The one-line version: this week's signal is money and machinery. Anthropic's price ladder has a hard cliff on August 31 — Sonnet 5's promotional $2/$10 rate reverts to $3/$15, a 50% jump, on September 1. Microsoft's Project Perception hit public preview on August 3, turning agentic security from a research demo into a product you can try. And venture money kept pouring into agent infrastructure — four disclosed rounds from $75M to $135M. For a team of one, all three point at the same to-do: re-price your stack and know what's now buyable.

1. Anthropic's price ladder and the August 31 cliff#

Here's the current Claude API ladder, per million tokens (input / output), straight from Anthropic's pricing page:

On top of that, Batch API is 50% off both input and output, and prompt caching drops cache reads to 10% of the base input price (a ~90% discount on cached input). Those discounts stack — but they apply to whatever the base rate is, so when Sonnet 5's base jumps on September 1, your batched, cached Sonnet-5 bill jumps with it.

The number that matters: Sonnet 5's $2/$10 is a temporary promo, and the page states it plainly — introductory pricing runs "through August 31, 2026, after which the standard pricing of $3/$15 per million input/output tokens will take effect." That's a 50% increase on both dimensions, arriving whether or not you touch a line of code.

What it means for founders: if you sized your agent budget on the promo, your Sonnet-5 workloads cost 50% more next month by default. This is the month to re-price the whole ladder, not just Sonnet — a lot of bulk work that's sitting on Sonnet belongs on Haiku 4.5 at $1/$5, and only the hardest reasoning needs Opus 5. Our companion decision piece walks the model-by-model call in Opus 5 vs Sonnet 5 vs Haiku 4.5: which Claude model for the agent job, and the build-it version is in the three-tier Claude model router that cuts your agent bill.

2. Project Perception enters public preview#

Microsoft announced Project Perception on July 27, and it entered public preview on August 3, 2026 (GeekWire, Axios). It's a coordinated agentic security system that runs three classes of agents: red-team agents that model how an attacker could move through a system, blue-team agents that triage and prioritize the resulting risk, and green-team agents that apply the fixes.

Underneath is MAI-Cyber-1-Flash, Microsoft's first cybersecurity-specialized model, built to shoulder the bulk of routine security work at low cost while a larger model handles only the hardest cases. In Microsoft's MDASH vulnerability-management harness, the combined system reportedly hit ~96% on the CyberGym benchmark at roughly half the cost of its prior configuration — a 90/10-style tiering play, cheap model on volume, expensive model on the edge cases.

What it means for founders: agentic security just moved from "interesting research" to "a preview you can point at your own systems." You don't have to build red/blue/green loops yourself to benefit from them anymore — the pattern is becoming a product. We unpacked the cost architecture behind it in Project Perception and MAI-Cyber-Flash's 90/10 model tiering — and the tiering logic there is the same one that should govern your own agent stack.

3. Capital keeps concentrating on agent infrastructure#

The rounds that closed as Q3 opened make the theme obvious — money is flowing to the agent control plane and to vertical agents:

And the freshest disclosure: on July 29, Freehand raised a $75M Series B co-led by Battery Ventures and NewRoad Capital Partners, for AI agents that manage supply-chain spend at Fortune 500 companies — a textbook vertical-agent bet (Forbes).

What it means for founders: capital is concentrating on the plumbing (control plane, cheaper inference) and on agents that own a specific workflow end to end. That tells you where the tooling — and the acquisition interest — is heading. Our full read on the split is in the agent-funding map: control plane vs. vertical bet.

The thread#

Three signals, one decision. The cost side (Anthropic's cliff) and the capital side (four agent rounds) both say the same thing: the agent stack is being priced and financed in real time, and the cheap-vs-premium math shifts again on September 1. The capability side (Project Perception) says the sophisticated patterns — tiered models, coordinated agents — are becoming things you buy, not things you build from scratch. For a solo founder, the leverage move is small and dull and worth an afternoon: put a swappable client in front of your model calls, re-price the full Claude ladder before the promo expires (start with which Claude model for the agent job and wire it up with the three-tier router), and note that the security pattern you'd never have staffed is now a preview away. If you missed it, last week's wire tracked the same current from the model-access angle.