There Is No GitHub Copilot X: What Copilot Actually Is in September 2026

GitHub Copilot X is not a product you can buy, and it never was. It was a March 2023 vision statement wrapped around five technical previews. GitHub's own framing at the time: "our vision for the future of AI-powered software development." The label was dropped once those previews either shipped under plain names or were killed outright.

So if you arrived here looking for Copilot X pricing or a Copilot X download, stop looking. What exists in September 2026 is a product line with four individual tiers and two organization tiers, a model picker holding more than thirty frontier models, an asynchronous agent that opens pull requests without you watching, a generally available command-line agent, an agentic code reviewer that has run 60 million reviews, and a billing system that was ripped out and rebuilt three months ago. The productivity story underneath all of that is considerably messier than the marketing.

What "Copilot X" actually was, and where each piece went

On 22 March 2023, GitHub published a post titled "GitHub Copilot X: The AI-powered developer experience." It was not a launch. It was a roadmap with waitlists attached, built around GPT-4, bundling five separate previews under one letter. Every one of them has since been renamed, absorbed, or retired.

The letter itself never appeared on an invoice.

There is a second stale name worth clearing up, because it shows up in a lot of 2024-era articles: Copilot Workspace. That was a separate GitHub Next technical preview announced in April 2024, a browser-based, task-centric environment that turned an issue into a plan and then into edits. GitHub Next closed it down in 2025. The issue-to-pull-request idea did not die with it. It came back as the Copilot cloud agent, which ships inside every paid plan today. If you are reading a review of "Copilot Workspace", you are reading about software that no longer runs.

Prices and tiers in September 2026

Six tiers, two of them for organizations. Prices below are from GitHub's own plans page and docs, in USD per month.

Read the credits column carefully, because it is doing the real work. One AI Credit equals $0.01. A Pro subscription at $10 hands back $15 of credits; Business at $19 hands back $19. The plan price is no longer a ceiling on what Copilot costs you. It is an entry fee plus a starting balance.

AI Credits: the billing change that reset everything

GitHub announced on 27 April 2026 that premium request units, the previous quota currency, would be retired. From 1 June 2026, Copilot meters GitHub AI Credits against actual token consumption at published per-model API rates. GitHub's stated reason was blunt: "Copilot is not the same product it was a year ago," and the company had been absorbing escalating inference cost.

What that means in practice:

• Free forever: inline code completions and next edit suggestions. Ghost text does not touch your balance on any plan.
• Metered: chat, agent sessions, cloud agent runs, Copilot CLI, Spaces, Spark, third-party coding agents, and the model portion of code review.
• Double-metered: Copilot code review on private repositories consumes AI Credits and GitHub Actions minutes, a change that landed the same day, 1 June 2026.
• No soft landing: the old behaviour where an exhausted quota dropped you to a cheaper fallback model is gone. When credits run out, budget policy decides what happens next: block the usage, or bill the overage.

Business and Enterprise credits pool at the billing-entity level rather than sitting in per-user buckets, which is friendlier to teams with a handful of heavy agent users and a long tail of people who mostly want autocomplete. GitHub also ran a promotional bump for existing organizations through the summer: $30 of credits for Business seats and $70 for Enterprise seats, against list prices of $19 and $39. If your invoice looked comfortable in July, look again. That window has closed.

The model picker: Copilot as a multiplexer

The single biggest structural difference between 2023 Copilot and 2026 Copilot is that GitHub stopped being an OpenAI reseller. The supported-models reference now lists more than thirty models across OpenAI's GPT-5 and GPT-6 families, Anthropic's Claude Sonnet, Haiku, Opus and Fable lines, Google's Gemini 3.x Flash releases, xAI's Grok 4.x, Moonshot's Kimi models, and Microsoft's own MAI-Code family. Availability is tiered: Free gets automatic selection only, Pro gets a picker, and the heaviest reasoning models sit behind Pro+, Max, Business and Enterprise.

