On Sept 22 Anthropic launched Claude Opus 5.5 and cut its API prices. Input went from $5 to $4 per million tokens, output from $25 to $20, and cache reads from 50 cents to 20 cents, which is the 60% cut in the subject line and the one that matters if you run agents (Anthropic announcement, Anthropic's cost breakdown). Anthropic says it runs about 30% faster than Opus 5 and has stronger defenses against prompt injection; those are the vendor's numbers, not ours. Subscribers do not get a cheaper bill; they get more usage inside the same five-hour limits. If you were already paying for Opus on the API, you just got a raise. If you were about to stop, look again.

The Sharp Take

The models are moving. The pricing is moving faster. The useful reaction is not to watch the rate card; it is to pull out the one job you shelved last month because Opus was too expensive for it, and run it again at the new rates. Either it pays now or you learn exactly why it still doesn't.

Top stories

1. Opus 5.5 shipped, and the cache-read cut is the real story

What happened. Anthropic released Claude Opus 5.5 on Sept 22 (Anthropic). Cache reads dropped 60%, input and output 20% (pricing page). Two footnotes worth knowing: when Anthropic's safeguards step in, a cybersecurity request can be answered by Opus 4.8 instead, and biology work needs its Life Sciences Verification program.

The part that matters. The headline is not the speed bump. It is the cache-read cut, and it pays out under one condition: your agent re-sends the same context, over and over, inside the cache window. On the API that window is five minutes by default, one hour if you pay for it; on a Claude subscription it is an hour (caching docs). So it is not "the bookkeeper who opens the same chart of accounts next month." It is the bookkeeper who closes one client's books in a single sitting, forty calls deep, with the chart of accounts in every call. Same for the investor screening thirty listings against one buy box in an afternoon. Session-shaped work, not calendar-shaped work.

For your business. If you run a batch job that repeats a big block of context (a policy, a rate card, a playbook, a codebase) across many calls in one session, that batch just got cheaper. If your Claude use is one question at a time, spread across the day, the 20% on input and output is your number, and it is still a raise.

Do this now. Pick your heaviest repeating Claude job. Run it once on Opus 5 and once on Opus 5.5, same prompts, same session. Record three things for each: total cost of the job, wall time, and whether the output passed your own check. Anthropic's own migration note says to run its prompt audit first because some old prompt patterns cost more on the new model (Anthropic). If 5.5 is cheaper and at least as good, switch that job's default. The rate on cache reads is 40% of what it was; what the job actually costs is what you are measuring.

2. Cowork is gone, Projects got memory, and most of you do not have it yet

What happened. Anthropic folded Cowork into the main Claude app on Sept 16 (Anthropic) and then announced a redesign of Projects: a Project becomes a shared memory that every thread adds to and draws from (Anthropic). Read the rollout line carefully. It starts with select Pro and Max subscribers who use cloud sessions in Claude Code and have no existing Projects, widens to more Claude Code users over the week, and only after that reaches Projects in the Claude app, then Team and Enterprise. Your existing Projects keep working as they do today.

The part that matters. Persistent memory is the change. You stop re-briefing your AI every morning. That is what makes the three-agent stack work at solo scale: one Project holds the standing context, three agents inside it do the jobs. It is also why the setup is worth doing now, before the memory upgrade reaches your account, because the Project you build today is the one that gets upgraded.

For your business. A bookkeeping firm keeps one Project per client with the chart of accounts and prior-month notes so every close starts warm. A real-estate investor keeps one Project holding the buy box and has it re-screen new listings without re-explaining the criteria. A consultant runs one Project called Proposals that already knows the rates and past scopes. One Project per recurring job, with the three roles from the guide working inside it.

Do this now. Take one recurring three-step task and set it up as a plain Claude Project this week, on whatever plan you have, using the guide in the Toolkit below. Attach the standing context once. Watch for this: the redesigned Projects landing in the Claude app for existing Pro and Max Projects. When it does, that Project carries the standing context forward; you update it when the job changes instead of re-explaining it every time.

3. Mollick cut a book trailer in 45 minutes without leaving ChatGPT

What happened. Ethan Mollick wrote up three of his own runs in "The Overhang" on Sept 18: GPT-6 Astra used Blender to build an animated 3D scene, wrote a script, generated voice, music and sound, and delivered a trailer for his book in 45 minutes. On a second run it drove a video generator through his browser and cut a 30-second promo. He says it ran within a normal paid plan's allowance. Where you get Astra matters: on Plus it is in ChatGPT Work and Codex, not plain Chat; Astra in Chat is Pro and up (OpenAI).

