上线超过六个月的模型贡献收入不足 20%
Key Highlights
Every's Context Window column reads Anthropic's S-1 and lands a blunt verdict: every frontier model's shelf life is about six months. The argument is that buyers and vendors alike must plan around constant turnover, because the model that pays the bills today will be obsolete before the fiscal year ends.
What Happened / The Evidence
The piece cites Evan Armstrong's reading: Anthropic depends heavily on its newest models — those older than six months drive under 20% of revenue. The $16.3 billion first-half 2026 revenue came before the two best-reviewed models, Opus 5.5 and Fable 5.1, shipped. In plain terms, the company is earning almost entirely on its latest release, with the back catalog contributing little. That is a strikingly concentrated revenue curve for a major AI lab.
Technical Details
The concentration is not accidental; it follows from how fast benchmarks and user expectations move. A model that led the pack in spring can feel dated by winter, and enterprises renegotiate or switch as newer options arrive. The column frames this as structural: the "best model" is a moving target, and the company riding it must ship continuously just to hold revenue flat. The six-month figure is a rule of thumb, not a law, but it captures the tempo.
vs Competitors
The same dynamic applies to OpenAI, Google, and others, though their revenue mixes differ. The differentiator is release cadence and the ability to migrate users forward without churn. A lab that ships a hit every six months stays solvent; one that stalls watches its share evaporate. The shelf-life lens also reframes "model moats" — they are temporary, and the real moat is the delivery engine, not any single checkpoint.
Industry Impact / Use Cases
For buyers, the lesson is to avoid locking architecture to one model version; build swappable inference so you can chase the frontier without rewrites. For investors and operators, the six-month clock is a risk metric: compute commitments made today must pay back before the model that justified them ages out. Every's framing is a useful reality check against the temptation to treat any single model as a durable asset.
Practical Notes
The six-month clock is a planning constraint, not a prophecy. For buyers, the implication is architectural: build inference that can swap models without rewriting the application, so chasing the frontier does not mean rebuilding. Abstract the model call, pin versions deliberately, and keep an eval suite that tells you when a newer model is actually better for your tasks. The labs that ship continuously will keep pulling revenue forward; your job is to capture the gains without being held hostage to any single release.
For operators and investors, the concentration of revenue in the newest model is a risk metric. Compute commitments made to justify a launch must pay back before that model ages out, or the economics break. The bottom line: "model moats" are temporary, and the durable asset is the delivery engine — cadence, migration tooling, and user lock-in through workflow. Every's framing is a useful corrective to treating any checkpoint as a permanent advantage; plan for turnover, and the turnover stops being a surprise.
What to Watch
Watch the release cadence of the major labs; if six months becomes the norm for a model to go from hero to has-been, roadmaps and hiring plans built on permanent advantages will keep breaking. Also watch enterprise contracts for model-portability clauses — buyers are starting to demand the right to swap, which pressures labs to compete on delivery, not lock-in.
Takeaway
The six-month shelf life is a planning constraint, not a prophecy. Build swappable inference, keep an eval suite, and treat model moats as temporary. The durable asset is the delivery engine — cadence, migration tooling, and workflow lock-in — not any single checkpoint.
Context
The column also implies a capital-discipline lesson for the labs themselves: spending billions on compute to train a model whose revenue half-life is six months demands a delivery cadence that keeps paying the bill. The firms that master continuous release will compound; those that ship episodically will watch their best work age out before it earns.
The firms that internalize turnover as a design constraint will stay agile; those that bet the roadmap on a single immortal model will keep rewriting it when reality arrives ahead of schedule.