Opus 5 Overtook Fable 5 in Enterprise Spend — Ramp Data

For years, the playbook was simple: ship a smarter model, raise the price, watch enterprises switch.

Fable 5 broke that pattern.

Ramp's August 2026 AI Index — built from anonymized spend across 70,000 US businesses — shows Anthropic's June flagship stuck at a low single-digit share of actual token volume and only ~11.4% of Anthropic-attributed dollars two months after launch. Meanwhile Claude Opus 5, released July 24, 2026 at half Fable's list price, has already overtaken Fable in enterprise spending, Ramp told the Financial Times and summarized in its public index.

TL;DR

What changed in enterprise model choice

Ramp economist Ara Kharazian summarized the shift: GPT-5.6 Sol is "really good, increasingly the choice for developers." Fable 5 "disappointed both in adoption and real-world application given price + data retention requirements."

Three forces explain the plateau:

1. Price gap — Fable at ~$10/M input is roughly double GPT-5.6 Sol and orders of magnitude above DeepSeek V4-Flash-class pricing. Enterprises optimize spend-per-merge, not benchmark bragging rights.

2. Good-enough tier — Opus 5 at half Fable's rate handles most agent harness workloads teams actually meter — code review, CRUD, doc refactors — without Mythos-class reasoning on every call.

3. Policy friction — Fable's 30-day retention (vs Zero Data Retention options on other tiers) adds procurement drag for regulated buyers already nervous about frontier cost.

Ramp's line: "With Fable 5, we've found a new upper bound for how much businesses are willing to spend on AI."

The cost math enterprises actually run

Ramp's data makes sense when you work the spreadsheet — not the leaderboard.

Example: 10M input tokens / month on coding assist

At 2× list price, Fable must deliver 2× the business value per token to justify default status. Ramp's 6% token share vs 11.4% dollar share gap shows enterprises bought some Fable — likely for tasks they believed required frontier capability — but routed volume elsewhere.

The token share below dollar share pattern means: teams that do use Fable consume fewer tokens at higher unit cost. That is consistent with "reserve frontier for hard problems" routing — but 6% of Anthropic tokens two months post-launch is a weak flagship signal.

Compare OpenAI: GPT-5.6 Sol at 25% of tokens suggests the "good enough + priced right" tier captured default volume. Fable did not replicate that dynamic inside Anthropic's own customer base.

How to implement model routing in your harness

Ramp's findings align with what explainx.ai has covered on why Fable 5 is not the best default. Implementation patterns:

1. Tiered routing by task type

User request → classifier (cheap/fast)
  ├─ boilerplate / tests / refactors → Opus 5 or Sonnet
  ├─ architecture / security review → Fable 5 (if ROI proven)
  └─ batch / offline → open-weight or GPT-5.6 Sol API

2. Escalation on failure

Start on Opus 5; escalate to Fable only when validation fails (lint errors persist, test suite red, human flags). ATT's routing used similar logic at enterprise scale — 56% spend reduction without quality collapse.

3. Retention-aware routing

Fable's 30-day data retention (vs ZDR options) adds procurement friction. Route regulated workloads to tiers with acceptable data handling before model capability enters the conversation.

4. Budget caps per tier

Finance teams increasingly set $/employee/month caps (top 1% at $7,400/employee). Harnesses that default to Fable burn caps on routine work. Explicit Opus default preserves budget for escalations.

What people are asking about Fable 5 adoption

"If Opus overtook Fable in dollars, is Fable dead?" No — 11.4% dollar share is a plateau, not zero. Some workloads still justify frontier spend. The story is default routing changed: newest ≠ automatic upgrade.

"Why does Anthropic still lead OpenAI in business adoption?" Ramp's July data: 43.5% vs 39.7% of paying US businesses. Anthropic wins accounts; OpenAI's GPT-5.6 Sol wins token volume within OpenAI. Platform choice and model-tier choice are separate decisions.

"Does this affect Claude Code / Max subscribers?" Ramp measures B2B card/bill-pay rails — API and enterprise contracts. Consumer Max plans may not fully appear. Direction still signals how procurement teams think about tier pricing.

"Should I downgrade my team's default from Fable to Opus?" If you have not benchmarked your task mix, yes — start there. Run A/B on merge rate, rework rate, and $/shipped feature before assuming frontier is required. See Fable vs GPT-5.6 vs open-weight comparison for task-level anchors.

