Copilot drastically cuts model permissions, developers are furious

GitHub really irritated developers this time.

Over the past week, GitHub Copilot made a round of adjustments to its individual subscription plans (Pro and Pro+) that can only be described as “cutting off the lifeline”: it removed access to premium models such as Claude Opus, tightened the free trial entry, and imposed stricter limits on usage. As soon as the news broke, hundreds of angry comments flooded the GitHub community announcement thread, while the developer forum LINUX DO saw multiple posts from “Copilot refugees” looking for help—a small‑scale wave of cancellations is underway.

What happened

In mid‑April, GitHub officially released an announcement titled “Changes to GitHub Copilot Individual Plans.” The key changes include:

• Pro subscriptions no longer support selecting top‑tier third‑party models like Claude Opus. Previously, users could freely switch between GPT‑4o and Claude Opus in Copilot Chat; now those options are drastically reduced.
• Pro+ subscriptions are also affected: the Opus model’s usage multiplier was raised to 7.5×, meaning the same allowance now yields far fewer calls.
• Free trials are fully suspended, officially due to “abuse issues.”
• The student plan (Copilot for Education) was adjusted in parallel and no longer allows manual selection of top‑tier models.

In short, GitHub either cut or effectively raised the price of the most valuable models inside Copilot through a multiplier system. For developers attracted by “$10 per month access to Opus,” this is essentially a unilateral breach of contract.

Developers’ anger is real

On the LINUX DO forum, several related posts quickly gained large numbers of replies and interactions. One user titled their post bluntly: “GitHub Copilot Pro subscription is dead—moving to Zhipu GLM,” writing:

This morning felt like a thunderbolt out of the blue—GitHub Copilot Pro is completely ruined. I’d even pay extra, but now they won’t let us use the models at all. The GitHub community is full of complaints—there are already over 90 comments below.

Another user was more radical, posting a screenshot of their PayPal complaint page and saying they had backed up all repositories locally and were preparing to leave the GitHub ecosystem entirely:

GitHub basically said: can’t stand it? Then leave. Fine—I’m leaving. Ticket submitted; refund should come soon.

Other users mentioned that GitHub has already opened a refund channel—which itself is telling. Platforms usually simplify refunds only when they anticipate large‑scale cancellations. It’s better to handle refunds gracefully themselves than to force users into PayPal disputes or credit‑card chargebacks.

This sentiment is not unreasonable. Many developers chose Copilot Pro/Pro+ because it offered a relatively cheap way to access Claude Opus and GPT‑4o—top‑tier models—for $10–$39 per month. That subscription meant being able to call these models directly within the IDE for code completion, refactoring, explanation, and generation. Now that the models are gone, the subscription’s core value has vanished.

Why GitHub did this

The answer is simple: losing money.

Large‑model inference costs are real. A model at the level of Claude Opus has a per‑token inference cost far higher than GPT‑4o‑mini or Claude Haiku. When GitHub offered a fixed monthly fee for “unlimited model choice,” savvy developers naturally concentrated usage on the most expensive models—like everyone in a buffet eating only lobster.

One forum user put it neatly:

Opus now has a 7.5 multiplier—counting by usage really is too unprofitable under current conditions. Pro+ is practically useless now too.

From a business perspective, GitHub’s move is understandable. Copilot is a key part of Microsoft’s AI strategy, but it cannot subsidize users indefinitely. The issue is execution—no sufficient transition period, no reasonable alternatives, and poor communication. “If you can’t stand it, leave” may or may not be an exact quote from support, but it captures the attitude users felt.

It also exposes a deeper problem: when developers tightly bind their workflow to a platform’s AI capability, unilateral policy changes pose huge risk. Today it’s Opus being cut; tomorrow it could be GPT‑4o; the next day, a price hike. Your productivity tools shouldn’t be built on ground someone else can pull away at any moment.

Where the “refugees” went

After canceling, developers need replacements. Forum discussions show three broad directions:

1. Domestic large models: Zhipu GLM as a popular choice

Several users mentioned the Zhipu GLM model family. One said, “Heard GLM is catching up with Opus—going to try it.” The GLM‑4 series has indeed improved markedly over the past year in code generation and logical reasoning, especially for Chinese‑language programming tasks. Its comprehension and generation quality now meet most daily development needs.

Zhipu’s pricing strategy is also friendlier. Compared with Copilot Pro+ at $39 per month, GLM API calls cost much less and avoid the “sudden model withdrawal” risk—what you call is directly the model you choose, without middle‑man policy flips.

