The Brief — The Great Intelligence Markdown

Two weeks ago the frontier looked like a luxury aisle. Now it’s a clearance rack. Anthropic dropped Opus 5 at Fable-5-adjacent intelligence for half the old price, OpenAI answered by slashing what it charges for the cheap end of GPT-5.6, and Anthropic’s Sonnet 5 is still running an introductory rate that expires at the end of the month. If your AI cost model is more than a quarter old, it’s wrong — and probably wrong in your favor.

Underneath the price war, the more interesting story is where the money is actually flowing: not to another chatbot, but to agents that do the boring back-office work nobody wants and to a browser that treats your open tabs as a to-do list. Here’s what mattered this week, and what a one-person shop should do about it.

Opus 5 lands, and the “frontier” gets cheaper

Anthropic shipped Claude Opus 5 on July 24, calling it a step-change for its top tier — state-of-the-art on coding and knowledge-work evals, and close to its own flagship Fable 5 at roughly half the cost. Pricing lands at $5 per million input tokens and $25 output, same as the Opus 4.8 it replaces, with up to 90% off via prompt caching. There’s also a Fast mode that runs about 2.5x quicker for double the base rate.

So what: “Frontier intelligence at half price” used to be a roadmap promise. Now it’s a Tuesday. If you priced out an agent workflow six months ago and shelved it because the token math didn’t work, re-run the numbers — the thing you couldn’t afford in January may quietly pencil out today.

OpenAI cuts, and the race to the bottom accelerates

OpenAI didn’t wait long. On July 30 it cut the price of GPT-5.6 Luna — its fastest, cheapest model — by 80%, and knocked 20% off Terra, the mid-tier. That’s on top of the three-model Sol/Terra/Luna lineup it launched in July, all with million-token context windows. Meanwhile Anthropic’s Sonnet 5, its cheaper agent workhorse, is holding an introductory $2/$10 per-million rate that jumps to $3/$15 after August 31.

So what: The labs are now competing on price at the volume end, which is exactly the end that matters when you’re running something on every inbound email or support ticket. Two moves: pin your high-volume, low-stakes jobs to the cheapest tier that still passes your eval, and if you’re building on Sonnet 5, lock in usage before the September bump.

Polar: a browser that runs your agents off your tabs

Kevin Jiang, who worked on Perplexity’s Comet browser, launched Polar on July 29 with a $5.7M seed led by Madrona (angels include ex-GitHub CEO Thomas Dohmke and Modal’s Erik Bernhardsson). The pitch: while everyone else builds AI browsers that book your flights, Polar targets knowledge work — schedule workflows, save prompts, and assign tasks to agents based on whatever tabs you have open. It’s free with daily credits; $20/month buys more.

So what: Most early users run Polar as a second browser purely for automation, not daily driving — which is the honest way to adopt a tool like this. Point it at one repetitive tab-heavy chore you already do (pulling data across dashboards, prospect research, competitive checks) and let it eat that one thing before you trust it with more.

The money is going to boring

Ellis AI, founded by Cadre co-founder Ryan Williams, raised a $10M seed (First Round, Khosla, Thrive, Harlem Capital, and Ariel’s Mellody Hobson) on July 31 to point AI agents at private-credit back offices — the month-end grind of downloading files from five systems, reformatting data, reconciling balances, and chasing discrepancies by hand. The agents flag and prepare; humans still decide. It’s the same pattern behind Mercor’s reported $20B-valuation talks and a wave of vertical raises: unglamorous, workflow-deep, industry-specific.

So what: The durable AI businesses right now aren’t horizontal “do anything” assistants — they’re the ones that own one painful, repetitive workflow in one industry nobody’s automating. If your niche still closes its books by hand, that’s not a complaint. That’s a product.

A cheaper token is still a token

One caution to balance the sugar rush: cheaper per-token pricing has a way of quietly inflating consumption. Reporting this summer found engineers who leaned hardest on AI were roughly twice as productive — but burned about 10x the tokens to get there, with per-developer usage rising nearly 19x in nine months. And independent agent benchmarks still show even top models flubbing the majority of real professional tasks.

So what: Falling prices don’t cap your bill — usage does. Set a monthly token budget the same way you’d set an ad budget, watch it weekly, and treat any agent you deploy as a junior hire on probation: useful, worth the money, and not yet allowed to sign off on anything unsupervised.

The takeaway for the week: the cost of being smart just fell again, the winners are picking narrow lanes, and the only bill that grows is the one you don’t watch. Price your stack like it’s August 2026, not last quarter — because it is.