Two weeks after Opus 4.7's release, the regression reports keep coming. Multiple developer forums, enterprise users, and reviewers continue to flag that 4.7 hallucinates more frequently than 4.6 on factual recall, occasionally produces less reliable code in long sessions, and has introduced new failure modes in workflows that worked smoothly under the prior model. Anthropic has not formally acknowledged a regression, and the official benchmarks still show 4.7 ahead. The disconnect between benchmark and practice is itself the lesson — and a familiar one.
The deeper lesson is structural, not about Anthropic specifically. Most enterprises have built their AI roadmaps around a single primary model vendor. When that vendor ships a model regression — or a price change, a rate-limit cut, a policy shift, a security incident, an outage — the organization absorbs it. This is exactly the supply-chain risk that procurement teams have been managing in semiconductors, components, and SaaS for decades. AI is not yet treated the same way, and it should be. The organizations doing this well are routing critical workloads across at least two model families (typically Anthropic + OpenAI, or Anthropic + Gemini), with abstraction layers that let them swap providers without re-engineering downstream systems.
Worth noting in this context: Writer, whose enterprise app I covered last week, is one of the few enterprise AI vendors that runs its own proprietary LLM family rather than wrapping OpenAI or Anthropic. From a supply-chain standpoint, that matters — Writer's roadmap and reliability are not contingent on whether GPT-6 ships on schedule or whether Opus 4.8 fixes a regression. For governance and risk teams running an AI buyer's checklist, "is this vendor an OpenAI/Anthropic wrapper, or do they own their model?" is a question worth asking out loud. It is not a deal-breaker — many wrappers add genuine value above the model layer — but it changes how you think about diversification in your AI portfolio.