On October 7 one of the most widely deployed open-source projects in mapping made agent-assisted code a sanctionable offence, and deleted the careful policy it had spent eight months running in order to do it. The replacement is three sentences long. The diff is seven lines added, one hundred and forty removed.

The reason was not bad code. It was arithmetic that a Harvard working paper had already put a number on, and that number is the most useful thing in this issue.

QGIS made agent-assisted contributions a sanctionable offence, and deleted LLVM's policy to do it.

QGIS — the open-source GIS that a great many governments and utilities run their mapping on — merged QEP428 on October 7 at 21:00 UTC. The GitHub API puts the change at +7 / −140 across one file. That file is now, in its entirety:

The 140 deleted lines matter more than the 7 added. They were QGIS's AI tool use policy, adopted on January 31 and credited in its own footer as "a quasi direct adaptation" of LLVM's AI Tool Use Policy — the human-in-the-loop standard a lot of projects copied. Use what tools you like, read everything the model wrote before you ask a human to, label it with an Assisted-by: trailer. It had a framework for "extractive contributions," a quote from Nadia Eghbal's Working in Public, and a golden rule that a contribution should be worth more than the time it takes to review.

It lasted eight months. LLVM's copy is still live and unchanged; we opened it on October 13.

Proposal author Nyall Dawson was direct about why, and the reason is not the one you would guess:

Not a security incident. Not a licence scare. Not a pile of broken commits. A maintainer left. Dawson also stated the intent plainly: start closed, reopen case by case, and the cases he named first were "ai can be used as a replacement for ide autocomplete" and "ai can be used for language translations."

Two details that show how fast this was done. The merged file still carries Even Rouault's name and email at the top as author — Rouault wrote the policy that QEP428 deleted, and he voted +1 for the deletion, from a GitHub account now named rouault-on-strike. And the header has two Date fields that disagree, 2026/01/31 and 2026/09/24, because the replacement was pasted over the top rather than rewritten. The live AI policy of a major open-source project is three sentences with a broken header.

The review bottleneck now has a number, and it is 3.45 days.

The study QGIS's maintainers were describing from experience already exists. Fiona Chen and James Stratton of Harvard have been circulating Artificial Intelligence in the Firm: Bottlenecks in Software Production since January 7, with a current version dated August 4. It uses a proprietary dataset from the engineering analytics platform Jellyfish: 300 million work events across 718 firms, combining GitHub activity, Jira issues and Google Calendar, in a staggered difference-in-differences design that exploits when each firm adopted.

Both AI assistants and AI agents raise coding productivity. Neither passes that productivity through to shipped software. The reason is downstream:

And the objection you would raise first is already closed. The pull requests are not bigger: the effect on lines added per PR is statistically indistinguishable from zero (−67.99, standard error 126.91, against a baseline of 1,168 lines). The same-sized change now takes half again as long to get through.

Here is the finding that no roundup we read carried, and the reason this issue exists:

AI assistants had no significant effect on review time at all. Agents added 3.45 days; assistants came in at −0.38 days with a standard error of 0.97, against their own baseline of 9.34 days.

Standard error is the measurement's own wobble — how far the answer would move if you ran the study again on a different sample of firms. When the effect you measured is smaller than its wobble, you have not found an effect; you have found noise. Agents: 3.45 against a wobble of 1.60, comfortably real. Assistants: −0.38 against a wobble of 0.97, which is a fancy way of writing zero. Same pattern for changes-requested and comments per pull request — agents moved both, assistants moved neither.

So the only large-scale measurement of the problem QGIS is reacting to draws its line exactly where Dawson said he would want to reopen the door. Autocomplete is not what breaks review, and the ban catches it anyway: QGIS prohibited "LLM, AI or agent assisted" contributions, and on October 6 two contributors asked for the word "written" so that autocomplete and IDE ranking would fall outside it. The text merged on October 7 does not contain it. Dawson has been consistent that this is a deliberate stop-gap rather than a permanent position, so this is not a gotcha — but the distinction the evidence supports was available, and the policy as merged does not make it.

The paper also quotes the people living it. One junior engineer:

Can AI review its way out of this? Not by March 2026 it could not. By then nearly 80% of firms had used AI tools somewhere in review, but only 23.3% of review comments were AI-generated, and only 10.8% of pull requests received even one AI-generated comment. (Ars Technica's write-up rendered that second figure as agents being "responsible for ... 10.8 percent of all pull requests," which is a different and larger claim than the paper makes. The paper's sentence is about comments received, not pull requests authored.)

