Anthropic's Best Model Can't Hold Users

Anthropic builds the best AI model in the world. Prediction markets agree — Polymarket gives the company 72% odds of holding the crown through year-end 2026. Fable 5 tops SWE-Bench Pro at 80.3%, a full 11 points ahead of its nearest competitor. It was so capable that the US government briefly pulled it from public access under export control rules.

📖 Read the full version with charts and embedded sources on ComputeLeap →

And yet, according to data published by the Financial Times this week, Fable 5 accounts for just 11% of enterprise spending on Anthropic's tools. Not 11% of the broader AI market — 11% of spend on Anthropic's own platform. The older, cheaper Opus 4.8 surged from 15% in early June to over 50% by late July. Anthropic's newer Opus 5, released at half Fable's price, overtook it within weeks.

View original post on X →

The thesis here is not that Anthropic is failing. The company's annualized revenue hit $65 billion in July, up from $47 billion in May. It has 6,000 customers spending $100,000 or more annually. The IPO filing targets a valuation north of $2 trillion.

The thesis is sharper than that: building the best model is necessary but insufficient for building the best business. And the gap between "best model" and "best product" is where margins go to die.

The Ramp Data Tells the Story

The numbers come from Ramp, the fintech company that processes corporate expenses for 70,000 businesses. Their AI Index offers one of the few windows into what enterprises actually spend on AI — not what they say in surveys, not what analysts project, but where the money moves.

Here is what it shows for Anthropic's model mix:

• Opus 4.8: ~50% of spend (surged from 15% in June)
• Opus 5: Mid-teens percent (overtook Fable within weeks of July 24 launch)
• Fable 5: ~11% (plateaued since June launch)

View on Techmeme →

The pattern is unmistakable. Enterprises are not choosing the smartest model. They are choosing the cheapest model that clears their quality bar. And for the vast majority of corporate AI workloads — summarization, classification, code review, document extraction — Opus 4.8 at $5/$25 per million tokens clears that bar just fine.

ℹ️ Fable 5 costs $10/$50 per million tokens — exactly double Opus 4.8 and Opus 5. At that premium, it needs to deliver 2x the value per task. For most enterprise workflows, it delivers maybe 1.1x.

As Simon Willison noted when surfacing the FT story, the Ramp data is an unusually reliable signal because it tracks actual corporate card spend, not self-reported usage.

The Price War Just Got Real

Anthropic's pricing problem is not happening in a vacuum. The competitive pressure is coming from three directions simultaneously, and it is accelerating faster than most observers expected.

The AI pricing landscape has undergone a structural shift in 2026. What was once a two-player market (OpenAI vs. Anthropic) is now a multi-front war where open-weight Chinese models, aggressive promotional pricing, and Anthropic's own model lineup are all compressing the premium that frontier intelligence can command.

From above: OpenAI is cutting prices. On August 21, OpenAI slashed GPT-5.6 Sol prices by over 20%, dropping input tokens from $5 to $4 and output from $30 to $20 per million. Sol now undercuts Claude Opus 5 on both input and output — and costs a fraction of Fable 5. The cut is promotional (through November), but it signals OpenAI's willingness to sacrifice margin for market share.

From below: DeepSeek is rewriting the cost floor. DeepSeek V4-Flash runs at $0.14/$0.28 per million tokens — 93% cheaper on input and 97% cheaper on output than Claude Sonnet 5. The broader trend is staggering: the price to run a GPT-4-class model has fallen roughly 1,000x in three years. Chinese open-weight models have turned what was once a frontier capability into a commodity.

From within: Anthropic is cannibalizing itself. When Anthropic launched Opus 5 on July 24 at half Fable's price, it immediately started eating Fable's already-thin share. This is not a mistake — it is Anthropic acknowledging the market reality. CNBC reported that Opus 5 outperforms Fable 5 on coding and knowledge work evaluations while costing 50% less.

The pricing map tells the full story. Here is what frontier inference costs per million tokens across the major providers as of this week:

Model	Input	Output	Provider
Claude Fable 5	$10.00	$50.00	Anthropic
Claude Opus 5	$5.00	$25.00	Anthropic
Claude Opus 4.8	$5.00	$25.00	Anthropic
GPT-5.6 Sol (promo)	$4.00	$20.00	OpenAI
DeepSeek V4-Flash	$0.14	$0.28	DeepSeek

That is a 71x spread between Fable 5 output and DeepSeek V4-Flash output. Even comparing frontier-to-frontier, GPT-5.6 Sol now undercuts Opus 5 by 20% on both dimensions. The walls are closing in from every side.

What the Community Is Saying

The Hacker News thread on this story hit 770 points with 679 comments — one of the most active AI discussions this month. The sentiment is brutal.

View on Hacker News →

The primary critique centers on Anthropic treating monetization like model training — constantly experimenting. Users describe unpredictable quota systems and frequent policy changes that erode trust. One startup founder reported having their Claude Team account banned without warning when a remote team member accessed it from overseas, forcing a weeks-long migration to OpenAI.

Multiple commenters described switching to alternatives:

• Qwen 3.8 running locally
• DeepSeek V4 via OpenRouter at a fraction of the cost
• Chinese models (Kimi K3, GLM-5.3) for specific workloads

The thread reads less like a pricing discussion and more like a trust discussion. Users are not just complaining about cost — they are complaining about an adversarial relationship with their AI provider.

