GitHub pauses Copilot Pro trials and tightens limits as providers grapple with demand

For all the utility of AI coding tools, software developers face a growing problem: usage limits and costs are really getting in the way. And that’s starting to show in how the major AI labs are tightening caps, tweaking plans, and pulling back access to key features and tiers. Behind that is a simple issue: heavier, longer-running usage is pushing these systems far beyond what earlier subscription models were designed to handle.

Last week, GitHub announced that it was pausing new Copilot Pro trials, which typically last 30 days, citing a “significant rise in abuse.” At the same time, the Microsoft subsidiary revealed that it was tightening usage limits and retiring the Opus 4.6 Fast model from its Copilot Pro+ plan, with users instead directed to the standard Opus 4.6 model or the Copilot Cloud Agent for similar capabilities. The same intelligence, just with less capacity for heavy or parallel workloads.

GitHub says usage has grown heavy, with high levels of concurrent demand putting pressure on its systems. Anthropic, meanwhile, has also been adjusting its own usage caps in recent weeks, while simultaneously restricting how Claude can be used with third-party tools such as OpenClaw, pushing that usage onto separate, pay-as-you-go pricing.

“We’ve been working hard to meet the increase in demand for Claude, and our subscriptions weren't built for the usage patterns of these third-party tools,” Anthropic’s Claude Code chief Boris Cherny wrote on X. “Capacity is a resource we manage thoughtfully and we are prioritizing our customers using our products and API.”

GitHub pauses for thought

In its initial guise, Copilot was really all about speeding up smaller coding tasks, like suggesting the next few lines of code – the kind of usage that tends to be short-lived and fairly predictable.

What’s happening now is much broader. Developers are keeping sessions open for longer, running repeated tasks, and using these systems to work through pull requests, test failures, and other loops that can run continuously. A single user can generate far more load than before, particularly when multiple processes are running at once.

GitHub’s response reflects that change. Pausing free trials reduces the flow of new users, while tighter limits and the removal of a higher-throughput model reduce the pressure from those already using the service. Indeed, models designed for higher throughput tend to be the ones that get hit hardest when demand spikes, especially when users are running longer or repeated tasks.

It’s worth noting that GitHub says trials will return once additional checks are in place, while Copilot remains available via a free plan, albeit with tighter usage limits than the Pro tier.

“This is a temporary pause,” the company wrote. “We are actively working on improved safeguards to prevent misuse of the trial system, and we will reopen trials once those protections are in place.”

Open and local AI stacks gain ground

While these various changes have a more direct impact on individual users, they also help explain why some of the companies building on top of these models are starting to reduce their reliance on them.

Platforms such as Cursor and Windsurf have been moving toward more home-grown models with tighter control over how they run, rather than relying entirely on external providers. That includes tuning their own systems for longer-running coding tasks and managing costs more directly.

And much of this, ultimately, is dependent on open-weight models, which give companies more flexibility over how and where they are deployed.

At the same time, there is growing interest in running more of this locally. Ollama allows developers to run models on their own machines rather than relying entirely on cloud services. And open-source projects such as OpenClaw have also emerged, aiming to replicate parts of the agent-style experience in a more self-contained setup. In fact, Ollama and OpenClaw make natural bedfellows for the local AI stack, with one handling model serving and the other orchestrating tasks on top.

However, that kind of agent-style usage is also a big part of what’s driving these changes. Tools like OpenClaw can run continuous, automated loops that go far beyond a typical developer session, placing much heavier demands on underlying systems. In effect, it creates a feedback loop of sorts, where more capable tools drive heavier usage, prompting providers to tighten limits, which in turn pushes developers to look for alternatives that give them more control over how these tools are run and scaled.

One response has been a growing interest in open coding platforms such as OpenCode and Kilo Code, which give developers more scope to build and run AI tooling on their own terms, with the freedom to choose and swap models, rather than being tied to the limits, pricing, or policies of the major platforms.

That all said, most developers still rely on the largest providers — and for good reason.

OpenAI, Anthropic, Google and others are leading the AI coding charge, with the strongest models, the deepest tooling, and the infrastructure to support them. But recent moves to cap usage, adjust pricing, and restrict access show that building on these platforms means living with whatever limits, pricing, and access those companies decide to impose.