BRIEF · AUG 21

Anthropic ahead on enterprise AI spend - but OpenAI is closing fast

Ramp data from 70,000 US businesses: Anthropic 44 percent, OpenAI 40 percent, and zero lock-in either way.

Ramp's data from 70,000 US businesses puts Anthropic at 44 percent of enterprise AI spend vs. OpenAI's 40 percent as of July 2026 - but OpenAI is growing faster in Q3. Vendor stickiness is near zero: raw model performance decides your bill, not platform lock-in.

Also in today's brief

Models & Releases

Cerebras CS-4 hits 4,400 tokens per second

Unveiled August 18, the CS-4 packs three WSE-3 Turbo wafers into one rack for 750 PFLOPs and 4,400 tokens per second per user on a 120B-parameter model - first shipments land in Q3. At that speed, an agent takes more reasoning steps in the same wall-clock time. Read our take →

Microsoft cuts MAI-Code Flash 73 percent and adds vision

The updated model rolled into every GitHub Copilot plan on August 11 with native image understanding, a 73 percent lower list price, and a 22 percent Terminal-Bench 2.1 gain for CLI tasks. The cheap default coding model just got materially better - re-check which model your team pins. Read the changelog →

Tools & How-To

Ramp opens its AI routing infrastructure to everyone

Three days after Stripe bought OpenRouter, Ramp opened its internal routing infrastructure spanning 8 AI providers - claiming 40 percent average inference savings on its own production traffic. Free through the end of 2026, US-only. Read our take →

ChatGPT can now read and send your iMessages on Mac

The August 20 Mac update lets ChatGPT read full iMessage, SMS, and RCS history and send texts as you - only in Work and Codex sessions, only on Apple Silicon, only behind a per-send confirmation OpenAI says not to disable. Read our take →

Quick Hits

•

Try Make. Cloud automation that connects the same apps as n8n without hosting anything - a fit for agentic workflows without a server to babysit. Try it →

•

Try Buttondown. The newsletter platform sending you this email. Clean API, no AMP, fair pricing. Try it →

From the Pondero Stack

n8n 2.35: Agent Builder test runs and per-tool approval - does it change the verdict?

n8n 2.35 adds Agent Builder test runs before live traffic and per-tool human approval on irreversible actions in Slack, Telegram, and n8n Chat. Verdict: still the pick for self-hosters escaping per-execution pricing, 4.5 out of 5. If you want cloud-only with zero servers, Make is the better pick. Read it →

How was today's brief?

★★★★★ Nailed it | ★★★ Solid | ★ Missed

Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.

Affiliate disclosure ·

Pondero earns commissions on some links. This does not affect our editorial picks. Full policy: pondero.ai/affiliate-disclosure/

Originally sent to subscribers of the Pondero Brief.View the original.

Pondero Brief: Anthropic leads OpenAI 44 to 40, but no one is locked in