GitHub Copilot — The Start of AI-Assisted Coding

On 29 June 2021, GitHub released GitHub Copilot—an AI code-completion product built jointly with OpenAI—as a technical preview. It sat in the sidebar of VS Code and could write entire function bodies from a comment or a signature. For the first time, mainstream developers were handing the wheel of their editor to an AI.

On 21 June 2022 it became generally available at US$10/month or US$100/year, free for verified students and maintainers of popular open-source projects. Within a few years the industry's development experience had been rebuilt in a way that could no longer be undone.

Codex Under the Hood

Behind Copilot ran OpenAI's Codex. The paper that introduced it — "Evaluating Large Language Models Trained on Code," 7 July 2021 — defines it as "a GPT language model fine-tuned on publicly available code from GitHub." On the paper's own HumanEval benchmark Codex solved 28.8% of problems in a single attempt where plain GPT-3 solved 0% (70.2% if allowed 100 samples per problem). In the narrow but valuable task of predicting the next few dozen lines of code, Codex was the clear winner of its moment.

IntelliSense and TabNine had offered completion before, but no shipping product had been willing to write whole functions in one shot. Copilot was the first mass-market product to cross that line.

Doe v. GitHub — Collision with Open Source

On 3 November 2022 a group of anonymous developers filed a class action in the Northern District of California — before Judge Jon S. Tigar — against GitHub, Microsoft, and OpenAI. Counsel: Matthew Butterick with the Joseph Saveri Law Firm. Two core claims.

1. Open-source licence breach. MIT, GPL, and Apache-licensed code on GitHub had been used as training data and was being regenerated by Copilot without attribution or licence inheritance—a breach of the licence contracts.

2. DMCA §1202 violation. Stripping Copyright Management Information (CMI)—author names, licence notices—from outputs is itself unlawful.

The case was pared back three times. On 11 May 2023 Judge Tigar dismissed most of the 22 claims with leave to amend, leaving DMCA §1202(b)(1) and (b)(3) and the open-source licence breach. On 3 January 2024 the state-law claims (unjust enrichment, negligence, unfair competition and others) fell as preempted by federal copyright law, and the §1202(b) theory was rejected on the reading that "the alleged DMCA violations can only occur when the copied works are identical" — "near-identical" and "modified" were not enough. Finally, in June 2024 (reported 8–9 July) the remaining DMCA §1202(b) claim was dismissed with prejudice; Tigar found unpersuasive the plaintiffs' reliance on a study showing only that Copilot "may theoretically be prompted by a user to generate a match to someone else's code."

What survives is two claims: open-source licence violation and breach of contract. No court has ruled that Copilot is unlawful, but the case became the first formal challenge to the unspoken premise that "public code = free training material".

The Copilot suit was the precursor to New York Times v. OpenAI (2023) and the wave of suits against generative-image models.

Productivity — Belief and Doubt

The much-quoted "55% faster" is GitHub's own number. GitHub Next and the Microsoft Office of the Chief Economist ran a 2022 controlled experiment in which developers implemented an HTTP server in JavaScript; the blog says 55%, the accompanying paper by Peng, Kalliamvakou, Cihon and Demirer says 55.8% (n=95, p=.0017), with mean completion times of 1h11m against 2h41m. The ~30% suggestion-acceptance rate is likewise GitHub's own figure, published with its 2024 Accenture deployment study. Neither should be read with the vendor stripped off.

The more independent evidence is Cui et al. (2025, Management Science), which pools three randomised controlled trials run at Microsoft, Accenture and an anonymous Fortune 100 company. Across 4,867 developers it finds a 26.08% increase in completed tasks (SE 10.3%), with less experienced developers adopting more and gaining more. Two of the six authors work at Microsoft.

Pointing the other way: in July 2025 the non-profit lab METR published a randomised controlled trial in which 16 experienced open-source developers worked on real issues in their own repositories. They forecast a 24% speedup beforehand and still believed they had been sped up 20% afterwards — while measuring 19% slower. The tools were mostly Cursor Pro with Claude 3.5/3.7 Sonnet, not Copilot. The gulf between perceived and actual speedup showed that tool choice and workflow design mattered far more than the marketing suggested.

The anxiety that "junior developers will be wiped out" runs in parallel. If completion can fill in more and more code, the rungs on which a beginner climbs to expertise get sawed off. From 2024 onward that worry has fused with a cooling hiring market into an unspoken industry-wide unease.

Cursor, Windsurf, Claude Code — The Insurgents Catch Up

2024 was the year AI coding IDEs exploded. Anysphere's Cursor forked VS Code and added multi-file editing and the "Composer" agent. Its annualised revenue passed US$100 million in January 2025 and US$500 million in June; on 13 November 2025, announcing a US$2.3 billion raise at a US$29.3 billion post-money valuation, the company said it had crossed US$1 billion in annualised revenue — its own figure, not an audited one. Codeium renamed itself Windsurf in April 2025. Sourcegraph Cody discontinued its Free and Pro tiers on 23 July 2025 to focus on enterprise. Anthropic shipped Claude Code as a research preview on 24 February 2025 and made it generally available on 22 May 2025.

GitHub went on the defensive. On 18 December 2024 it shipped Copilot Free — 2,000 code completions and 50 chat messages a month. The individual plan has been US$10/month since GA; during 2025 it was folded under the name "Copilot Pro," and on 4 April 2025 GitHub added Copilot Pro+ at US$39/month with 1,500 premium requests. It also rolled out Copilot Workspace (spec → implementation → PR end-to-end) and Copilot Agents (autonomous task execution). On 1 June 2026 every plan moved to usage-based billing: premium request units gave way to AI Credits drawn down by token consumption, with Pro at $10/month including $10 of credits and Pro+ at $39 including $39. Flat-rate completion pricing had stopped fitting a product whose agents burn tokens.

A Whole Industry Pivots

Five years after Copilot's preview, nearly every part of the software industry has bet on AI dev tools. The 2025 Stack Overflow Developer Survey: 84% of developers use or plan to use AI tools (76% in 2024); 51% of professionals use them daily; but only 29% trust the accuracy of the output, down 11 points year on year, while 46% actively distrust it, up from 31%. Adoption at a record high, trust at a record low. "I use it but I don't believe it" is the honest median.

Copilot will be remembered less as a single product than as the product that rewrote the industry's defaults. Having an AI live inside the editor became normal, and the ripples it set off across licensing, productivity, and the labour market are still moving.