Copilot's $40,000-for-$40 Reset, the Framework Era Ends & 'Make the Change Easy'

For weeks this roundup has tracked capabilities — workflows, harnesses, the labs walking down the stack. Today the conversation snapped back to who pays for the compute underneath all of it. GitHub reset Copilot's pricing and the timeline lit up; Uber put a dollar cap on coding agents; and the recurring subtext — that a year of near-free inference was a subsidy, not a business model — finally got said out loud. Threaded through it: Jerry Liu's frank admission that the framework era he helped build is over, Matt Pocock's tiny prompt tweak that's quietly making agents better, and a renewed argument about who actually owns the software an agent depends on.

The Subsidy Era Ends

The biggest story of the day was a pricing change, and the surprise was who defended it. GitHub reset Copilot's limits, and the cancellation notes started flying — one quote-tweeted by Theo read "I successfully burned all of my monthly tokens in under 3 days thanks to your garbage new pricing model." Theo's response (155K views, 1,116 likes) was the math, not the outrage: "In order to hit the limit of your $40 Copilot plan, you have to do at least $60 of inference. The previous limit structure was entirely broken. You could do $40,000+ of inference for $40. You're not hurting them by cancelling your sub."

His longer version (with a breakdown video) sketched the trap: "1. They charged per message because nobody knew what a token was. 2. 'Messages' got longer and more expensive because of agents. 3. they delayed changes because people like you would flip out. 4. It went too far and they ran out of compute." And the part that reframes the whole debate — when a reply argued that cancelling still hurts revenue, Theo shot back: "Not when they're compute constrained and the freed up compute can be sold to enterprises for literally 1,000x more revenue."

The replies did the analysis:

  • The cleanest framing, from @18_priyansh: "40k to 60 is a 666x cap shrink. CAC subsidy is done, IDE habit is the moat. devs running >$60/mo inference are arbitraging copilot overage vs byo-key direct. which path bleeds first?"
  • The "this is actually good" take, from @AlecMey25715067: "this is actually a win for the AI coding space. Trying to sell a request-based AI agent is a complete detriment to the industry... They were forced to actively make the product worse to cut costs."
  • The bankruptcy counterfactual, from @MyFleetingDream: "If a startup did the GHCP pay-for-requests pricing model for the last year, they would've gone bankrupt the week after autopilot mode came out."

Theo then poured gasoline on it (51K views): "Copilot's biggest issue isn't cost. It isn't product... They have the insurmountable problem of having the actual stupidest user base of any AI product. The average Lovable user understands basic software development better than Copilot users." That earned a deserved pile-on — the strongest rebuttal, from @dave_mostoller: "Many people work for large companies that restrict what their employees can use, and usually, that's copilot... Another one of your bad takes that only considers indie devs and startups, and ignores the majority of devs that work for enterprise." But buried in the noise was the genuinely interesting structural point, from @4nanei: "copilot didn't win those seats, it got bundled into them. every dev who'd never choose an ai tool now has one. that's the opposite of lovable, which earned each user. so is a captive base the problem here, or the moat nobody else can buy?" (See also @bretgreenstein's data point: Copilot is used by "only 3% of M365 users.")

Simon Willison supplied the enterprise mirror. His post (237K views): "Uber reportedly now caps coding agents at $1,500/month per employee per tool — seems sensible to me, but it's also an interesting hint at the value Uber thinks these tools are providing" (writeup). The replies turned a billing limit into a thesis about measurement:

  • @alokitwrites: "The cap feels like a measurement forcing function. If a tool is worth more than $1,500/month, the team should be able to point at shipped work, not just usage. The awkward bit: the bill shows up per seat, but the value shows up in project throughput."
  • @ilmAI_: "the $1500 is a measure of what they'll tolerate paying, but not what the tools return... measuring whether the output actually got better beyond # of PRs/commits is hard so cap the bill."
  • @CollinWilkins7 sketched the new org chart: "give TL/SR $1500 CC license and limit jr to $500 because they won't maximize... have a QA / reviewer agent in pipeline."

Put together, the day's two pricing stories rhyme: the consumer subsidy is being clawed back because it was never sustainable, and the enterprise buyers who are sustainable are now busy figuring out what a coding agent is actually worth against a salary line.

The Framework Era Is Over

Jerry Liu — who built one of the most-installed pieces of AI plumbing of the last three years — sat down and said the era he helped create is finished. The framing came from Conor Bronsdon's Chain of Thought episode (watch · listen): "The agent harness ate the abstraction layer. The patterns @llama_index used to wrap (query rewriting, reasoning loops, provider glue) have solidified into Claude Code, OpenClaw, and Manus. The abstraction moved up, out of Python classes and into natural-language skills and MCP tools."