The churn is relentless. In the weekly release notes for 31 August 2026 alone, GitHub added Claude Fable 5.1 for Pro+, Max, Business and Enterprise, and began rolling Gemini 3.8 Flash across every paid tier. Any article that tells you "Copilot runs on GPT-4" is describing 2023. Any article that names one specific model as the Copilot model is describing a snapshot that will be wrong within a month.

The agent layer, feature by feature

Agent mode (synchronous)

This is the one inside your editor. You pick "Agent" in Copilot Chat, and it edits files, runs terminal commands, reads test output and iterates while you watch. It is available in VS Code, Visual Studio and JetBrains IDEs. The JetBrains implementation reached general availability for custom agents, sub-agents and plan agent on 11 March 2026, which closed a long-standing gap where serious agentic work was VS Code only.

Cloud agent (asynchronous)

Formerly called the coding agent, now documented as the Copilot cloud agent. You assign it an issue or launch it from the Agents view, and it works on GitHub itself inside a GitHub Actions-powered environment: it researches the repository, writes an implementation plan, makes commits on a branch, runs tests and linters, reviews its own diff, and opens a pull request. Included in all paid plans; Business and Enterprise admins must enable the policy first.

Its limits are published, and they matter more than the capabilities:

• One repository per task. No cross-repo changes.
• One branch and one pull request per assignment.
• A hard 59-minute execution ceiling that cannot be extended.
• GitHub-hosted repositories only.
• Branch protection rules can block it unless Copilot is granted an explicit bypass.

Those constraints define the shape of work you should hand it. A well-scoped bug fix, a batch of unit tests, a dependency bump, a mechanical refactor inside one service: good. A change that spans three repositories and a schema migration: not this tool.

Copilot CLI

Generally available since 25 February 2026, for all Copilot subscribers including Free. It runs on Linux, macOS and Windows via PowerShell or WSL, in an interactive mode with a plan mode, or programmatically with a single prompt for scripting. You can define custom agents as .agent.md files that carry their own tools, MCP servers and instructions, and attach lifecycle hooks to them. preToolUse can deny or rewrite a tool call before it executes; postToolUse handles what happens afterwards. It bills by token like everything else agentic.

Code review

This is Copilot's quiet success story and the number GitHub is proudest of. In a post on 5 March 2026, GitHub reported 60 million Copilot code reviews since the April 2025 launch, roughly 10x growth, more than one in five code reviews on the platform, and over 12,000 organizations running it automatically on every pull request. It says 71% of reviews produce actionable feedback at an average of 5.1 comments, while the other 29% come back clean. GitHub treats that silence as a feature rather than a failure.

The same post is unusually candid about trade-offs: adopting a more advanced reasoning model improved positive feedback rates by 6% while review latency rose 16%. Since then, severity labels (high, medium, low) arrived in May 2026, repository-level AGENTS.md support went generally available on 18 June 2026, coverage expanded to bot-authored and very large pull requests on 27 August 2026, and on 1 September 2026 Copilot gained the ability to approve a pull request when an administrator authorises it.

Agent HQ, custom agents and AGENTS.md

At GitHub Universe on 28 October 2025, GitHub announced Agent HQ. Rather than defending Copilot as the only agent you may run, it opened the platform to competitors' agents from Anthropic, OpenAI, Google, Cognition and xAI, available inside a paid Copilot subscription. The orchestration surface, branded "mission control", spans github.com, VS Code, GitHub Mobile and the CLI, and lets you assign the same task to several agents in parallel and compare the results. For enterprises there is a control plane with agent allow-lists, model access policies, branch-level permissions and audit logs.

Behaviour is configured in source control, not in a settings panel. Copilot reads AGENTS.md at the repository root and in nested folders, plus .github/copilot-instructions.md, scoped .instructions.md files, and even CLAUDE.md and GEMINI.md. Custom agent personas live in .github/agents. If your team keeps repeating the same review comment, that is a line in a file, not a habit to nag about.

Chat surfaces beyond the editor

In August 2026 GitHub put cloud agent sessions into Microsoft Teams and Slack in public preview. Mentioning the GitHub app starts a session in the channel; anyone in the thread can add context and steer, and participants with repository write access can push it to actually make changes. Sessions consume AI Credits, with sandbox usage billed separately. VS Code picked up Agent Merge in public preview the same month: an agent that takes a pull request the rest of the way by resolving review feedback, failed checks and merge conflicts.

