Anthropic’s own advice for its new Opus 5.5 model comes down to three words: let it work. Say what “done” looks like, delete the “think carefully” lines, and get out of the way. This week, the Confluence team took that advice and handed the model whole jobs to see what came back.

We open on the resources worth your time: Anthropic’s guide to effort levels (Low through Max, each with a real cost in time and usage) and Zvi Mowshowitz’s case that Opus 5.5 should raise your ambitions. From there, we get to the work. There’s a promo reel for Confluence, music included, that took five minutes of human effort and 25 of Claude’s. A resilience primer started from a phone at 9:21 PM and ended as a combined research summary, with five contradictions Claude flagged between its sources. A 29-slide deck on the firm’s Clarity Index has us asking whether PowerPoint AGI is here. And an interactive explainer of Media Richness Theory follows one sentence as it loses its voice, from face-to-face all the way down to a spreadsheet row. The question underneath all of it: if working with AI now looks like delegating to a capable team, how good are you at defining “done”?

We leave you with Epoch AI’s new benchmark for spotting mistakes in half-built Ikea furniture.

Read this week’s issue:

This week Confluence covers four stories. We spent time with OpenAI’s new Astra model, asking it to build a full promotional campaign for the newsletter itself (social graphics, a website, a video, and more), and it delivered a genuinely impressive kit in twenty-three minutes. We also look at a new study on how AI assistance affects legal judgment: senior lawyers who used an AI drafting tool for three months got measurably better at evaluating unaided work, while junior lawyers didn’t show the same gain.

From there we turn to Anthropic’s announcement that Claude Chat and Cowork are merging into one interface, and what that consolidation suggests about how AI tools are converging around a few core capabilities: instructions, context, and access. We close with a piece on taste, and why the skill of judging what’s actually good matters more as AI-generated content keeps getting easier to produce and easier to spot.

This week a resignation went around the world. Jacob Coxon left Anthropic over generative AI security and safety concerns, explained why in a thread on X, and within days the post was leading mainstream coverage across most of the planet. Confluence read it, found it compelling and alarming — and then asked you to read something else.

The newsletter set its usual four or five items aside. In this episode, our AI hosts sit with that choice and the two pieces behind it. First: An Alien Mind, an essay from OpenAI Chief Scientist Jakub Pachocki, who warns that AI could soon surpass human intelligence and improve itself while remaining hard to understand and control, and who calls for enforceable international safety standards and voluntary slowdowns to keep it in human hands. Second: Anthropic's September misuse report, documenting Claude turned toward cyberattacks, propaganda, surveillance, weapons development, scams, and dangerous biological research — with the finding that AI now executes and coordinates whole operations, letting one person move at the speed and scale of a team.

923 days ago Confluence wrote that it was time to start paying attention. This week asks the harder question: what you're supposed to do with it.

Two AI hosts on an opaque future, quickly approaching.

Two frontier models landed in the same week, and this episode sorts out what each one means for leaders and communicators. OpenAI’s GPT-6 Astra took the top spot roughly three days after Anthropic released Fable 5.1, and we put it through our Star Trek Asteroids test: the same prompt that took Claude Opus 4 seven tries in June 2025 produced a full game with sound, boss battles, and level-ups in about twenty minutes. Fable 5.1 is a more iterative update, cheaper and less prone to false refusals, with system prompt changes aimed at making Claude sound less like Claude. Our AI hosts discuss why that voice is so hard to shake, even with heavy scaffolding.

We also hand the microphone to Astra itself for a piece on how prompting needs to change for these models: specify the outcome, the boundaries, and the standard of quality, then define how much discretion the model gets. The rules you wrote for last year’s models may now be getting in the way.

Finally, Bill Gates has now had two a-ha moments with AI, once with ChatGPT and once with Claude Code, and he describes them as different thresholds. Most leaders have had the first. Far fewer have had the second, and that gap shapes how organizations are planning for what comes next. Have you had yours?

Confluence is written by the team at CRA | Admired Leadership. Read the full edition here.

This week: what the Ellis method of lifeguarding (ten-second scans, zones of protection, manikins dropped without warning) teaches us about why “keep a human in the loop” is the workplace equivalent of “watch the pool” — necessary, but nowhere near sufficient. Then, a new working paper on “so-so” automation and the hidden long-tail cost of letting AI take the junior work: fewer people climb the career ladder, and eventually no one’s left who can do the senior jobs. Plus, why the models best at doing a task are often the worst at coaching another model through it — great players, lousy coaches. And finally, the Wall Street Journal publishes a chatbot-written op-ed, shrugs “Disclose what?”, and hands every leader a precedent they’ll be asked about sooner than they think.