Our caveats, because they are real: this is a working paper, not peer-reviewed, and its data stops in March 2026 — before Opus 5.5, before the current agent generation. The time split is partly imputed, though the authors validate it against a survey of 100 engineers on Prolific, who self-report spending 30% of their week writing code against the 42% the data implies.

Meta's Muse Wonder scores four points higher and costs 2.8 times more per task.

Meta launched Muse Wonder and Muse Wonder Full on October 12, fourteen months into what it calls a ground-up overhaul, trained at its gigawatt-scale Prometheus data centre. Artificial Analysis measured both the same day: Intelligence Index v4.3 at max reasoning effort puts Muse Wonder at 52 and Muse Wonder Full at 56, against Muse Spark 1.3 at 48, on a six-week release gap. The gains concentrate in agent work — Terminal-Bench 4.0 goes from 33% to 57% on Full, and GDPval-AA v2.1 climbs 176 Elo.

Meta's launch post contains no prices. Artificial Analysis supplies them, and they are the story:

Before → After. Muse Spark 1.3 billed $1.25 input / $0.15 cached / $4.25 output per million tokens. Muse Wonder bills $4 / $0.20 / $20. Output per token roughly 4.7 times higher. Input token use also went up 28%, from 6.03M to 7.69M per Index task.

Put those together and cost per Artificial Analysis Index task goes from $1.60 to $4.52 — 2.8 times — for four Index points. Input alone is $3.43 of that $4.52, or 76% of the bill; the output you read is the cheap part. The model is genuinely more efficient on the way out, at 54k output tokens per task against Spark 1.3's 60k, and it still costs nearly three times as much, because it reads far more to get there.

And the headline number is not purchasable. Muse Wonder Full — the 56, the 57% on Terminal-Bench — is not priced per token and is only available to Muse Code subscribers. The model you can call through the Meta Model API is the 52. If you are putting Meta next to Anthropic or OpenAI on a spreadsheet, check which of the two you actually have a rate card for.

Credit where due: Muse Wonder Full cut its hallucination rate from 33% to 23%, which Artificial Analysis notes is the second-lowest of any model scoring 45 or more on the Index, behind Gemini 4 Argon.

xAI put Grok chat and Cursor coding on the same quota.

The X Subscription went live on October 12: Grok, Grok Bot, Cursor and X Premium on one plan. Free, Plus at $8, Premium at $30, Super at $100, Ultra at $200. xAI's table says Premium delivers what SuperGrok plus Cursor Pro plus X Premium+ did, "previously $90/month combined."

The mechanism is one sentence and it is worth reading twice:

One pool, spent "whether you use them in the Grok apps, the Grok tab on X, or Cursor." xAI promises "at least twice as much coding usage with our models as the comparable Cursor plan," and for teams 3x the coding usage at a lower per-seat price than old Cursor Teams. That is a real cut if coding is all you do. If your team also generates video and images in Grok — which the same plan covers — those draw down the allowance your agent loop needs, and nothing on the announcement page gives the exchange rate between a rendered video and an hour of Cursor.

One line deserves separate attention from anyone running Claude or GPT inside Cursor: third-party models stay available on Premium and above, but "usage is billed separately at API list prices, through top-ups." Outside the pool, at list, as a top-up. Existing SuperGrok, Cursor and X Premium subscribers keep their current plan and price unless they switch.

We could not open x.ai/pricing — it returns a Cloudflare challenge to us — so the per-tier multipliers (4x, 15x and 50x the Plus allowance) come from the announcement page only, and we have not seen the usage table behind them.

Anthropic's own arithmetic deflates its scariest result: two million devices buys about 40 H100s.

Anthropic published research on self-replicating models on October 12, and the headline writes itself: AI that spreads across smart ovens and off-brand Android TV boxes, with no datacentre to switch off. The findings are real. A 9B model got inference running on 8 of the 9 hardware types tested. They bought 15 Android TV boxes at roughly $55 each and sharded a 26B model across them. In a contained fleet, GLM-5.2 completed full replication, reaching 8 of 8 nodes in 4 of 5 trials.