What changes on 28 September 2026

A changelog dated 28 August 2026 flagged three near-term changes, and they are worth putting in a calendar:

• Unification, no earlier than 28 September: Copilot Chat on github.com, Copilot Chat in GitHub Mobile and the cloud agent collapse into a single service under one policy that is enabled by default. Cloud agent moves onto Sandbox for a faster cloud experience. Chat data retention extends from 28 days to the lifetime of the account, and opting out costs you Copilot on web and mobile entirely.
• Code review default flips on 28 September: the default effort level moves from Lite to Balanced. Organizations that want Lite must select it explicitly beforehand. Since effort correlates with tokens, this is a cost change dressed as a quality change.
• Seat billing, 1 September and 1 October: new Business and Enterprise seat assignments require payment per seat, and from 1 October existing card and PayPal customers are charged upfront per seat. Removing a seat does not produce a prorated refund.

Adoption: which numbers hold up

Satya Nadella said on Microsoft's FY2025 Q4 earnings call on 30 July 2025 that there were 20 million GitHub Copilot users, and GitHub confirmed to TechCrunch that this was an all-time figure. That distinction matters: it counts anyone who ever signed in, not monthly or daily actives, and Microsoft does not publish the retention split. On the same call Nadella said 90% of the Fortune 100 use Copilot and that Enterprise customers grew 75% quarter over quarter.

Widely repeated third-party trackers put paid Copilot subscribers at roughly 4.7 million as of Microsoft's January 2026 quarterly call. I could not confirm that figure in a primary Microsoft document, so treat it as unverified. The number I would trust most as a usage signal is not a headcount at all: 60 million code reviews and one in five pull request reviews on GitHub, because GitHub published it about its own platform, where it can actually count.

The evidence, including the parts nobody puts on a landing page

What the vendor research says

GitHub's own randomized controlled trial on code quality recruited 243 developers with at least five years of Python experience, randomly assigned Copilot access, and analysed 202 valid submissions (104 with Copilot, 98 without) on a web-server task scored by ten unit tests and by blind expert review. Developers with Copilot were 53.2% more likely to pass all ten tests. Blind reviewers scored Copilot-assisted code better on readability (+3.62%), reliability (+2.94%), maintainability (+2.47%) and conciseness (+4.16%), and were 5% more likely to approve it. An earlier GitHub-affiliated trial found a 55.8% speed-up on a self-contained task. Both are real studies with published methods. Both are also vendor research on the vendor's own product, on tasks the vendor chose.

What independent research says

METR published a randomized controlled trial in July 2025 that produced the most uncomfortable result in this field. Sixteen experienced open-source developers worked 246 real issues in their own repositories, projects averaging 22,000+ stars and a million-plus lines, with AI randomly allowed or forbidden per task. They forecast a 24% speed-up. Afterwards they believed they had gained 20%. Measured, they were 19% slower with AI. Economics and machine-learning experts asked to predict the outcome were off by even more.

Be fair to the study's own caveats. The tooling was early 2025, mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet. The developers were experts in codebases they knew intimately, which is close to the worst case for AI context. The sample is sixteen people. METR itself now labels the finding historical and rewrote its experiment design in February 2026, partly because agentic tools spread so fast that recruiting AI-free control conditions became hard. The result is not "AI does not work." It is that self-reported speed-up is not evidence, and neither is your own feeling of momentum.

Google Cloud's 2025 DORA report, based on roughly 5,000 technology professionals plus 100+ hours of qualitative work, found 90% of respondents using AI at work with a median of two hours a day, a quarter of the workday. More than 80% believed their productivity had risen. Yet 30% reported trusting AI output "a little" or "not at all". DORA's framing is that AI is an amplifier: it magnifies whatever your delivery system already is, good or bad. The follow-up ROI report published in early 2026 added two concepts worth stealing. One is the J-curve, where teams dip before they gain and positive return is not guaranteed. The other is the verification tax, the ongoing cost of checking whether generated code is correct, secure and architecturally sane.