On X, the reactions were equally pointed. Austen Allred, with 338,000 views on his post, put it bluntly:

View original post on X →

And in a separate post that drew 2,630 likes, he noted that AI adoption across most companies is far slower than the X timeline suggests — which means the addressable market for a $10/$50 frontier model is even smaller than Anthropic's bulls expect.

View original post on X →

The Polymarket Paradox

Here is the sharpest irony in the entire AI industry right now.

Polymarket traders give Anthropic 72% odds of having the best AI model at year-end 2026. On $784,000 in trading volume, the market's conviction is high. OpenAI sits at 6%. Google at 7%.

But the Ramp data shows that having the best model translates to 11% of your own enterprise spend.

This is the paradox at the heart of the frontier model business: the market for "the best" is tiny. Most AI workloads do not need the best model. They need a model that is good enough, fast enough, and cheap enough. And the definition of "good enough" keeps rising as mid-tier models improve.

It is the same dynamic that played out in databases (most workloads run on Postgres, not Oracle), in cloud computing (spot instances outsell reserved capacity), and in every mature technology market. The premium tier exists, and it matters, but it never captures the volume.

⚠️ Contrarian Corner: The bull case for Fable 5 is timing, not pricing. Enterprise adoption curves for premium tooling always lag — GPT-4 took six months to overtake GPT-3.5 in API spend. Jay Hack argues on X that the bottleneck for most valuable use cases today is not model intelligence but organizational readiness. Fable 5 may simply be ahead of its market. And Anthropic's $65B annualized revenue proves the overall business is accelerating — even if the flagship is not yet the revenue driver.

The Real Business Is Not the Model

This is where the story gets interesting — and where Anthropic may actually be ahead of the narrative.

Look at what Anthropic has done in the last 90 days. Not the model launches (everyone does those). Look at the business model shifts. The company appears to understand something that neither its critics nor its cheerleaders have fully articulated: the per-token API business is a transitional phase, not the endgame.

1. Claude Code at $60/user/month — bundling the frontier model into a developer tool with a fixed subscription price. This is the Adobe Creative Cloud playbook: stop selling the engine and start selling the vehicle.

2. Opus 5 as the volume play — deliberately launching a cheaper model that competes with Fable. This is Anthropic telling the market: "We know you won't pay $50/MTok for output. Here's an almost-as-good option at $25."

3. Enterprise plans with Fable included — since July 20, Max and Team Premium plans include Fable 5 at 50% of limits. This turns Fable from a standalone product into a bundle feature — a familiar SaaS move.

The pattern suggests Anthropic is pivoting from "pay per token for the best model" to "pay a subscription for the best platform." That is a fundamentally different business, and it is probably the right one.

The precedent is instructive. Microsoft did not win enterprise software by selling the best spreadsheet — it won by bundling Excel, Word, Outlook, and Teams into a platform subscription. Salesforce did not win CRM by having the best database — it won by building an ecosystem of apps and integrations that made switching costly. Anthropic may be executing the same playbook in real time: Fable 5 is the halo product that establishes credibility, Claude Code is the adoption wedge, and the subscription bundle is the actual revenue engine.

What This Means for You

If you are building on AI APIs today, the Fable 5 data should change how you think about your stack:

💡 For developers: Build model-routing into your architecture now. Use Fable/frontier for complex reasoning tasks (novel code generation, multi-step research, safety-critical decisions) and route 90% of volume through Opus/Sonnet-class models. Tools like OpenRouter Fusion can automate this. Anthropic itself showed that a Fable orchestrator with cheaper model workers achieves 96% of full-Fable performance at 46% of the cost.

For CIOs and procurement teams: The AI token economics have matured. Think in cost-per-task, not capability-per-benchmark. A model that scores 80% on SWE-Bench Pro versus one that scores 69% may not matter when your use case is summarizing meeting notes. Benchmark the actual workload, not the model.

For investors watching the Anthropic IPO: The Ramp data is a feature, not a bug. It shows that Anthropic has pricing power across its model family — Opus 4.8's surge proves enterprises are spending more on Anthropic overall, just not on the most expensive tier. The question is whether the market will value Anthropic as a premium model company or as a platform company. The answer determines whether $2 trillion is cheap or expensive.

The Lesson History Keeps Teaching

Every generation of technology produces this exact dynamic. Intel made the best x86 processors for decades, and AMD ate their lunch on price-performance. Oracle built the best database engine, and Postgres captured the volume. Nvidia makes the best GPUs, and the market keeps asking whether the cost is sustainable.

The AI model market is not different. It is just younger.

Anthropic's Fable 5 is the best AI model available today. The prediction markets confirm it. The benchmarks confirm it. The brief government ban confirmed it in the most dramatic way possible.

But the best model is capturing 11% of its own maker's enterprise revenue. The cheaper alternatives are winning — not because they are better, but because they are good enough. And in technology markets, "good enough" has always been the most dangerous phrase in the English language.

The question is not whether Anthropic survives. At $65 billion in annualized revenue with a model family that spans every price point, it will. The question is whether the frontier model business — the idea that the most capable AI commands premium pricing — is a sustainable category or a transitional phase.

The smartest move for Anthropic is the one it already seems to be making: stop trying to sell the best model and start selling the best system. Models commoditize. Platforms compound. The Ramp data is not a verdict on Anthropic's future — it is a signal that the future is not in per-token pricing. The companies that figure this out first will own the next decade of AI infrastructure. The ones that keep optimizing for benchmark bragging rights will own a beautiful trophy and a shrinking margin.

Based on the Ramp data, the market has already answered. It just hasn't told Anthropic yet.

Originally published at ComputeLeap