Then Anthropic does the multiplication, and the result is the opposite of alarming. Scaled to the Kimwolf botnet's observed peak of over two million devices: about 133,000 replicas, roughly 80,000 tokens per second across the entire swarm. Their own comparison:

The largest Android TV botnet anyone has observed, fully converted to inference, is worth about one rack. The demo ran at 0.39 tokens per second, and a single inference setup takes 71 to 150 hours.

One correction, because the secondary coverage inverted it. Several summaries, including the Chinese daily briefing this issue drew its leads from, reported the entry vector as unauthenticated ADB ports — advice-shaped, as if you should go and audit your own fleet. That is not a finding about the world; it is a simplification Anthropic built into the test rig deliberately, and they flag it: "Our evaluation keeps the cyber-exploit element of the replication trivial (there's no authentication over the ADB port), which makes this an upper bound on real-world success." They add, fairly, that real IoT security is often not much better — Mirai spread on default manufacturer passwords.

The forecast is the part to diarise. Anthropic plots small open models against the Epoch Capabilities Index, fits a trend of 13.7 points per year at an R² of 0.95, and concludes a sub-30B model should reach the capability floor for replication by the end of 2026. As of October 13 the floor and the size ceiling do not intersect. That is a statement about this quarter, not about next year.

80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.

Read the report

Production got cheap. Absorption did not — and absorption is what you are actually buying.

Nothing above is a story about a model being good or bad. Every one is a story about the thing standing downstream of the model, waiting to take delivery. For QGIS that is a reviewer, and the resource it ran out of was a person, not a quality bar. For Muse Wonder it is the input side: the bill went up 2.8 times partly because the model reads 28% more, and reading is 76% of the invoice, so you are paying for consumption rather than generation. For xAI it is one shared glass that your video and your agent loop both drink from. For Anthropic's worm it is physics — two million compromised devices add up to one rack, and no model improvement changes that.

The line to repeat to a colleague: capability is getting cheaper and the capacity to absorb it is not, so your shipping rate is set downstream of the model. Which means the highest-leverage number in your stack is probably not on any rate card.

The 20-minute job this week: stop measuring your agents' output and measure your absorption. Pull the last 30 merged pull requests in your main repo, record the days from submitted to merged, and split them by whether an agent wrote the change. You now have your own version of the 7.03-day baseline and your own 3.45. If the gap is there, the cheapest fix is not a better model — it is whatever gets a reviewer to yes faster, and the honest candidates are smaller diffs, machine-checkable gates, and a written rule where your codebase currently contains two contradictory patterns. Then run ls -R .claude/ and see how many of the four levers you actually ship.

⚠️ The strongest argument against all of this: "the reviewers are the bottleneck" is also precisely what a guild says when its leverage is under threat, and review capacity is the one constraint here that automation can genuinely attack. Machine-generated review comments stood at 23.3% as of March 2026, which is where the data stops — a bottleneck at 23% penetration with an immature tool is not the same kind of object as a bottleneck at physics. QGIS's own thread carries the uncomfortable version: contributors pointed out the ban also catches LLM-assisted security auditing, in a project where a maintainer states that 100% of submitted security advisories have been written with AI since the current system started. Closing the door protects the reviewers you have. It does not get you more of them.

Who on your team could answer the question QGIS will now ask?

The penalty in that three-sentence policy is not a rejected pull request. It is "code reversion and contributor rights sanctioning" — merged work backed out, contributor rights curtailed. QGIS is one project, but it adopted LLVM's policy in January because LLVM's was the obvious one to copy, and plenty of others copied the same text. The question is what the next project does when it decides the compromise did not work.

So: if an upstream maintainer asked how a specific patch from your organisation was written, who would answer, and from what record? Not a policy — a record. Tessl's Ian Johnson ran the same ticket three times against a deliberately messy three-era repo. The bare agent walked straight past the correct 2026 domain layer to a 2023 god file, reached into a frozen module twice, and passed 1 of 8 checks. With a charter written, same ticket, 8 of 8. His line is the one to sit with: everything that would have stopped the agent "was a file I had not written yet."

If you ship patches upstream, those files are now your paperwork too.

The best sentence we read this fortnight is in the QGIS thread, from a contributor named Amalvir who had submitted two AI-assisted bug fixes and turned up to vote for the policy banning them:

Which brings us to our one ask.

What is the gap in your own stack between submitted and merged? Reply with the two numbers — your baseline and your agent-written changes — and I will read every one. Next issue is better because of them.

If this was worth your time, forward it to the person on your team who does the reviewing.