Ramp's June 2026 AI Index put Anthropic at 41% of enterprise AI spend. OpenAI at 32.3%. First time Anthropic led. Then, three days later, Fable 5 and Mythos 5 went offline by government export control order.

These two events are not directly related. But they arrived back to back, and the combination says something useful about how builders should be thinking about AI provider concentration right now.

What the Ramp data actually shows

Ramp's spend index tracks real corporate card and bank account transactions across their customer base. Not a survey, not a press release, not analyst estimates. Actual money flowing to API providers, aggregated.

When Ramp shows Anthropic ahead, it means enterprises are paying Anthropic more than they're paying OpenAI — at least across Ramp's customer set. That's a meaningful signal.

The shift happened fast. Earlier in 2026, OpenAI led by a comfortable margin. The Claude Fable 5 launch — now offline — likely contributed to the June spike. Enterprises evaluating Fable 5 would have bumped API spend during the evaluation window, which was exactly June 10-12. Which means the 41% figure may already be softening following the shutdown. I don't know that for certain. But it's the obvious question to sit with.

What the outage actually looks like for builders

Let me be precise: Fable 5 and Mythos 5 are offline. Claude Opus 4.8 — launched May 29 at $5/M input and $25/M output — is still running normally.

So: if your production stack was on Opus 4.8, nothing changed operationally. If you had migrated to Fable 5, or were evaluating it for production, you lost that model. No warning, no migration path offered, no restoration timeline.

Anyways, the real question isn't "is Anthropic down?" — they're not, not fully. The question is: what's your fallback when any capability tier you depend on disappears without notice? Fable 5 was probably in the roadmap at a lot of shops. It's not available today.

The Mistral Medium 3.5 row is worth special attention. 128B dense weights, $1.50/M input, open-weight with a permissive license. If you want to eliminate provider shutdown risk entirely, the self-hosting path is viable now in a way it wasn't a year ago.

What I think about provider concentration

Most builders I know have a multi-provider setup in theory. In practice, production routing at most shops is 90%+ to one provider. Switching costs are real. Model evaluation takes time. That's not crazy — until the week it matters.

The government ordering a frontier model offline is a new category of provider risk. Most concentration-risk discussions are about pricing surprises or operational failures. Export control suspension is a different animal. It can move in 24-48 hours. It doesn't require a technical incident. The political dynamics driving it are largely invisible to builders until the letter arrives.

The part I keep coming back to is the investor angle. Amazon holds more than $8B in Anthropic. Amazon's CEO reportedly raised concerns with White House officials on June 11. The shutdown order arrived the next afternoon. You can construct an innocent reading of that sequence — legitimate security concern, appropriate escalation, government acting on the merits. You can construct a less innocent one. I genuinely don't know which is right.

What I do know: an investor in your AI provider can apparently influence government action on that provider's models. That's now in the threat model. It wasn't before.

Source spread

• Ramp — AI Index June 2026 — builder. Real transaction data; primary source for the 41% / 32.3% enterprise spend figures.
• Anthropic — Fable 5 / Mythos 5 suspension statement — builder. Confirms the export control order; thin on timeline and technical detail; disputes jailbreak severity framing.
• Fortune — The Amazon-White House backstory — skeptic. Best-sourced reconstruction of the Jassy escalation path; named officials, dated calls.
• Mistral — Medium 3.5 launch — builder. Primary source for 128B open-weight specs, $1.50/M pricing, and license terms.

Pros & cons

What's real:

• Anthropic's enterprise momentum was genuine and data-backed before this week. Ramp's data is transactional. The lead was real.
• The outage is partial. Opus 4.8 is running. For shops not already on Fable 5, nothing changed operationally.
• Mistral Medium 3.5 at $1.50/M and open-weight is a meaningful self-hosting option. Not a drop-in replacement, but a real hedge.

What deserves a side-eye:

• The June 41% number was almost certainly boosted by Fable 5 evaluation spending during June 10-12. If the model stays offline, expect that figure to reverse in July's data.
• "No restoration timeline" from Anthropic is a commercially uncomfortable position. Indefinitely offline while competitors operate normally is a compounding problem for their enterprise relationships.
• The investor-to-government escalation path is now documented. That belongs in your vendor risk framework. Not as a reason to leave Anthropic — as a fact about how the AI platform market actually works.

Further reading

• Ramp AI Index June 2026 — enterprise spend data; source for 41% / 32.3% figures
• Anthropic — Fable 5 and Mythos 5 suspension statement
• Fortune — How a warning from Amazon led the White House to shut down Anthropic's Mythos model
• Mistral — Medium 3.5 open-weight launch — specs, pricing, license terms
• Anthropic — Claude Opus 4.8 — the Anthropic model still running