So what survives? "Context quality. And Jerry bet the company on it, looking to become the best company in the world at extracting data and context locked in PDFs." The episode lays out the wager: "Why 95%+ accuracy is the real bar in legal, insurance, and finance, and why pointing Opus or GPT at a million PDFs gets you ~80%" — plus "how you survive disrupting your own product while companies 1,000x your size try to do it to you." One sharp reply, from @ggooodluckk: "In legal/insurance/finance, context extraction is not just retrieval quality; it is whether the system can preserve source, hierarchy, conflicts, and audit trail well enough for a human to trust the next step." And @baibaida: "jerry admitting the era he helped build is over takes guts. but the pivot makes sense — when the model eats the wrapper, the only moat left is the data underneath it."

Jerry kept poking at the same nerve all day. His shower thought, riffing on a Stanford study where law professors preferred AI tutoring answers 75% of the time: "If 1. AI is smarter than humans at law, therapy, etc. 2. Humans still like talking to other humans. Then: Humans are just an AI wrapper. Everyone should just regurgitate what Claude tells them in real time. Long Cluely." He doubled down: "humans are just claude code for claude code." The replies refused to let "wrapper" be the whole story:

  • @Haggis: "the part claude can't replace is accountability. nobody wants legal advice from a thing that can't be disbarred. humans stay in the loop because someone has to be holding the bag when it's wrong."
  • @subramanya: "the human part is probably not raw answers anymore. it is taste, trust, and being accountable when the Claude answer is wrong."
  • @ktchn_ngnr42 closed the loop back to harnesses: "human was the harness. the whole 'prompt engineering → harness engineering' is just human having taught models to do what they used to do."

This is the same "abstraction moves up" thesis that's run through the week — yesterday it was the labs eating the app layer; today it's a framework author conceding the wrapper layer and re-staking his claim on the one thing the harness can't generate: clean context.

Prompt Craft & the Skill Factory

Matt Pocock's smallest prompt of the week may be his most useful. The tweak (62K views, 1,200 likes), borrowed from Kent Beck: "Before implementation, look for opportunities to prefactor the code to make the implementation easier. 'Make the change easy, then make the easy change.'" ("Prefactor" = refactor pre-implementation.) The field reports were enthusiastic — @ChrisMasterton: "HOLY MOLY! What an amazing and significant difference this makes on very diverse projects — one SaaS, the other a game" — and the thread quickly turned into a survey of prior art: a /tidy-first skill based on Beck's Tidy First, a /to-design skill slotted between to-prd and to-issues, and Angie Jones tagging Kent Beck himself: "look you just might make into a skill 😆."

The skeptics were the valuable part. @DevaBuilds named the failure mode: "Models will happily prefactor when the direct change would have been 5 lines. Do you constrain it or does the prompt let the model decide when restructuring is actually worth it?" And @dionysuzx proposed a one-word fix: prefactor to make the implementation "simpler," not "easier" — since "easier" can license more complex code. The judgment call, in other words, is exactly the thing the prompt offloads to the model.

The companion theme was how skills actually travel between people. Matt floated a distribution mechanism (56K views): "1. Create an npm package with your skills. 2. On postinstall, run a script that symlinks them to .claude/skills. Simple, versioned, intuitive. Any reason this wouldn't work?" Two reasons, immediately:

  • @jherr (Jack Herrington) — "Because all of the package managers are blocking postinstall scripts now?" — which Matt conceded: "👍 checks out." (The collision with this week's supply-chain anxiety is the point; see the next section.)
  • @Everlier offered the heavier-but-safer pattern: a "company-wide LLM gateway (Anthropic/OpenAI-compatible) which supports Skills, Prompt Injection processing, MCPs" with telemetry and the ability to "block specific unsafe actions" before the tool call reaches the client.

Matt's portability hedge — "Symlink .claude/skills to .agent/skills" — nods at the multi-harness future, but the postinstall hole means the cleanest-looking idea is the one that's hardest to ship safely.

And the prompt-as-craft thread closed on a human note. Thariq shared (95K views) that Anthropic's now-circulating "teacher" prompt — the one that won't let you leave a Claude session until you actually understand the work — was written by Suzanne Wang, "a writer in her little spare time" (her short story Mall of America is, fittingly, about the human condition that happens to involve AI). The reply that landed, from @Lbdev__: "the real failure mode is never bad code, it's code you can't re-derive a week later. A prompt that refuses to end the session until you've proven you understood is the actual fix." And @garg__suhani: "Good prompts are often just good writing in disguise."

