Why Anthropic's Most Expensive Model Is Its Least Popular

In Anthropic's model menu, Fable 5 is the most expensive option: $10 per million input tokens, $50 per million output — twice the price of Opus 5. Yet Ramp's July spending data shows Fable 5 taking just 8% of total Anthropic model spend, far behind Opus 4.8 — a model already a generation old — at 28%.

On August 23, the Financial Times published "Anthropic's best AI model struggles to attract users as cheaper tools thrive", which quickly hit the Hacker News front page with 482 points. The piece describes a paradox: Anthropic's revenue is still climbing — roughly $65 billion annualized in July, per sources, up from $47 billion in May — yet its most powerful, most expensive model is its weakest segment.

The numbers: a frontier model at 8%

The Ramp AI index aggregates billing data from about 70,000 companies and shows how Anthropic model spend split in July:

Model	Share of spend (July 2026)
Opus 4.8	28.0%
Sonnet 4.6	8.3%
Fable 5	8.0%
Opus 4.6	6.9%
Sonnet 5	3.6%
Opus 5	3.5%

Two things jump out. First, the frontier model Fable 5 — what Anthropic positions as its peak — only ties Sonnet 4.6 in real-world spend. Second, Opus 4.8, the previous generation, still holds nearly a third. Note that Opus 5 only launched on July 24, so its 3.5% covers just over a week rather than the full month. That makes it more striking that it still overtook Fable 5 in business spending in that short window.

Why the most expensive model loses

It isn't about quality. Fable 5 remains Anthropic's strongest model on several benchmarks, and Anthropic's own system card admits Opus 5 is "not more capable than Fable 5 overall". The problem is arithmetic.

A model that's 3% better on a benchmark but twice as expensive in production has to prove that 3% pays for itself in real work. For most tasks — autocomplete, summarization, support replies, test generation — it doesn't. Opus 5 landed exactly on that point: "close to the frontier intelligence of Fable 5 at half the price," in Anthropic's own framing, and "designed to be used every day."

Pressure also comes from outside. OpenAI's GPT-5.6 Luna now sits at $0.04/$0.24 per million tokens after a late-July price cut, while DeepSeek and Kimi K3 sell near-equivalent models a tier cheaper. When the price floor keeps dropping, a $10/$50 tier gets harder to justify for anyone not running genuinely hard tasks.

Anthropic manages its own scarcity

Anthropic's response shows it understands the problem. Since July 20, within subscription plans, Fable 5 is capped at 50% of weekly usage for Max and premium Team plans, while Pro and standard Team users access it through separate credits. Fable 5 gets no dedicated allowance — it draws from the same weekly pool as other models, so using Fable burns the allowance faster.

Meanwhile, Opus 5 ships with a cost dial: an effort parameter with five levels from low to max, letting customers control how long the model thinks before answering. It's Anthropic selling bill control inside the product itself — an admission that for many teams, predictable cost matters more than a few benchmark points.

What to Know

This isn't an "Anthropic is losing" story. Revenue is still rising, and the company expects Q3 to be profitable under the same model it used to declare Q2 profitable. What's changing is the definition of "best": from "highest benchmark score" to "best value per task".

• Model selection is a portfolio problem, not a single-model problem. Default to the cheapest model that clears your quality bar for the task; reserve the frontier model for the 10-20% of genuinely hard tasks where the margin pays for itself.
• Benchmark margin isn't ROI. A model leading SWE-bench 97% vs 95% doesn't justify paying double across your whole pipeline. Measure on your own data.
• Watch spend, not just quality. Anthropic adding an effort dial and rationing Fable is a signal: providers are passing the marginal cost of the frontier tier back to you. Teams that don't measure price-per-task will overpay.
• The floor keeps dropping. OpenAI cut Luna's price 80% three weeks after launch; Opus 5 already costs a third of what Opus 4.1 did. Locking into an expensive model makes your costs relatively more expensive every quarter.

---

The Fable 5 story isn't a product failure. It's a sign that the premium tier of AI is being squeezed from both sides: cheaper models below are good enough for most work, and the provider itself has to throttle consumption of its priciest model. For developers, the takeaway is simpler than it looks: the best model is no longer the most powerful one — it's the one that's worth the money for the exact job you're doing.

---

Content assisted by AI (Amy 🌸). Reviewed by the author.