Enterprises pick cheaper Claude models over Anthropic's flagship Fable 5

A flagship with a small share of its own platform

Anthropic's strongest model is not where its business customers spend. According to figures published by the Financial Times and drawn from Ramp's AI Index — a dataset built from corporate card spending across 70,000 businesses — the flagship Fable 5 model captures roughly 11% of enterprise spending on Anthropic's own tools. That share has been flat since the model's June launch, while the older, cheaper Opus 4.8 jumped from about 15% of spend in early June to more than half by late July. Opus 5, released on July 24 at half Fable 5's price, overtook the flagship within weeks and now sits in the mid-teens.

The technical case for Fable 5 is strong, as a widely shared dev.to analysis notes. It leads the SWE-Bench Pro benchmark at 80.3%, eleven points ahead of the nearest rival, and was at one point withdrawn from public availability under US export controls. Prediction-market traders on Polymarket put a 72% probability on Anthropic holding the best model at the end of 2026, against 6% for OpenAI and 7% for Google, on roughly $784,000 of trading volume.

Price beats peak for most workloads

The spending mix points to a simple purchasing rule: companies appear to select the least expensive model that meets their quality threshold. Routine enterprise tasks — summarising, classifying, reviewing code, extracting data from documents — run acceptably on a mid-tier model priced at $5 per million input tokens and $25 per million output tokens. Fable 5 costs exactly double, $10 and $50, so it must return twice the value per task to pay for itself; the dev.to analysis estimates the uplift for typical workflows at closer to 1.1x.

Simon Willison, surfacing the Financial Times piece, argued the Ramp numbers are a particularly trustworthy signal because they record what companies actually paid rather than what they report in surveys.

Pressure from above, below and inside

OpenAI cut prices for GPT-5.6 Sol by more than 20% on August 21, lowering input tokens from $5 to $4 and output from $30 to $20 per million. The reduction is promotional and runs through November, but it leaves Sol cheaper than Claude Opus 5 on both input and output.

At the bottom of the market, DeepSeek's V4-Flash runs at $0.14 and $0.28 per million tokens — which the dev.to post calculates as 93% cheaper on input and 97% cheaper on output than Claude Sonnet 5. The same analysis estimates the cost of running a GPT-4-class model has fallen roughly a thousandfold in three years as Chinese open-weight models commoditised that tier.

Anthropic also faces pressure from its own catalogue. CNBC reported that Opus 5 beats Fable 5 on coding and knowledge-work evaluations while costing 50% less. Across the market, the gap between Fable 5's output price and DeepSeek V4-Flash's output price now spans a factor of 71.

Trust complaints compound the pricing story

A Hacker News discussion reached 770 points and 679 comments, making it one of the busiest AI threads of the month. Commenters criticised unpredictable usage quotas and frequent policy shifts, with one startup founder reporting that a Claude Team account was banned without warning after a remote employee logged in from another country, prompting a weeks-long move to OpenAI. Others described switching to locally run Qwen 3.8, DeepSeek V4 through OpenRouter, or Chinese models such as Kimi K3 and GLM-5.3 for particular workloads. On X, Austen Allred — whose post drew 338,000 views — suggested that AI adoption inside most companies is far slower than social media timelines imply, which shrinks the addressable market for a premium frontier model.

The bull case

Timing may be the issue rather than price. Premium tiers historically lag in enterprise adoption — GPT-4 needed about six months to pass GPT-3.5 in API spending — and Jay Hack has argued that organisational readiness, not model intelligence, is the bottleneck for valuable use cases. Anthropic's overall business is also accelerating: the dev.to analysis cites annualised revenue of $65 billion in July, up from $47 billion in May, 6,000 customers spending at least $100,000 a year, and an IPO filing targeting a valuation above $2 trillion.

Why it matters

The lesson for the wider AI market is that benchmark leadership does not automatically convert into revenue share — not even within a vendor's own customer base. If the best model available earns 11% of spend on its own platform, frontier capability alone is a weak moat. Buyers keep raising their definition of good enough while cheaper options improve from below and rivals cut prices from above. Vendors competing on intelligence alone may find that distribution, reliability, predictable pricing and trust decide who actually captures the volume.