September 15, 2026: GitHub adds Copilot Auto tiers as small reliability issues clear

Today is September 15, 2026, and GitHub’s newest visible product move is a Copilot Auto update that gives users a simple way to steer the tradeoff between cost and quality.

GitHub’s Sept. 14 changelog says Copilot Auto now has three tiers: efficiency, balance, and intelligence. The same entry says usage is billed on the model Auto actually selects, and paid subscribers get a 10% discount. In plain language, Auto is a router. It picks a model for each prompt. The tier changes the router’s bias, but billing still follows the model that was actually used, not just the label you picked. That matters because it gives GitHub a new control surface for Copilot experience and spend without making users manage every model choice manually. It also makes the product feel more configurable without requiring users to think about every underlying model switch. [1]

GitHub Docs now describe Copilot auto model selection with task optimization as generally available in Copilot Chat on GitHub.com, VS Code, Copilot CLI, and the GitHub Copilot app. But the tier controls themselves are only available in VS Code, Copilot CLI, and the app. So the feature is broader than the control surface. That distinction should matter to enterprise admins and rollout planners, because the same product capability can be exposed unevenly across clients. [2]

On the operations side, GitHub Status shows the platform as All Systems Operational as of 11:17 UTC today. That is the current baseline. But the incident feed shows two small disruptions around the edges. First, GitHub recorded a minor Copilot AI Model Providers incident on Sept. 15, tied to intermittent errors affecting Claude Fable 5.1 in some Copilot products and IDEs. GitHub says it was resolved by 11:17 UTC. Second, GitHub recorded a Sept. 14 minor Actions incident where larger runner jobs for some customers were slow to start, and that resolved at 19:35 UTC. [3] [4] [5]

The important point is not that the whole platform was down. It was not. The point is that the model-provider layer and the Actions runner layer are still capable of producing user-visible friction even when the overall status page is green. For Copilot, that is a reminder that model selection is not just a product feature. It is also an operational dependency chain. If a specific provider or model degrades, the user can feel it even if GitHub.com itself looks healthy. That makes the status page useful, but not sufficient, for understanding the actual experience developers may have on a given day. [3] [4]

There is also a security thread worth keeping in view. GitHub’s Sept. 9 changelog says npm now applies a temporary 72-hour security hold after a recovery-code sign-in for all accounts, and that the hold pauses publishing and token creation. That is a concrete hardening step against account takeover and malicious publishing. It is not proof of an active attack campaign, but it is a meaningful supply-chain safeguard because it creates a cooling-off period before an attacker can immediately use a recovered account to ship packages or mint tokens. [6]

So the day’s picture is not dramatic, but it is coherent. GitHub is making Copilot easier to steer by cost and quality, while also showing that the surrounding infrastructure still has small points of failure that can briefly affect model access or build jobs. At the same time, npm is tightening account security in a way that slows one common path to supply-chain abuse.

For developers, the practical question is simple: when Auto is available in your client, does the new tiering actually change behavior in a way you can observe on real tasks, and does the billing line match the expected model mix? For enterprise admins, the next question is whether those settings can be managed consistently across the supported clients and whether policy controls arrive in the same places as the feature itself. GitHub Docs answer part of that, but not all of it. [1] [2]

What I would watch next is straightforward. First, whether GitHub publishes any follow-up or root cause detail on the Claude Fable 5.1 Copilot provider incident. Second, whether the new Copilot Auto tiers show up on GitHub.com itself, or stay client-specific as the docs currently imply. Third, whether the Actions runner delay recurs in the next status window. And fourth, whether npm’s 72-hour hold is followed by any broader account-security changes or adoption notes. Those four signals will tell us whether today’s updates are just small product and reliability adjustments, or the start of a larger operational pattern.