Signal #92: GitHub Killed the Copilot Seat

Aligned shows you what buyers are doing between meetings inside your most important deals then surfaces the risks, openings, and next steps. No more guessing where a deal stands or finding out a champion went quiet too late. Just real visibility, so you can act before it's too late.

Start Using Aligned Free

GitHub stopped selling Copilot by the seat. A Croatian study shows why ChatGPT skips most company websites, a Sofia lab priced every finished robot task, and Poland's AI regulator can now fine. You're starting to pay AI for finished work, not for heads.

Happy Friday, friends.

For three years you bought AI the way you bought Slack: one licence per person, paid monthly, whether anyone used it or not. This week four separate things pointed the same way. The tools are starting to be priced, measured and policed by the work they finish.

Start with GitHub if you sign off on software budgets. Forward the Croatia study to whoever runs your website.

On 28 October GitHub told its Enterprise customers they no longer need a separate Copilot subscription. They get the whole suite, including the Copilot app, code review and automations, and pay only for the AI credits they use. GitHub also made Copilot Fusion generally available, a router that sends each task to whichever model it thinks fits. GitHub says Fusion comes within two points of Claude Code on Opus 5.5 Max at about half the cost per task.

A seat is a guess about how much AI each person will use, and a credit is a meter. Under seats, the developer who never opens Copilot costs you as much as the one running agents all night. Under credits you pay for output, so the expensive model only runs when the cheap one isn't good enough. The catch is that a meter has no ceiling, and a team that loves the tool can outspend any seat price. Anthropic shipped the other half of the answer the same day: reports that show managers what their team's token spend produced.

Before your next renewal, pull last quarter's Copilot usage and count the seats nobody touched. Existing customers can stay on their current plans for now, so there's time to run both numbers side by side. If credits win, agree a monthly limit per team before the first invoice, not after.

Most Croatian company websites are close to invisible to ChatGPT, for a boring reason. RankRush, a Croatian AI-visibility platform, went through 157,000 sources cited in 14,319 AI answers about Croatian companies, and Netokracija published the results on 27 October. When AI cited a company's own site, blog and advice pages took 21.9% of those citations and price lists took 0.2%, mostly because the prices sat in PDFs. Sixteen of the 17 company sites examined had no developed structured data.

This matters because the click is moving. Pew data cited in the study shows people click a link in 8% of searches with an AI summary, against 15% without one, while AI search grew 386% in a quarter on Claude and 370% on Gemini. The AI picks the page that answers the question most directly, not the one Google ranks highest, so a small firm with one clear page can beat a big one with a slick homepage. You could call this SEO with a new name. Partly, but the study found AI checks your site's claims against reviews, maps and directories, which old SEO ignored.

It's a weekend job. Put your prices on the page as text, add company and FAQ markup, and write the one page that answers the question customers ask right before they buy. RankRush's Marin Zlatarić thinks early AI citations will be harder to dislodge than Google rankings ever were, which makes the next few months the cheap ones.

The best AI model tested finishes about a third of ten-step robot jobs. That's the headline number from LADDER, a public benchmark that INSAIT, the AI institute at Sofia University, launched with NVIDIA on Monday. On the lighter test GPT-6 Astra finishes 62.5% of tasks at $6.62 per finished task, while Gemini 4 Argon finishes 55.9% for $1.33. On the hardest tests, eight to ten steps in a row, the best system completes 34.8%.

The unit is the story: LADDER scores finished jobs and what each one cost, not how smart the answer sounds. Read that way, GPT-6 Astra is 6.6 points better than Gemini 4 Argon and about five times more expensive per job, which makes it the wrong pick for most warehouses. The counterargument is that the expensive model wins wherever a failed job is costly, and that's fair. But now you can put a number on that trade instead of guessing, and a Bulgarian lab built the scale.

Anyone pitching you a robot "powered by frontier AI" should now get two questions: what does it cost per completed task, and how often does it finish a ten-step job? The leaderboard updates as new models land. Bookmark it before your next vendor meeting.

Since Wednesday, a Polish regulator can inspect your company and ask how your AI works. On 28 October the enforcement chapters of Poland's Act on Artificial Intelligence Systems came into force, and KRiBSI, the new Commission for AI Development and Safety, became the country's AI market watchdog. It can demand documentation, order fixes, suspend a system and impose fines. The law covers companies that use AI as well as those that build it, so buying from a vendor doesn't move your obligations onto the vendor.

Poland is the largest market most CEE startups sell into, and one of the first in the region with a single AI regulator that has these powers. For anyone selling there, the cost of an AI feature now includes the logs and documents that prove what it did. Founders will say this is Brussels paperwork arriving early, and some of it is. The upside is that KRiBSI can also give individual opinions on whether a use case counts as high-risk, so you can ask before launch instead of finding out from a fine.

Make a list this month of every AI system your product runs on or ships, and flag anything that touches hiring, credit or health. Check that the logs exist. Where a use sits near the high-risk line, request the opinion now.

Dev: Your coding agent's built-in browser can read your SSH keys. Forever Security showed that one rigged file or page can reach your machine through the embedded browsers in Cursor, Copilot, Codex, VS Code and Claude Desktop. In Codex and VS Code its tests pulled SSH keys, .env files and npm tokens, and neither was patched as of 29 October. Cursor and Copilot (1.1.29) are fixed. Update today, and don't open unreviewed projects in agent previews.

Productivity: Granola now types for you, free. The meeting notetaker added voice typing on Mac: you press Fn, talk, and it cleans up the text and drops it into any app, from email to CRM notes to prompts. Words are unlimited and free, with iOS and Windows coming. Try it on your post-call follow-ups this week.

Marketing: Google opens Gemini and Search to your app. Google's new Connected Apps Platform lets you plug your service into Gemini and Search through an MCP server, so a user can design a flyer in Canva or order from Instacart without leaving the answer. It's a US-only waitlist preview for now. Build the MCP server anyway, because the same one works for Claude and ChatGPT.

Dev: Mistral Large 4 weights are public. Mistral released the weights of its 1-trillion-parameter model (about 52 billion active per task) on Hugging Face on 29 October. A quantised version runs on four B200 GPUs, and it scores 79% on the CyberGym-E2E bug-patching test. If you want a European frontier model you host yourself, you now have one; read the licence on the model card before reselling it as an API.

Next edition soon,
Çelik