest of the economy catches up last. The piece's practical takeaway, credited to engineer Thorsten Ball: a model that's wrong 10% of the time doesn't tell you which 10% — so you pay for that gap in retries and rework. Every's own prescribed habit matches: read the Vibe Check on every release, and if relevant, switch your default and test it against one hard task you already know well (a "personal benchmark") rather than trusting benchmark marketing.

Curation section

Mapping against Ray Data Co

The sharpest connection is Thorsten Ball's line, relayed secondhand in this issue: a model that's wrong 10% of the time doesn't tell you which 10%, so you pay for the gap in retries. That's the exact premise behind RDCO's entire fresh-eyes-critic architecture (verify-vault-write, verify-strategic-output, verify-dispatch, station-critic) and the feedback_workflow_agent_output_integrity memory's seven factory failure modes — the model can't self-report its own error, so an independent gate has to catch it instead of trusting the producer's confidence. Second, the piece's "personal benchmark" habit (test each new release on one hard task you know well before switching your default) is already how the founder operates in practice: feedback_delegation_model_effort_pairing records the explicit 10-06 call to run everything on Opus 5.5 over Fable 5.1, and reference_claude_plan_max20x tracks the Max 20x plan itself as the vehicle for that kind of frequent re-evaluation. The "shoggoth strategy" debate (specialist front, generalist range behind it) is weaker — it's genuine career-positioning color relevant to project_credibility_for_phdata_sales, but the note shouldn't oversell it as a load-bearing RDCO insight; it's a useful frame, not a new decision.

Related