The Financial Times reported this month that Anthropic’s US customers are drifting away from Opus 4.8, the company’s most powerful model, in favor of cheaper alternatives — including, awkwardly, Anthropic’s own mid-tier Sonnet 4.6. The numbers are stark: Opus 4.8 costs $5 per million input tokens and $25 per million output tokens, while Sonnet 4.6 runs $3 and $15. In head-to-head comparisons, users picked Sonnet over Opus 59% of the time.
The read from the peanut gallery is predictable: the AI bubble is deflating, nobody wants to pay for quality, the race to the bottom is on. But that framing assumes Opus 4.8 was supposed to be a product people buy. It wasn’t.
The High-Effort Default Is a Tell
Look at how Anthropic actually shipped Opus 4.8. The model defaults to “high effort” — meaning it spends more tokens per task than its predecessor. Anthropic’s own release notes warn that Opus 4.8 “may use more tokens than Opus 4.7 on some tasks.” So the cost gap between Opus and Sonnet isn’t just the sticker price. It’s compounded by the fact that the flagship burns more tokens to do the same job.
That’s not an accident. It’s a design decision. If you wanted Opus to be a volume product, you’d default it to the cheapest setting that still wins benchmarks. Instead, Anthropic shipped a model that is, by default, the most expensive way to do anything. The only reason to do that is if volume was never the point.
And here’s the tell: Anthropic is already discounting the flagship. The company touted that Opus 4.8 is 61% cheaper per token than Opus 4.7. That’s not the behavior of a company defending a premium product’s margins. That’s the behavior of a company managing a halo. You don’t cut the price of a product you expect people to buy at volume; you cut the price of a product whose job is to stay in the conversation.
The Benchmark Economy
Opus 4.8 posts an 88.6% on SWE-bench Verified and 89.9% on GPQA Diamond. Those numbers are real, and they’re impressive. But ask yourself: who actually needs a model that scores 88.6% on SWE-bench? The answer is almost nobody — and that’s fine, because the number isn’t for users. It’s for investors, for the press, for the enterprise procurement officer who needs to justify a vendor choice to a CFO.
The flagship model is a spec sheet with a token price attached. It exists to win benchmarks, generate headlines, and anchor the brand. The actual revenue comes from Sonnet, which is cheaper, faster, and good enough for the office work that makes up the vast majority of AI usage. When Anthropic’s own announcement for Opus 4.8 leads with a Databricks executive praising the model’s “step change in agentic reasoning,” that’s not a product launch. That’s a press release for the brand.
One trader I spoke to on a trading desk put it bluntly: “Opus is the concept car. Sonnet is the sedan. Nobody drives the concept car to work.”
Stratification, Not Commoditization
The “race to the bottom” crowd has this backwards. What’s happening isn’t commoditization — it’s stratification. The frontier model and the workhorse model are different products with different economics, and the frontier model is increasingly a prestige good. The $200-a-month Claude Pro subscriber isn’t paying for Opus’s benchmark scores; they’re paying for the feeling of having access to the best. The enterprise buyer running Sonnet at scale isn’t choosing the cheaper model because they’re cheap; they’re choosing it because it does the job.
If this read is right, a few groups have to revise their priors. Investors who value Anthropic on the strength of Opus are pricing the wrong asset — the flagship’s benchmark scores tell you almost nothing about the revenue line. Enterprise buyers who assume “best model” means “best product” are confusing a marketing artifact with a tool. And the FT’s framing — “struggles to attract users” — assumes the goal was to attract users. What if the goal was to attract attention?
The real risk for Anthropic isn’t that users reject Opus. It’s that the gap between the spec sheet and the product people actually use grows so wide that the spec sheet stops mattering. When the concept car stops generating buzz, the sedan has to sell itself. And sedans, as any automaker will tell you, are a much harder business.