Cursor's 15x spread means the next AI coding dollar goes to review, not seats

*Vendor averages hide who actually merges the code, and the queue that forms behind them*

Opening read

The most useful number today is the fifteen-fold gap between Cursor's heaviest users and its median user, and my read is that it turns every per-seat return calculation for AI coding tools into a guess. Output is piling up in a few people and a few agents, which moves the pressure to whoever reviews the code. The rest of the day confirms the running stories: Anthropic's first outside evaluator is paid by Anthropic, stolen developer credentials remain a leading way into the systems those agents touch, and Alibaba's Qwen 3.8 weights landed while it previews the next architecture.

AI coding: the average hides a 15x spread

The reasonable read on AI coding tools is that they lift the whole team, and the customer stories back it. Coinbase says its average engineer merges 55% more pull requests (proposed code changes submitted for review) than at the start of 2026, and at Grab more than a third of merge requests now involve Cursor. Cursor Blog (https://cursor.com/blog/coinbase) Cursor Blog (https://cursor.com/blog/grab) A smart buyer takes that as a rising tide and sizes next year's seat budget to it. I held a version of that view too.

The shape of the data breaks it. Cursor's own July post put its top 1% of users at 46 times the AI-assisted lines of code per day of the median active user, and its Spring 2026 developer-habits report found the same tier landing roughly fifteen merged pull requests for every one the median author did. Cursor Blog (https://cursor.com/blog/cfo-council) Nate Jones (https://natesnewsletter.substack.com/p/scale-ai-developer-productivity) A mean drawn from that distribution is a statement about a handful of people. The far end is startling: one engineer, Lauren Tan, reported 1,000 merged pull requests in July and 2,462 in August. Nate Jones (https://natesnewsletter.substack.com/p/scale-ai-developer-productivity) Whether those changes are the kind your team needs, the evidence cannot show. What the counts do raise is a hypothesis: nobody reads at that pace, so either review is thinning out or it is becoming the queue.

The same skew shows up between companies, and the review strain is already visible in some places. Inside Cursor, cloud agents (AI agents that write code on hosted machines rather than a developer's laptop) authored about 10% of merged pull requests in December and more than 60% by this month, while Jellyfish's analysis of 20 million pull requests through late 2025 found autonomous agents under 2% of merges. Cursor Blog (https://cursor.com/blog/cloud-agent-environment) Cursor Blog (https://cursor.com/blog/self-hosted-machines) AI Engineer (https://www.youtube.com/watch?v=WqZq8L-v9pA) Those are different curves, and a frontier shop's case study says almost nothing about your team's shape. On the review side, one AI Engineer talk reported that since AI coding tools spread in early 2026, review comments have grown longer and more merges go through unreviewed. AI Engineer (https://www.youtube.com/watch?v=Ib5GBkD555M) That is the constraint starting to bind, not a forecast of it. Nate Jones puts the consequence in his title: the finished work waits, since the bottleneck has moved from writing code to checking it.

For a business leader approving a seat expansion, the stake is that a per-seat model of the return overstates it when the return sits with a few users. For an engineering leader, the stake is that the next constraint may be review capacity, and buying more generation would make it worse. Here is the check I would run before signing either budget. Pull the merged pull requests from your main repository for the last 90 days and the 90 days before that. Count merges per author, then compare the top author with the median; separately, compute the median time from opening to merge in each window. If a handful of names carry most of the total, the seat is the wrong unit and the deal should be priced on the people actually using it; the exact ratio matters less than the concentration. If merges rose while time-to-merge also rose, that is your evidence that review is the constraint, and I would fund reviewer time and review tooling before adding a single seat. If time-to-merge held flat, the review worry does not apply to you yet.

Outside testing: the evaluated still pays the evaluator

Earlier this month Anthropic proposed, and OpenAI matched, a commitment to embed outside evaluators with employee-level access inside the labs, and this brief has treated who picks and pays those evaluators as the question that matters. Today's evidence sharpens it. Both commitments now stand, OpenAI has added a call for international coordination, and Anthropic's first named evaluator is Accenture's Faculty unit, which will red-team the models (attack them to find failures) and run alignment assessments. Don't Worry About the Vase (https://thezvi.substack.com/p/the-quest-for-embedded-evaluators) Anthropic pays for that work directly, and the two firms have each put a floor of $1 billion, spread across five years, behind building out evaluation capacity. Anthropic News (https://www.anthropic.com/news/accenture-embedded-evaluation) So far, then, the lab picked and pays its own evaluator. No equivalent naming or funding disclosure from OpenAI appears in today's evidence, so the same charge cannot yet be laid against it. A business leader who cites lab safety attestations in vendor risk reviews should file an Accenture-signed report as vendor-funded evidence, because the company being evaluated is the one writing the check.