Stack Overflow's developer survey shows the same split from the other side: 84% using or planning to use AI tools, but trust in accuracy down to 29% from around 40% the previous year, and the top frustration, cited by roughly two-thirds of respondents, is output that is "almost right, but not quite". Almost-right is the expensive failure mode. It survives review.

Security and maintainability

Veracode's Spring 2026 update to its GenAI code security research is the most direct evidence that model progress has not been security progress. Across 80 coding tasks in Java, JavaScript, C# and Python, four vulnerability classes, and more than 150 models tracked longitudinally, syntax pass rates climbed above 95% while the security pass rate sat at about 55%, essentially flat since 2024. Java scored 29%. Cross-site scripting passed 15% of the time and log injection 13%, against 82% for SQL injection and 86% for insecure cryptography. Newer frontier releases stayed within the margin of error of older ones. Only reasoning-focused models reached 70–72%.

GitClear's June 2026 report, "The Maintainability Gap", analysed 623 million code changes from 2023 to 2026. Refactoring collapsed from 21% of changed lines in 2022 to 3.8% in 2026. Block duplication rose 81%, from 40.3 to 73.0 per million changed lines. Within-commit copy/paste rose 41%, error-masking constructs 47%, two-week churn 15%, while cross-file function calls, a proxy for reuse, fell 35%. GitClear's summary is the honest one: the throughput is real, and so is the debt it accrues.

None of this is an argument against using Copilot. It is an argument for where to spend the time it gives back: review, tests, security scanning, and the refactoring the tools are statistically least likely to do for you.

How Copilot compares in September 2026

The competitive picture changed more than Copilot did. Cursor's parent Anysphere was acquired by SpaceX in an all-stock deal that closed on 14 August 2026, converting Anysphere shares into 389,289,254 SpaceX Class A shares at an implied $60 billion. Cursor's pricing page now advertises Grok limits alongside frontier models. Claude Code ships inside every paid Claude plan. OpenAI's Codex is not sold separately at all; it rides inside ChatGPT tiers, and OpenAI split the old $200 Pro plan into $100 and $200 variants on 9 April 2026.

Two honest caveats about market-share claims. First, the widely circulated figures showing Copilot's share sliding while Claude Code climbs come from developer surveys with self-selected respondents, and different surveys disagree by more than ten points. Second, the framing of a single winner is wrong anyway: Agent HQ exists precisely because GitHub concluded developers will run several agents, and it would rather be the place they are governed than the loser of that fight. Copilot's structural advantage is not model quality; it rents the same models everyone else does. The advantage is that pull requests, issues, Actions, branch protection, audit logs and SSO already live on GitHub.

Picking a tier without wasting money

A few practical rules that follow directly from how credits now work:

• If you mostly want autocomplete, Free is genuinely enough at 2,000 completions a month, and completions never burn credits on any plan. Pro's $10 buys you model choice, code review and the cloud agent, not faster ghost text.
• Do not buy Max as an insurance policy. $100 for 20,000 credits only pays off if you are running long agent sessions most days. Start on Pro, watch consumption for a month, then move.
• For organizations, set budgets before you widen access. The fallback-to-cheaper-model safety net is gone; without a budget policy, an enthusiastic team can generate overage at published per-token rates.
• Check the 28 September code review default. If your repositories run Copilot review on every pull request, the shift from Lite to Balanced changes token consumption across your whole PR volume.
• Write an AGENTS.md before you write prompts. It is read by agent mode, the cloud agent, the CLI and code review, and it is the cheapest way to stop repeating yourself.
• Instrument outcomes, not vibes. METR's finding is that developers misjudged their own speed by nearly 40 percentage points. Cycle time, change failure rate and review load are measurable; feeling fast is not.

FAQ

Is GitHub Copilot X still available?

No, and it never was a separate purchasable product. "Copilot X" was a March 2023 vision statement covering five technical previews. Chat and the CLI shipped under plain names, pull request assistance became Copilot code review, the docs feature became Copilot Spaces, and Copilot Voice was cancelled in 2024. Today you buy plain GitHub Copilot in one of six tiers.

