TL;DR
- The Financial Times reported that Anthropic’s most capable model is struggling to attract users while cheaper tools take the bulk of real-world work, a pattern visible for months.
- Opus costs roughly ten times what Sonnet costs per million tokens, and for drafting, summarising, classifying and answering questions, Sonnet is good enough, so usage clusters on the cheaper tiers.
- The same shape appears elsewhere, with GPT-4o Mini doing the bulk of OpenAI’s API work and Gemini Flash carrying most of Google’s API traffic while the flagship models get the demonstrations.
- The advice is to pick the cheapest model that reliably does the job and invest in the surrounding system; the FT piece is paywalled, and whether the agentic market can sustain frontier investment remains open.
In plain English
Model providers sell tiers of the same family at different prices. Anthropic’s runs from Haiku, which is fast, cheap and limited, through Sonnet as the everyday workhorse, up to Opus as the most capable and most expensive. Pricing is per million tokens, where a token is roughly a fragment of a word, so cost scales with how much text goes in and comes out. A tenfold price gap between tiers means the capability difference has to be worth ten times as much for the top tier to make sense on routine work.
The system around the model is where the durable investment sits: the instructions you write, the tools you connect, the criteria you evaluate output against, and the integration code. That layer carries over when you swap models, and models will be swapped as costs fall, new tiers arrive and open-weight alternatives close the gap. Anthropic’s counter-move is to push Opus towards agentic use, meaning complex multi-step tasks where the cheaper tier genuinely cannot finish the job.
The Financial Times reported this week that Anthropic’s most capable AI model is struggling to attract users, as cheaper tools capture the bulk of real-world usage. The full article is behind the FT paywall, but the headline points to a pattern that has been visible for months.
Anthropic’s lineup runs from Haiku (fast, cheap, limited) through Sonnet (the workhorse) to Opus (the most capable, the most expensive). In practice, most usage clusters around Sonnet and Haiku. The API pricing tells you why: Opus costs roughly ten times what Sonnet costs per million tokens. For the majority of tasks people are actually doing — drafting, summarising, classifying, answering questions — Sonnet is good enough. And “good enough at a tenth the price” wins almost every time outside a benchmark.
Nobody is arguing Opus isn’t more capable. The question is whether the gap in capability justifies the gap in cost for what most people need day to day. For most workloads, the answer is evidently no.
The pattern repeats across the industry. GPT-4o Mini does the bulk of OpenAI’s API work; GPT-4o is the one they show off. Google’s Gemini Flash carries most of the API traffic. The most capable model in every lineup is the one fewest people use day-to-day. The labs release the most capable model, and most customers pick the cheapest one that does the job.
The reason this matters beyond industry commentary: if you’re adding AI to a business process, the model you start with is rarely the model you’ll keep. Costs drop, new tiers appear, open-weight alternatives close the gap. The investment that holds its value is the system you build around the model. The instructions, the tools, the evaluation criteria, the integration code. That layer carries over when you swap the model.
Anthropic’s response to this pressure has been to push Opus into agentic use cases — complex, multi-step tasks where the capability difference justifies the cost. That’s a reasonable bet: there are tasks where Sonnet genuinely can’t do the job, and the people running those tasks care about quality more than token price. Whether that market is large enough to sustain the investment in frontier models is the question the FT article is trying to answer.
For anyone running a business with AI rather than studying the industry: choose the cheapest model that reliably does the job, build the system around it so you can swap later, and spend the saved budget on the work the model still can’t do. Most people will keep using the cheaper model because it does what they need.