Scale · founder · 7 min read
Your AI Coding Tool Just Started Keeping Your Data Longer
GitHub Copilot chat retention goes from 28 days to the life of your account on Sep 28. And Claude Fable 5.1 now requires retention to run at all.
Everyone watched the price this week. GitHub Copilot’s promotional credits expired today, exactly as scheduled, and we put the arithmetic on a page two weeks ago.
While that was happening, GitHub published two changes to how long your data is kept. Neither is a price change. Neither got covered. Both are defaults.
If you have ever pasted a customer record, a contract clause, a support ticket, or a half-finished pricing model into a chat window — and you have — this is the more consequential story.
Change one: 28 days becomes forever
On September 28, 2026, GitHub merges three products into one. Copilot Chat on github.com, Copilot Chat in GitHub Mobile, and the Copilot cloud agent become a single experience under a single policy.
Buried in the “what’s changing” list is this: Copilot on github.com migrates to the agent sessions experience, and as a result chat data will be retained for the life of the account instead of 28 days.
That is not a rounding change. Twenty-eight days is a window that eventually forgets. The life of the account is a permanent record of every question you asked your AI about your own business.
Three things make it sharper:
- It’s on by default. No action is required to be moved onto it.
- The opt-out is all-or-nothing. GitHub is explicit: if you opt out of the unified experience, you or your team “lose access to Copilot on github.com and GitHub Mobile” once it launches. You cannot keep the 28-day window and keep the product.
- It arrives as a feature announcement. The retention change sits in the third bullet of a section about product convergence, under a headline about policies and billing.
Business and enterprise admins have until September 28 to decide. Put it in a calendar.
Change two: a model that won’t run without retention
The second one is stranger, and it breaks a norm.
Claude Fable 5.1 went generally available in Copilot today, September 1. It’s Anthropic’s newest Mythos-class model, aimed at long-horizon autonomous coding, and it’s available on Pro+, Max, Business and Enterprise.
Here is the part that matters. In GitHub’s words: Claude Fable 5.1 requires data retention by default. When you use it, Anthropic retains data including prompts and outputs in order to operate the safety classifiers that detect harmful use. Every other Claude model in Copilot — except Fable 5 and Fable 5.1 — continues to run under zero data retention.
To be fair to both companies, two caveats belong here immediately. Retained data is not used to train Anthropic’s models. And the policy is off by default at the org level; an admin has to switch it on, and GitHub states plainly that enabling it “constitutes acknowledgment of this requirement.” That is a clean consent flow, better than most.
But the structural fact stands: the most capable model in the picker is the one you cannot use privately. Capability and privacy have been put on opposite ends of a slider, and the slider is in your admin’s hands, not yours.
The exemption has an expiry date
There is a carve-out, and it is time-boxed. Eligible enterprise customers can use Fable 5.1 and Fable 5 with zero data retention through the end of the calendar year, under what GitHub calls a time-bound exemption, while Anthropic rolls out Enterprise Frontier Safeguards.
After that, continued use requires EFS — which enables automated safety monitoring and, GitHub says, gives eligible customers more control over how their data is retained, including the option to use their own storage and encryption keys.
So: December 31, 2026 is a real date on which a privacy arrangement expires and is replaced by something not yet shipped. Eligibility is determined by your GitHub account team, and GitHub Support explicitly cannot help. If zero retention is a commitment you have made to your own customers, that conversation starts now, not in December.
Why a non-technical founder should care
You may not run an enterprise. The principle scales down anyway.
Chat logs are a business record. Whatever you type into an AI tool about your roadmap, your churn problem, your unresolved legal question, or your co-founder situation is stored somewhere with a retention policy you did not read. Ask yourself what a 24-month-old transcript of your own strategy questions is worth to someone who obtains it, and treat the answer as your threshold.
Retention is now a purchasing criterion, not a compliance footnote. If you have signed a DPA with a customer, or sell into healthcare, finance, education or the EU, “our vendor keeps chat data for the life of the account” is a sentence you may have to disclose. Check what you’ve already promised.
The safety-versus-privacy trade is going to spread. Anthropic’s reasoning is legitimate — the classifiers that catch misuse of a highly capable model need to see something to classify. Expect other labs to reach the same conclusion about their frontier models. The pattern to watch for is exactly the one here: capable model, retention required; older model, retention optional.
What to do this month
Four things, none of which takes long.
Find your retention window. For every AI tool your team pays for, answer one question: how long is my chat data kept, and can I change it? Most vendors publish this. Several bury it. If you have to search for more than ten minutes, that’s your answer about how much they want you to know.
Diary September 28. If you use Copilot on github.com or GitHub Mobile, decide before that date whether the unified experience is acceptable. Opting out means losing both surfaces entirely.
Decide the Fable 5.1 question explicitly. Not by leaving a policy toggle at its default, but by deciding whether the best coding model is worth retained prompts for your particular business. For a lot of small teams the honest answer is yes. Make it a decision anyway.
Stop pasting secrets into chat. This one predates today’s news and outlives it. See what your coding agent is allowed to touch and env files and coding agents.
The seat price didn’t move. It never does anymore. What moved was how long the record lasts — and unlike a price, that one is hard to walk back.
Sources read directly: GitHub Changelog — Upcoming changes to GitHub Copilot policies and billing (August 28, 2026), GitHub Changelog — Claude Fable 5.1 is generally available in GitHub Copilot (September 1, 2026).
Related guides
founder · 8 min read
NewResearchers put a virus in an agent's memory file. It spread to the next agent.
New Anthropic/EPFL research says AI agents can pass self-replicating instructions through files like CLAUDE.md. The fix is one paragraph.
founder · 9 min read
NewYour Agent Can Now Spend Your Money. Nobody Can Prove You Said Yes.
AWS shipped autonomous agent payments this month. The infrastructure works. The paper trail that would settle a disputed charge doesn't exist yet.
founder · 8 min read
35 Security Holes in One Month: Why Vibe-Coded Apps Are Getting Riskier in 2026
35 new CVEs in March 2026 were traced to AI-generated code. Here's what happened and what founders need to do about it.
Enjoying this guide?
Get weekly practical guides, plus tool updates and implementation playbooks.