In partnership with
Global hiring doesn't come with a playbook. Join Oyster's webinars and events to learn how leading companies are hiring, expanding, and staying compliant across borders.
Cognition shipped a coding model last week priced per task, $1.97 a task, while everyone else still bills per million tokens. Odd unit to pick. I think they picked it on purpose, and I only worked out why after watching my own run burn. I'll come back to it at the end.
This week
Fable went dark at 78% context occupancy with a nasty refactor still open. So I asked it to write the plan, then handed the plan to Claude Code: 7 tasks, one subagent per task, a review gate behind each one. What I watched wasn't a lesson about models. It was a lesson about blast radius.
Task 1: "add these 4 SQL tables verbatim, write a test that checks they exist." Fully specified, zero judgment. It went to Haiku, the cheap tier: 76K tokens for the implementer, about 108K for its reviewer, approved on the first pass, done.
Task 3: "wire the new maintenance function into every place that writes to the database." Twelve call sites. Same mid-tier model as three other tasks. The reviewer caught a summary table being recomputed on every row inside a loop over thousands of rows, instead of once per batch. The fix introduced a workaround somewhere else, and the re-review caught that the fix had reintroduced the exact same per-row bug, just relocated to another file. Final bill: implementer 227K, review 160K, fix 164K, re-review 160K, second fix 122K, second re-review 140K. Six dispatches, roughly 970K tokens, for 25 lines of plan.
Task 4: comparable size, better specified, fewer moving parts. Two dispatches, approved immediately.
The verdict: the model tier is not your bill. The review loop is, and its length is set by how many places a task touches and how tightly it was specified going in. Reaching for the top tier "to be safe" on a mechanical task buys you nothing. A cheap model on a wide, vague task bounces through just as many fix-and-re-review cycles, arguably more.
Before your next dispatch, count the call sites. Task 3 touched 12 and cost 6 dispatches; Task 4 touched a handful and cost 2. Anything past 3 files gets split into single-responsibility pieces before it goes out, and budget 4x the implementer's first-pass number for anything non-trivial. That multiplier is the review doing its job, not waste: the per-row hammering it caught would have shipped to production in silence.
One brand shipped 30+ landing pages last week. No developers.
A DTC brand briefed Viktor inside Slack: one landing page per Meta ad group, mapped to a different headline variant. He wrote the code, deployed each page to their subdomain, posted the URLs back in #marketing, and now monitors performance across the set.
Their content team uses him to draft email flows, generate creative variants, and audit Klaviyo segments every Friday. Their growth lead uses him to catch spend anomalies before the day starts.
20,000+ teams now have the same setup: one AI employee across every marketing tool. A teammate who ships work in Slack and Microsoft Teams.
In the news
Cognition released SWE-1.7 on July 9, RL post-trained on Kimi K2.7 Code inside the Devin harness. It scores 42.3% on their FrontierCode 1.1 benchmark, against 46.5% for Claude Opus 4.8 and 43.0% for GPT-5.5. The score isn't the interesting part: the price is posted at $1.97 per completed task, not per million tokens.
This week on the blog
What Is Your AI Maturity Level? A 3-Question Test Based on Claude Code's Adoption Framework: the 5 adoption steps Boris Cherny mapped, and the one thing the score doesn't measure.
Humans Just Became Cheaper Than Software: what changes when an agent's marginal cost drops below a license fee.
23 Side Businesses to Start With Claude Code This Summer: 23 ideas you can actually execute, not a trend list.
Revenue Stack: July 2026
Same rule applies to your content or monetization stack: the expensive part is never the tool you pick, it's the number of round trips a badly scoped job forces on you afterwards. Scope first, pay once.
What decides which model tier you reach for?
Phil
PS: reply and tell me the last task you dispatched to an agent knowing it already touched ten files.
Back to that $1.97 per task. Billing per token bills the one number neither you nor I control directly. Billing per task admits what my run showed in black and white: what sets the price isn't the model doing the work, it's how many times it has to come back to it.
You're receiving this because you signed up at rentierdigital.beehiiv.com.