Credential theft: the door into your agents is a stolen token

Infostealers, malware that harvests saved passwords and login tokens, are a leading way attackers first enter cloud, code, and AI environments, and three families (Lumma C2, RedLine, and Vidar) account for 85.7% of the incidents Wiz found in NordStellar's data. About half of all infostealer families now collect GitHub credentials, and they lift live session tokens, so multi-factor authentication does not help once a session exists. Wiz Research (https://www.wiz.io/blog/infostealer-incursion-cloud-ai-credentials) This is not a new pattern. Microsoft's August ChainDrop writeup pointed at stolen maintainer credentials as the entry, so the route predates the September 25 report. Microsoft Security Blog (https://www.microsoft.com/en-us/security/blog/2026/08/04/chaindrop-supply-chain-compromise-anatomy-self-propagating-worm/) The stake for an engineer building agents is direct: an agent holding long-lived keys on a laptop is the target, and the decision is short-lived credentials before any new tool integration.

Open weights: Qwen 3.8 shipped, and Alibaba is previewing Qwen 4

Alibaba promised open weights (model files you can download and run yourself) for Qwen3.8-Max the week after its early August announcement, and it delivered: the 2.4T-A95B weights reached Hugging Face on August 12 and the 27B on August 14. Latent Space (https://www.latent.space/p/ainews-qwen-38-max24t-and-27b-new) OpenRouter Blog (https://openrouter.ai/blog/insights/qwen-3-8/) The kept promise matters less than the label on the next model. Alibaba is pitching Qwen3.8-Flash-Next as an experimental first look at the design Qwen 4 will be built on, which tells an engineer choosing a self-hosted model two things at once. OpenRouter Blog (https://openrouter.ai/blog/insights/qwen-3-8/) Run 3.8 now, because the weights are real and downloadable. But keep tuning of the machines that serve it shallow, because the preview says the next generation changes the layout, and nothing in today's evidence says when Qwen 4 arrives.

Augur calls

• Call: Anthropic will publicly name at least one embedded evaluator organization other than Accenture by 2026-11-30.
• Criteria: An Anthropic newsroom post, or a report from a major outlet, identifies a second named embedded evaluator with an agreement in place; no such announcement by the due date is a miss.
• Due: 2026-11-30
• Odds: 70%

• Call: Alibaba will publish downloadable Qwen 4 weights on Hugging Face by 2027-03-31.
• Criteria: A model repository under the Qwen organization on Hugging Face named as a Qwen 4 release with weight files available; a preview or API-only release does not count.
• Due: 2027-03-31
• Odds: 60%

What to watch

• Anthropic said on September 18 that more evaluator partnerships would follow within weeks, and the names will show whether any evaluator is neither a paid vendor nor a customer. Anthropic News (https://www.anthropic.com/news/accenture-embedded-evaluation)
• Dario Amodei's scheduled White House dinner with President Trump lands in the same week the evaluator commitments firmed up; any readout on oversight is the thing to watch. TechCrunch AI (https://techcrunch.com/2026/09/27/anthropics-ceo-is-about-to-have-dinner-with-president-trump/)
• Hugging Face's Transformers library added Qwen4-Exp support in August, so the model code exists before the weights do. Transformers Releases (https://github.com/huggingface/transformers/releases/tag/v5.16.0)
• Sakana AI's PC-ALM trains 1,000-layer networks without backpropagation (the standard way neural networks learn), and Sakana says it may prove useful on neuromorphic chips (hardware built to mimic how brain neurons signal), where running these dynamics is cheaper than on GPUs; the test is a result at a size someone ships. Sakana AI (https://sakana.ai/pc-alm/)
• Reply bots that plagued Twitter have surfaced on Bluesky, and a reply bot checker Simon Willison flagged uses Bluesky's freely available API (the open interface other programs can query) to spot them by reply speed and absence of original posts. Simon Willison's Weblog (https://simonwillison.net/2026/Sep/27/bluesky-bot-check/)

Verification notes

Cursor's 15x and 46x figures come from the vendor's own user data, and Lauren Tan's monthly counts are self-reported with nothing in the evidence about the size or acceptance rate of those changes. Nate Jones (https://natesnewsletter.substack.com/p/scale-ai-developer-productivity) The review-quality decline comes from a single conference talk without a stated dataset. Wiz's 85.7% is its read of NordStellar's dataset, and the embedded evaluator arrangements are announced commitments with no published evaluation yet.

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-09-28?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-09-28)