Comparison with the "newest model wins" era

Historically, each Anthropic flagship pulled spend upward at launch — teams upgraded because capability jumps justified price jumps. Fable 5's June 9 launch followed months of frontier hype, but by August:

• Token volume stuck at 6%
• Dollar share at 11.4% — better, but still minority
• Opus 5 at half price overtook Fable in spend within ~4 weeks of July 24 launch

That breaks the automatic-upgrade assumption. Enterprises now treat models like compute SKUs — route by cost-per-acceptable-result, not badge prestige.

Why Claude Opus 5 feels worse to work with covered UX and harness friction separately. Ramp adds the finance lens: even when Opus UX is debated, wallets prefer the cheaper tier for volume work.

Vercel AI Gateway and other data sources

Ramp is not the only signal. Vercel's AI Gateway sample (cited in Heise and other coverage) showed Fable at roughly 13.2% of all model spend across providers — a different population (developer-heavy, gateway-routed traffic) but the same direction: minority share for Anthropic's flagship.

Triangulating sources:

None of these alone is definitive. Together they suggest Fable is not the default routing target for volume enterprise work in August 2026 — consistent with Opus overtaking Fable in Ramp dollar spend.

Subscription vs API economics

Ramp's data skews toward API and enterprise billing visible on corporate cards. It undercounts:

• Claude Max / Pro subscribers paying flat monthly fees regardless of per-token economics
• Annual enterprise contracts with committed spend not fully visible on Ramp rails
• Self-serve developers on personal cards mixed into small-business buckets

That caveat cuts both ways. Power users on Max plans may love Fable for interactive coding while their company's API spend routes to Opus. The Ramp story is specifically about how businesses allocate marginal API dollars — which is what matters for harness defaults and procurement policy.

For teams choosing between GPT-5.6 Sol API price cuts and Anthropic tiers, Ramp's cross-vendor comparison (GPT-5.6 Sol at 25% OpenAI tokens vs Fable at 6% Anthropic tokens) is the actionable signal.

Why this matters if you build with Claude

Model routing is now the default architecture — not "always Fable."

This connects to explainx.ai's earlier coverage of why Fable 5 is not the best default and Opus feeling worse to work with — UX and harness issues aside, finance teams are voting with wallets.

Anthropic still leads 43.5% of paying US businesses vs OpenAI's 39.7% in Ramp's July data (top 1% spend post) — but within Anthropic spend, the flagship is not where the volume went.

Caveats

• Ramp measures card/bill-pay rails — not every enterprise contract.
• Vercel's AI Gateway sample (cited elsewhere) showed Fable at ~13.2% of all model spend — different population, same direction: minority share.
• Power users on Max plans may not show up fully in per-token economics.
• Opus overtook Fable in dollars; Anthropic has not published a public token split.

What builders should do this week

1. Audit your harness default model — if Fable is hard-coded, run a two-week Opus default trial with escalation rules.
2. Instrument $/successful-task, not $/token — tokens are a cost input; shipped features are the output.
3. Document retention tier in your security review — Fable's 30-day policy may block regulated use regardless of IQ.
4. Compare subscription economics — best coding subscription under $20 posts matter when finance caps per-seat spend.
5. Plan for routing as architecture — token budget planning is not optional when frontier tiers cost 2×.

The Ramp data is a snapshot, not a forecast. But the direction is clear: enterprise buyers optimize spend-per-merge, and Opus 5 at half Fable's rate is winning the default slot.

Related on explainx.ai

• Top 1% spend $7,400/employee — Ramp AI Index
• Why Fable 5 is not the best default model
• Fable 5 vs GPT-5.6 vs open-weight comparison
• Anthropic 30-day Fable retention vs ZDR
• GPT-5.6 Sol official API price cut
• Best AI coding subscription under $20
• Token budget planning for agents
• ATT routes 56% of AI coding spend via model selection

Sources

• Ramp AI Index — August 2026
• Financial Times coverage via Implicator AI summary — August 23, 2026
• Heise — Fable 5 spend stagnates

Spend shares are Ramp's August 2026 snapshot; Anthropic pricing per public rate cards. Follow @explainx_ai for model economics.