The part that matters. The capability is not new. What Mollick shows is the gap between what the model can already do and what most subscribers ask it to do. For a shop owner: the ceiling on one paid subscription is well above your current use of it. The browser step is available to anyone on a paid plan. The Blender step is one writer's result until somebody outside a Pro plan reproduces it.

Try this. Open ChatGPT Work on Plus or above. Pick three product or listing photos you already own. Give Astra one prompt: "Cut a 30-second vertical promo from these photos, with on-screen text I approve first." Review the storyboard before it renders. If the cut is watchable, run it once more with one taste note. Treat it as an experiment: Mollick's 45 minutes is his result, not a promise, and any video generator it reaches for may have its own costs. Budget an hour and your normal plan, and tell me what you got.

Watch for this. A reproducible screen recording of Astra driving Blender or another desktop app on a Plus plan, not a Pro plan. When that lands, the desktop workflow is one every paid subscriber can run, and we will say so.

Receipts

Lunduke Journal counts AI-assisted patches at 17.25% of patches submitted to the Linux kernel this month, with the community adopting an "Assisted-by:" line in commit metadata (their count, not ours). Governor Newsom's Sept 18 executive order gave California agencies 60 days to recommend a frontier-AI "kill switch" and to define loss-of-control incidents (Office of the Governor). Hinton, Russell and 100-plus signatories asked labs the same week for embedded, non-commercial auditors; hours later Anthropic named Accenture its first embedded evaluator, each side expecting to invest at least $1 billion over five years (Anthropic, CNBC), which critics called exactly the commercial entanglement the letter warned about. If we're wrong about any of it, show us the receipts and we'll correct it.

The Experiment

We ran the Handoff Audit on this build to pick its own next task. Three real weekly jobs: read the briefing and pick stories, verify every quote against a primary source, compose the poll and split-test subject lines. The audit scored each 1 to 5 on repeatable, boundable, checkable, reversible-if-wrong, owned-credentials. Verification scored 25 straight because "does the URL exist, does the number match, and does the sentence around it say what we say it says" is checkable in under two minutes and reversible before publish. Story selection and the poll each scored 21. The pick: verification, first. What failed on the first pass: the audit hedged and returned all three tasks when we gave it vague descriptions. The fix, one line in the kickoff prompt requiring one-sentence tasks, is in the download. The delegation brief it produced is what this issue was checked against; two claims got softened on that pass.

The Toolkit

New today: The Handoff Audit (free skill, three-page PDF). Give it three tasks you already do every week. It hands back the one to delegate first, a delegation brief, and the two-minute test you run on your next pass to know it worked. Four readers of the Sept 21 issue asked for exactly this. Get the free Handoff Audit (the button below says $0.00 because it is free; same link). The paste-ready text is on that page too, so you copy from the page and read from the PDF. Verified 2026-09-22 on Claude Pro; the run in The Experiment above is the actual test.

The Handoff Audit
The Handoff Audit
The Claude Project instruction set that picks your first task to hand off and hands you back a delegation brief with the Monday test. Three-page PDF.
$0.00 usd

Runs it: The Solo Operator's 3-Agent Stack (free guide). Do, Check, Ship, one Project per recurring job. About 45 minutes to set up. Get the guide. Verified 2026-09-22 on Claude Pro.

The poll

Two takes on the Opus 5.5 pricing news, both real drafts from this build:

A. Opus 5.5 cut cache reads 60%. If your agent re-sends the same context all session long, rerun that job at the new rate this week.

B. Anthropic released Opus 5.5 on Sept 22 at $4 input, $20 output, 20 cents per million cache reads. Fable 5.1 quality at Opus 5 speeds, per Anthropic. Try it on your heaviest job this week.

Which of the two takes above (A: the punch, B: the numbers) would you forward to a peer who runs a small team? Results reported Friday with sample size.

Login or Subscribe to participate

One ask

If you already run one Claude Project for something recurring, tell me which task lives in it. One sentence. That is the reader signal that decides which guide gets built next. Just hit reply.

The Closer

Every year the frontier gets cheaper, and every year the same question comes back: cheaper for what? The rate card does not change your business. The job you retest because of it might. Pick the one thing you ruled out last month and run it again.

If you have a problem worth solving and a way to reach the people who have it, reply and tell me about it. That could be a report you rebuild every month or a process eating your week. It could also be twenty years in an industry, the connections that come with it, and a customer list waiting for somebody to build the thing. I take on a few of those a year.