2. Open‑source models + local deployment

Copilot’s substitute doesn’t have to be another cloud service. For developers with sufficient hardware, locally deploying open‑source models is increasingly feasible. Open‑source code models such as DeepSeek Coder V3, Qwen2.5‑Coder, and CodeLlama already perform quite well. Combined with open‑source IDE plugins like Continue and Tabby, you can build a fully self‑controlled code‑assist toolchain.

The benefits are obvious: no subscription fees, no platform risk, data stays local. The downsides: needing GPU resources, manual deployment and tuning, and a performance ceiling still below that of Opus‑level proprietary models for now.

But the trend is clear. Open‑source code models are improving fast—the gap that was obvious six months ago has already shrunk to “good enough for most scenarios.”

3. Other CLI tools and plugins

The forum also saw many recommendations for non‑Copilot code‑assist tools. Some suggested “using other CLIs or plugins,” demoting Copilot to a fallback rather than a primary tool. New‑generation AI programming tools such as Cursor, Windsurf, and Cline each have their strengths. Their pricing may not be cheaper, but they offer better flexibility in model choice.

An interesting pattern: many users adopt a “multi‑legged approach”—using a locally deployed open‑source model for daily autocompletion and switching to a cloud top‑tier model for complex tasks. This hybrid setup balances cost and capability well.

The broader context

Zooming out, Copilot’s adjustments reflect the AI programming‑tool market entering a “bubble‑squeezing” phase.

Over the past two years, platforms raced to capture developer mindshare by offering top models at low or even loss‑making prices. GitHub’s $10‑per‑month Opus access, Cursor’s $20‑per‑month for 500 premium model calls—all commercially unsustainable. Once subsidies end and normal economics resume, price hikes or feature cuts become inevitable.

For developers, this prompts serious reflection: how should your AI‑assisted workflow be built to stay stable when platforms change policy?

A few key principles:

• Don’t put all eggs in one basket. Have both primary and backup tools; minimize switching cost.
• Track open‑source progress. Code generation is one of the fastest‑moving open‑source domains—today’s “not good enough” model might be sufficient in three months.
• Assess local‑deployment feasibility. If your team has spare GPU resources, deploying a local code model may have higher ROI than you expect.
• Distinguish “nice‑to‑have” from “mission‑critical.” If your workflow stops without a particular AI tool, your dependency level may be too high.

The opportunity window for domestic models

From an industry‑competition standpoint, Copilot’s misstep opens a rare opportunity for domestic models.

Previously, when Chinese developers chose AI coding tools, Copilot was the default—its deep integration with VS Code and GitHub was hard to rival. But now that this default choice falters and users reevaluate alternatives, domestic models gain serious consideration.

Zhipu GLM, DeepSeek, Tongyi Qianwen and other domestic models have made real progress in coding ability. More importantly, they have natural advantages for developers working in China—better network access, stronger Chinese comprehension, and more reasonable pricing. A responsive, accurate, cost‑effective local model can be more reliable than an overseas service prone to outages or sudden policy changes.

Of course, to truly absorb these “refugees,” domestic models still need stronger tooling integration. A good model API alone isn’t enough—developers also need solid IDE plugins, smooth workflow experiences, and stable service quality. China’s AI programming‑tool ecosystem still has room to improve here.

If you’re looking for a multi‑model solution, API aggregation platforms like OpenAI Hub provide a useful path—invoking GPT, Claude, GLM, DeepSeek and more through one unified interface format. No separate integration per model, low switching cost. In today’s environment—where any platform might change overnight—flexibility matters more than loyalty to any single provider.

Final thoughts

The most ironic part of this Copilot incident is that GitHub—the world’s largest open‑source community—has lost developer trust through closed and opaque commercial decisions. The users pushed away may end up discovering more open, flexible alternatives.

The market has never lacked choices; what it lacked was motivation to compare them seriously. This time, GitHub may have done domestic models and the open‑source community a favor.

---

Sources:

• Discussion of GitHub Copilot Pro subscription changes — LINUX DO forum users sharing experiences with Copilot Pro changes and switching to GLM
• Copilot refund discussion thread — LINUX DO forum users discussing Copilot refunds and cancellation experiences
• Copilot refugee replacement‑solution discussion — LINUX DO forum thread on alternative tools and workflows for displaced Copilot users