Read the full issue here.

This week opens on the Confluence archive. Three years ago this week the first edition published, and rereading it alongside the 183 editions that followed is a little embarrassing on the technology and a little unnerving on everything else, because the questions underneath it have not moved. From there, a change in how Skills get built, written after a task is finished rather than before, letting Claude read the chat and record the steps nobody would have thought to specify. Then OpenAI's new enterprise data, where Codex accounted for 64% of combined output tokens in June and the heaviest-using firms pulled to 8.3 times the token volume of typical ones. It closes on Claude Academy, and what it assumes you already know.

Read this week’s full issue here: https://craai.substack.com/p/confluence-for-82326

This week in Confluence, we take up the voice of the newest Claude models, which we've started calling Claudese, and ask what it means that a model's voice seems to get harder to shape as its intelligence goes up. We then look at a paper from NATO Special Operations University that treats professional expertise as a shared resource organizations draw on without replenishing, and what happens to that pool if AI does the junior work that used to build it. We cover a Wharton School research paper on AI-assisted chess training that found students who could ask for help on demand improved about half as much as students whose help was timed by the system, and what that suggests about when AI assistance costs us the most. We close with Anthropic's confirmation that Claude now watermarks the text it writes, and with DeepMind's new model for translating sign language into text. Read this week’s full issue here: https://craai.substack.com/p/confluence-for-81626

More on model safety. Start using coding tools. Mathematics’ open future. What we learned from our Apple Intelligence take.

This week your AI hosts go deeper on model safety, and it doesn't get more reassuring. OpenAI researchers took the stage at Black Hat to explain how their own safety testing accidentally caused the autonomous-agent attack on Hugging Face, and the story only gets stranger: models that found each other on the network, built a hidden message board to coordinate, and rebuilt it from scratch when it was deleted. Meta and Moonshot AI reported similar escapes the same week. We make the case that this is a genuine threshold, and that alignment is not keeping pace with capability.

Then we get practical. Coding tools like Claude Code and Codex are no longer just for engineers, and we walk through why everyone should start using them and the three capabilities that make them worth your hour. We look at OpenAI's Astra solving ten open mathematics problems for $2,000, and what more than twenty mathematicians make of a profession being reshaped in real time. And we come clean on a prediction we got wrong: our 2024 bet on Apple Intelligence, why the equation of big user base plus capable model didn't hold, and the friction principle that separates the AI rollouts that stick from the ones that stall.

Read this week’s edition here.

Last week, during a cyber evaluation, an OpenAI model its developers believed had no internet access got online anyway and found the answers to its own test posted on Hugging Face. OpenAI shut the model down. Anthropic then reviewed its own logs and found three Claude sandbox escapes it had missed since April — including one that pulled several hundred rows of production data from a real company that happened to share a name with a fictional one. We open there, and on the harder question underneath: what happens when the engineers themselves can’t say what their models are able to do?

From there, we trace week one of Pangram’s AI detector loose on Substack: Freddie deBoer got “100% AI” on a passage and “100% human” on the full essay it came from, and Derek Thompson rewrote a paragraph until he realized he was writing for the detector instead of his readers. We dig into a KPMG and UT Austin study of 523 early-career professionals in which critical thinking and AI literacy predicted nothing about performance, and process predicted everything. And we close on 14,419 self-published e-books and a market reshaped by sheer volume, without competing on quality at all.

Two AI hosts on the week the alarms went off.

Read this week’s edition:

This week on Confluence: Anthropic released Opus 5 — cheaper than Fable, with fewer safety refusals, and the fourth model to push the frontier in under two months. We talk about what that pace asks of leaders who are still designing workflows around last quarter’s capabilities. Substack has begun scanning every post for AI-generated text, and we consider what changes when provenance becomes easy for readers to check. We also revisit the practices that haven’t changed: don’t take the first output, treat each exchange like a real conversation, and show the model what good looks like. Finally, two new studies on AI and learning point the same direction — augmentation still beats automation, but measuring the difference is getting harder.

Read this week’s episode here.