How much does GitHub Copilot cost in September 2026?

Free is $0, Pro is $10 a month, Pro+ is $39, Max is $100, Business is $19 per seat and Enterprise is $39 per seat. Each paid plan includes a monthly AI Credit allowance, from 1,500 credits on Pro up to 20,000 on Max, and one credit equals $0.01. Enterprise also requires a separate GitHub Enterprise Cloud subscription.

What are GitHub AI Credits, and do code completions consume them?

AI Credits replaced premium request units on 1 June 2026 and meter actual token usage at published per-model rates. Code completions and next edit suggestions do not consume credits on any plan. Chat, agent sessions, cloud agent runs, the CLI, Spaces, Spark, third-party agents and the model portion of code review all do, and code review on private repositories also consumes GitHub Actions minutes.

Which AI models can GitHub Copilot run?

More than thirty, from OpenAI, Anthropic, Google, xAI, Moonshot and Microsoft. Free users get automatic model selection only; Pro adds a picker; the heaviest reasoning and Opus-class models require Pro+, Max, Business or Enterprise. The list changes almost weekly, so check GitHub's supported-models reference rather than any article, including this one.

Does GitHub Copilot actually make developers faster?

It depends heavily on the task, and the honest answer is contested. GitHub's own trial found developers 53.2% more likely to pass all unit tests on a scoped task, and an earlier study found a 55.8% speed-up. METR's randomized trial of sixteen experienced maintainers working real issues in codebases they knew well found them 19% slower with AI, even though they believed they were 20% faster. Constrained, well-specified work benefits most; deep work in a familiar large codebase benefits least.

Copilot, Cursor, Claude Code or Codex — which should you use?

Copilot if your workflow is organised around GitHub pull requests, if you need org-level policy and audit logs, or if you work in JetBrains, Xcode or Eclipse. Cursor if you want a single opinionated editor and its in-house Composer models, now under SpaceX ownership. Claude Code if you live in a terminal and already pay for Claude. Codex if you already pay for ChatGPT and would rather not add a bill. Many developers run more than one, which is exactly why GitHub built Agent HQ to host rivals' agents.

Sources

• GitHub Copilot X: The AI-powered developer experience (GitHub Blog, 22 March 2023)
• GitHub Copilot · Plans & pricing
• Plans for GitHub Copilot (GitHub Docs)
• GitHub Copilot is moving to usage-based billing (GitHub Blog, 27 April 2026)
• Supported AI models in GitHub Copilot (GitHub Docs)
• About Copilot cloud agent (GitHub Docs)
• About GitHub Copilot CLI (GitHub Docs)
• GitHub Copilot CLI is now generally available (Changelog, 25 February 2026)
• 60 million Copilot code reviews and counting (GitHub Blog, 5 March 2026)
• Copilot code review can now approve pull requests (Changelog, 1 September 2026)
• Introducing Agent HQ: Any agent, any way you work (GitHub Blog, 28 October 2025)
• Upcoming changes to GitHub Copilot policies and billing (Changelog, 28 August 2026)
• GitHub Copilot weekly releases, August 31 (Changelog)
• Major agentic capabilities improvements in Copilot for JetBrains (Changelog, 11 March 2026)
• Sunset notice: Copilot knowledge bases (Changelog, 20 August 2025)
• Does GitHub Copilot improve code quality? (GitHub Blog research)
• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (METR)
• We are changing our developer productivity experiment design (METR, 24 February 2026)
• DORA: State of AI-assisted Software Development 2025
• Spring 2026 GenAI Code Security Update (Veracode)
• The Maintainability Gap: 2026 AI Code Quality Research (GitClear)
• Mind the gap: Closing the AI trust gap for developers (Stack Overflow, 18 February 2026)
• GitHub Copilot crosses 20M all-time users (TechCrunch, 30 July 2025)
• Cursor pricing
• Claude pricing
• SpaceX to acquire Cursor for $60 billion (CNBC, 16 June 2026)
• GitHub kills 'Copilot Voice' tool (Visual Studio Magazine, March 2024)