Two months after its June 2026 launch, Anthropic’s flagship Fable 5 model accounts for just 11.4% of companies’ Anthropic spending — and a mere 6% of tokens — according to Ramp data covering 70,000 U.S. businesses, as reported by the Financial Times this week. Claude Opus 5, released in late July at exactly half the price, has already overtaken the flagship in enterprise spending. For a model positioned as Anthropic’s most powerful offering ahead of an expected $2 trillion IPO, the data is awkward.
The Numbers Don’t Support the Flagship Narrative #
Ramp’s data tells a specific story: Fable 5 was priced at $10 per million input tokens and $50 per million output tokens — roughly double GPT-5.6 Sol and 70 times more expensive on input than DeepSeek V4-Flash ($0.14/M). Enterprises looked at that pricing and routed accordingly. Fable 5 is capturing premium pricing — 11.4% of dollars from just 6% of tokens — but it isn’t capturing volume. Opus 5, at $5/$25, overtook it within a month of its late July launch.
Anthropic’s overall numbers are healthy enough to prevent panic. U.S. enterprise adoption reached 43.5% of tracked businesses versus OpenAI’s 39.7%. Annualized revenue hit $65 billion in July, up from $47 billion in May. However, average prices paid for Anthropic models dropped approximately 25% since mid-July — which means growth is volume-driven, not premium-tier-driven. The company is doing well by selling a lot of Opus, Sonnet, and Haiku. Fable 5 is not the revenue engine the model tier implied it would be.
The Price Gap Is the Real Story #
The core issue is not that Fable 5 underperforms. It may well be excellent. The issue is that the price gap between Fable 5 and alternatives has grown large enough that the burden of proof has inverted. Developers no longer ask “why use a cheaper model?” — they ask “why use the expensive one for this specific task?” For most tasks, there is no compelling answer when the price differential is 70x.
Miles Clements, partner at Accel, put it plainly: “Most people don’t need to operate at the frontier. The era when customers preferred only frontier models was not a durable era.” Mantas Lukauskas, AI tech lead at Hostinger, called recent pricing shifts “the first real test of whether advanced models could maintain price premiums in competitive markets.” Fable 5 appears to be failing that test outside of specific high-complexity autonomous workloads — which, by Anthropic’s own positioning, is where it belongs. The question is whether those workloads are large enough to matter financially.
There’s also an under-reported headwind: Fable 5 requires 30-day data retention — a policy that analysts flagged as a genuine competitive disadvantage for privacy-conscious enterprises compared to models without such requirements.
Model Routing Is the Developer Takeaway #
The practical response to this pricing landscape isn’t to pick the cheapest model and call it a day — it’s to implement model routing. Route simple queries to DeepSeek V4-Flash or Claude Haiku. Route medium complexity to Sonnet or Opus. Reserve Fable 5 for tasks that genuinely need it: multi-day autonomous agent runs, complex cross-repository reasoning, and high-stakes decisions where a wrong answer carries real correction cost.
At scale, the math is compelling. Research from 2026 shows that 70–80% of tasks in typical AI workflows can be handled by smaller models. Routing versus frontier-only approaches reduces aggregate cost per turn by 80–95%. At 100K–500K daily active users, the routing-versus-flagship gap comes out to $200K–$400K per month in savings. Practical routing implementations range from manual per-request selection to managed services like Komilion or Martian — the right choice depends on the cost of an incorrect routing decision in your specific product.
Related:[NVIDIA NeMo Switchyard Cuts Agent Costs 74% With Open Model Routing]
Key Takeaways #
- Anthropic’s Fable 5 captured only 11.4% of enterprise Anthropic spending and 6% of tokens two months post-launch, per Ramp data on 70,000 U.S. companies
- Claude Opus 5 — launched at half the price in late July — overtook the flagship in enterprise spending within a month
- The input token price gap between Fable 5 ($10/M) and DeepSeek V4-Flash ($0.14/M) is 70x; at these multiples, most tasks can’t justify the premium
- Model routing is the architectural response: send 70–80% of tasks to smaller, cheaper models and reserve the frontier tier for genuinely complex workloads
- Average prices paid across Anthropic’s model catalog dropped ~25% since mid-July 2026, meaning volume growth is outrunning per-unit value — a dynamic investors will scrutinize ahead of the IPO