GitHub Copilot Review

A lot has changed since Copilot was just an inline autocomplete tool. Halfway through 2026, it is an agentic platform with autonomous coding agents, cloud-based background workers, a CLI, custom agent configs, and third-party agents from Anthropic and OpenAI. But it has also picked up baggage: a privacy update that turned individual users into default training-data contributors, a billing overhaul replacing simple request counts with token-based credits, and the removal of premium models from the Student Developer Pack.

This review captures what Copilot looks like in mid-May 2026. Everything here is verified against GitHub’s official docs, changelog, and plan pages.

The State of Copilot in 2026

Copilot is available as a plugin for VS Code, Visual Studio 2026, JetBrains IDEs, Xcode, Neovim, and Eclipse. It also lives on github.com as Copilot Chat, as a cloud agent in GitHub Actions, and as a standalone CLI. Third-party agents - Claude by Anthropic and Codex by OpenAI - are assignable alongside Copilot’s own coding agent. No other AI coding tool covers this many surfaces under one subscription.

The feature set has four pillars: inline completions, Copilot Chat, agent mode, and cloud agents. Agent mode is the big one - it reads your codebase, plans multi-file changes, edits files, runs terminal commands, and iterates until it thinks the task is done. Cloud agents do the same unattended in a GitHub Actions sandbox and open a PR when they finish.

Agent mode hit GA on VS Code and JetBrains in March 2026. The April Visual Studio update added cloud agent sessions that launch directly from the IDE, expanded MCP support for custom agents, and gave enterprises per-organization access controls. Code review comment grouping landed May 12, making AI review feedback less noisy on large PRs.

Plans and Pricing

Free ($0): 2,000 completions and 50 premium requests per month. Lighter models (Haiku 4.5, GPT-5 mini), plus Copilot CLI. Genuine value for casual open-source work.

Pro ($10/month): Unlimited completions, 300 premium requests, access to Claude Sonnet 4.6, Claude Opus 4.6, GPT-5.2-Codex, and cloud agent. The workhorse plan. Annual billing ($100/year) grandfathered for existing subscribers.

Pro+ ($39/month): Everything in Pro plus 1,500 premium requests (5x), priority access during peak times, all models. Additional requests cost $0.04 each.

Business ($19/user/month): Pro features plus org-wide policy controls, IP indemnification, centralized billing.

Enterprise ($39/user/month): Business features plus custom model fine-tuning (preview), metrics dashboards, knowledge bases for grounding responses in internal docs, dedicated support.

The June 1 Billing Shift

On April 27, 2026, GitHub announced that all Copilot plans move to usage-based billing on June 1. Instead of counting individual requests, GitHub AI Credits now measure consumption - one credit equals $0.01. Pro gets $10 in monthly credits, Pro+ gets $39. If you run out mid-month, you wait for the reset or pay overages.

Token consumption varies by model. A quick chat with a lightweight model costs fractions of a cent. An active agent, by GitHub’s estimates, burns roughly $1 per minute. A Pro+ subscriber with $39 in credits gets about 39 minutes of continuous agent mode before hitting the cap. Code completions remain unlimited on all paid plans.

For organizations, there is a transition window (June 1–September 1): 3,000 credits/user/month for Business, 7,000 for Enterprise. After September 1, billing is fully consumption-based with no fixed allowance.

The change is fair in principle - a one-line chat and an hour-long agent session should not cost the same. But it introduces cost unpredictability where none existed. You now think about credit burn rates, which is the opposite of the “set it and forget it” experience that made Pro popular.

The Privacy Policy Shift

On March 25, 2026, GitHub updated its interaction data policy. Starting April 24, data from Free, Pro, and Pro+ users - prompts, completions, acceptances, rejections, and context - is used by default to train GitHub’s AI models. Users must manually opt out. Business and Enterprise data is excluded.

The rollout was a banner on github.com, a blog post, and a policy update. Many developers missed it until news sites and Reddit picked it up. A community thread with thousands of replies made clear developers wanted opt-in, not opt-out. If you do not want your interactions training models, go to your Copilot settings and disable the toggle - most people still do not know it is there.

Model Selection

Copilot’s model lineup has expanded significantly. As of May 2026:

• Pro and above: Claude Sonnet 4.6 (GA since February 17, 2026), Claude Opus 4.6, Claude Sonnet 4.5, GPT-5.2-Codex, and others from OpenAI and Anthropic.
• Free: Haiku 4.5, GPT-5 mini, and a limited set of lighter models.
• Students (removed March 2026): GPT-5.4, Claude Opus, and Claude Sonnet were pulled from the Student Developer Pack. GitHub said this was to sustainably keep Copilot free for students globally, but the move generated sustained community backlash.
• Deprecations: Claude Sonnet 3.5 retired November 6, 2025. A January 13, 2026 changelog announced additional deprecations for select Anthropic and OpenAI models.