Ownership, Supply Chain & the Claw at Scale

Armin Ronacher wants the Debian model back. His post (29K views, 658 likes) argued that a generation raised on npm and PyPI misses how open source used to work: "When Debian or a Linux distribution ships a dependency they take responsibility of it. If there is a security issue and it's not fixed upstream, they fix it for their users... They own their software including all the things it depends on." The payoff is about shifting the burden off the lone maintainer: "It does not put so much focus and attention on the one overworked developer who just happened to have too much of the world depend on their library. Remember: they carry a responsibility they never signed up to and never got compensated for."

The thread was unusually thoughtful:

  • @mylifcc crystallized it: "Debian absorbs upstream security debt invisibly; npm tweets shame at whoever's holding the bag."
  • @faust_: "npm made everyone forget that dependencies used to come with a human contract, not just a license file."
  • @tnkuehne drew the practical lesson: ownership "makes you think twice about the number of dependencies you want to add."
  • The honest counterpoint, from @juliknl: does that include the infamous Debian OpenSSL "fix" that introduced a severe security hole? Ronacher didn't flinch: "I too did not always appreciate what Debian did with my packages but that's just how open source works." (He separately pointed readers at Andrew Tridgell's rsync and outrage essay, urging people to "adjust your priors" on the vibecoding-a-core-tool drama.)

It's a neat collision with the section above: the npm-postinstall skill idea is convenient precisely because npm asks nobody to own anything — which is also why package managers are now blocking postinstall and why maintainers keep getting shamed instead of supported.

Meanwhile, OpenClaw kept compounding. Steipete reported (67K views) that, npm plus Docker, GitHub, internal deployments and forks, "real number is more in the 10-20 million downloads/week," attributing the jump to nothing flashy: "We made it faster, improved release quality, added more useful features and extended compatibility to more models." The week also brought a GitHub "Fastest Growing OpenSource App of All Time" award and the video of his Microsoft Build talk, "Build the thing that builds the thing" (BRK245).

The more useful moment was a security one. When a reply flagged a "Five OpenClaw 0-Days let Attackers Hijack Trusted AI Agent Access" headline, Steipete didn't deflect — he triaged it in public: "The headline is a bit hotter than the actual risk: these were medium-severity mutable display-name allowlist bugs in adversarial chats, not unauthenticated remote takeovers. They required specific channel config plus same workspace/tenant/contact access, and the fixes gate name matching behind explicit opt-in... But ya know, that doesn't get as many clicks." It's the right shape of response for a project that spent its first six months warning people not to run it and now ships to enterprises — and it lands squarely in the same "who owns the security of what you depend on" question Ronacher was asking.

Around the Ecosystem

  • Town raises $55M and earns swyx's highest compliment — accidental adoption. He wrote (17K views) that Town — "the Devin for Everything Else" — "organically spread to @liamcbride and the rest of our team with no further hyping or enablement from me. this never happens!" Town came out of beta with a Series A led by a16z; co-founder Jean-Denis Greze replied graciously. The thread's recurring read: organic, un-mandated spread is "the only honest demand signal in enterprise software."
  • LLMJunky's agent-tooling drop. The Codex App for Linux now supports Remote Control (github.com/am-will/codex-app); an open-source OnShape CLI lets agents "create and modify CAD workflows from the command line — sketches, extrudes, variables, exports" (github.com/am-will/onshape-cli); and the Hermes Agent landed a desktop app. CAD-by-agent and remote-controlled coding apps are both early, but they're the kind of "agent reaches into a domain tool" surface that's about to multiply.
  • Lee Robinson pushes back on the "builder" merge. Responding to the podcast-circuit line that "engineering, product, and design are all merging into a 'builder' role," he wasn't buying it: "This feels like an oversimplification and podcast talking point. Reality is a lot more complex." (It pairs with his earlier, much-quoted benchmarks rant — "benchmarks matter but they are not a substitute for extensively testing the model yourself with real work" — and his note that he's personally found Opus 4.8 "really good" despite the benchmark grumbling.)
  • The MAI tech report keeps drawing eyes. From the Microsoft Build fallout, swyx amplified a note that Microsoft's in-house model "uses zero synthetic data or distillation from previous models... reasoning, agentic behavior, tool use are all learned fully during post-training" — "one of the most transparent for a model at this scale."

Sources: RSS + thread scans of the accounts in TASK.md (Pocock, Theo, Thariq, LLMJunky, Mitsuhiko, bcherny, Steipete, swyx, Simon Willison, Karpathy, Jerry Liu, potetotes, Lee Robinson) for the ~24–36h window ending June 4, 2026. Engagement figures are point-in-time from Nitter.