One persistent issue: model availability is inconsistent across surfaces. Users report certain Claude models appearing in Copilot Chat on github.com but not inside Visual Studio 2026, or vice versa. GitHub says it is working toward parity but has not delivered it yet.

Agent Mode, Cloud Agents, and Custom Agents

Agent Mode (GA, March 2026): Give it a task - “refactor the auth module to support OAuth2” - and it reads files, plans changes, edits across files, runs tests, checks errors, and iterates. You approve changes along the way. It has matured substantially since its February 2025 preview.

Cloud Agent (GA): Runs agent mode in a GitHub Actions sandbox, triggered by assigning a GitHub issue to Copilot. It researches the repo, creates a plan, makes changes on a branch, and opens a PR. Starting June 1, cloud agent runs consume Actions minutes on top of AI credits. The April 27 update made startups 20% faster with optimized runner images.

Custom Agents (expanding): Configured with .agent.md files under .github/agents/. Define specialized behaviors, system prompts, tool access, and MCP server connections. First shipped in Visual Studio 2026 in March, expanded with MCP support in April, gradually rolling to VS Code and github.com.

Third-Party Agents: Claude by Anthropic and Codex by OpenAI are assignable alongside Copilot’s coding agent on github.com. Model selection for third-party agents launched April 14, 2026.

Copilot CLI

Copilot CLI reached GA on February 25, 2026. Integrated with the GitHub CLI, it handles command explanations, script generation, build-error debugging, and agent sessions from the terminal. The January 2026 update added full model selection parity with the IDE, /memory commands for context management, and Autopilot mode that chains steps without confirmation. For terminal-native developers, it is a credible alternative to Claude Code - though Claude Code leads on deep agentic work (80.8% vs 56% on SWE-bench).

Copilot Workspace

Copilot Workspace, the experimental browser-based environment from GitHub Next, was shelved in November 2024. The roadmap issue confirmed the team took lessons from the public preview and redirected focus to the Copilot coding agent. The coding agent and cloud agent are its spiritual successors.

Code Review and Vision

Code Review: Automatically comments on PRs with AI-generated feedback. Now groups similar comments together (May 12, 2026 update) to reduce noise. A premium feature available on Pro and above. Starting June 1, code review runs also consume GitHub Actions minutes.

Vision: Copilot Chat supports image attachments. Paste a screenshot of a UI mockup and ask for the component. JPEG, PNG, GIF, WEBP supported. Availability depends on the selected model - GPT-4o supports it, some older models do not. Vision in Visual Studio 2026 has been inconsistent, with users reporting the feature appearing and disappearing across updates.

Copilot vs Cursor vs Claude Code

Copilot vs Cursor: Copilot scores 56% on SWE-bench to Cursor’s 51.7%, but Cursor completes tasks 30% faster. Cursor wins on multi-file context and the AI-first editing experience. Copilot wins on IDE coverage (six vs one), GitHub integration, PR reviews, and price ($10 vs $20/month). Cursor edges ahead for inline completions in a VS Code fork; Copilot fits multi-IDE workflows and GitHub-native teams.

Copilot vs Claude Code: Claude Code is a terminal-native agentic tool with 80.8% SWE-bench accuracy, parallel agent teams, and deep reasoning. Stronger for complex multi-file tasks. Copilot is stronger for IDE work, PR reviews, and team management. Claude Code charges by API usage (can get expensive); Copilot is a fixed subscription with a free tier. For heavy terminal-based agentic work, Claude Code wins. For everyday IDE development, Copilot is more practical.

What Still Needs Work

Billing transparency. GitHub needs a real-time credit tracker in the IDE. Watching credits tick down during an agent session will be stressful without it.

Model consistency. Different models available in different places is a recurring frustration that has not been resolved.

Agent guardrails. Agent mode generates plausible but wrong refactors. The self-review feature helps but is no substitute for human review.

Student model removal. Cutting premium models from the Student Developer Pack without a replacement tier saved money but burned goodwill.

Privacy defaults. Making data training opt-out rather than opt-in was the wrong call. GitHub should let users choose to contribute.

Verdict

GitHub Copilot in mid-2026 is the most broadly available AI coding platform out there. It works in more IDEs, supports more models, and offers more deployment surfaces (IDE, browser, CLI, cloud) than any competitor. The agent features shipping in 2026 make it meaningfully more capable than it was a year ago.

But the June 1 billing change and the April 24 data policy shift have introduced friction that did not exist before. Copilot is no longer “just turn it on and forget about it.” You now need to think about credit consumption, overage risk, and whether your interaction data is training the next generation of models.

For individual developers on VS Code or JetBrains, Pro at $10/month remains good value - if you do not run agent mode continuously. Heavy agent users land on Pro+ at $39/month and may still hit the cap. For teams, Business at $19/user/month with IP indemnification is the safe enterprise pick. For everyone: check your privacy settings.

Copilot is still a strong recommendation. It just demands more attention than it used to.