GitHub Copilot Review 2026: Four Tiers, 1M Context, and the New Credits Math

The flat "$10 a month, unlimited" Copilot is gone. As of June 1, 2026, every Copilot plan runs on GitHub AI Credits, billed against your actual token use, and there are four individual tiers instead of two (GitHub Copilot plans, checked 2026-06-27). The new question is not whether Copilot is worth $10. It is which tier's credit budget survives the way you actually code, because the two headline features that shipped this month, a one-million-token context window and configurable reasoning, both burn credits faster than the defaults do.

Here is the one thing to leave with. Code completions and next edit suggestions are still unlimited and free of credit draw on every paid plan. Everything agentic (chat, the CLI, cloud agent, third-party coding agents) now spends credits, and 1M context plus high reasoning spends them quickly. Pro at $10/mo includes 1,500 credits, which is fine for autocomplete-plus-occasional-chat but tight for daily agent work. Pro+ at $39/mo, with 7,000 credits and premium models including Opus, is the real working tier for anyone living in agent mode (GitHub Copilot plans, 2026-06-27). We rate Copilot 4.2 out of 5 in June 2026, up from 4.1, and below is exactly why.

What changed since the last review

The previous review (rating 4.1, last updated May 1) predated all three of these. Each is dated and sourced.

• Usage-based billing went live June 1, 2026. Premium request units are gone. Every plan now includes a monthly allowance of GitHub AI Credits, consumed by token use across input, output, and cached tokens at each model's published rate (GitHub blog, billing transition). The conversion is fixed: 1 AI credit equals $0.01 USD (GitHub docs, usage-based billing).
• A 1M-token context window shipped June 4, 2026. It is live in VS Code, Copilot CLI, and the Copilot app, letting you work across large codebases and long documents without losing context. GitHub's own note: a larger context window or higher reasoning level "will consume more AI credits per interaction" (GitHub changelog, June 4 2026).
• Configurable reasoning levels arrived the same day. You can dial speed against depth and unlock extended thinking for hard architectural or debugging problems, on the same three surfaces, with the same credit caveat (GitHub changelog, June 4 2026).
• Claude Fable 5 came and went in three days. It went generally available in the Copilot model picker on June 9, then an editor's note dated June 12 confirmed access was "suspended across all GitHub Copilot experiences" following Anthropic's announcement, with all other Claude models (Opus 4.8, Sonnet 4.6, Haiku 4.5) unaffected (GitHub changelog, Fable 5). The trigger was an export-control order after a jailbreak. More on what that signals below.

One more practical note from the same plans page: new Pro, Pro+, and Max sign-ups are "temporarily paused as we ensure a high-quality experience" (GitHub Copilot plans, 2026-06-27). Existing Pro and Student users can still upgrade to Pro+ and Max; brand-new paid sign-ups may have to wait. Plan around it.

The four tiers and who each one fits

Prices and credit allowances are current per GitHub Copilot plans and the billing docs, both pulled 2026-06-27. The total-credits column splits into base credits (fixed, matched to your subscription) plus a flex allotment (a variable top-up GitHub adjusts as model economics shift).

Plan	Price/mo	Base + flex = total credits	Best for	Where you hit the ceiling
Free	$0	2,000 completions, auto model only	Trying it, light autocomplete	No paid models, no model picker, chat is rationed
Pro	$10	1,000 + 500 = 1,500	Autocomplete plus occasional chat	A few heavy agent sessions burn the month's credits
Pro+	$39	3,900 + 3,100 = 7,000	Daily agent work, premium models, audit logs	Sustained 1M-context days still drain it by month-end
Max	$100	10,000 + 10,000 = 20,000	High-volume, all-day agent workflows	Designed not to, but overage may be capped
Source	plans	billing docs	pulled 2026-06-27	pulled 2026-06-27

Free is a genuine on-ramp, not a demo. You get 2,000 completions a month plus access to Haiku 4.5 and GPT-5 mini through auto model selection (GitHub Copilot plans, 2026-06-27). What you do not get is the model picker or any premium model, so it is autocomplete and light chat, full stop.

Pro at $10/mo keeps unlimited code completion and next edit suggestions, adds model selection, third-party agents (Claude Code and Codex), and the cloud agent, then meters everything agentic against 1,500 credits (GitHub Copilot plans, 2026-06-27). For a developer who mostly accepts inline suggestions and asks Copilot Chat a handful of questions a day, that budget is comfortable. Turn on 1M context or extended reasoning as a habit and it stops being comfortable fast.

Pro+ at $39/mo is where the product gets serious: premium models including Opus, audit logs, and 7,000 credits, described by GitHub as "4x+ included usage than Pro" (GitHub Copilot plans, 2026-06-27). This is the tier for someone in agent mode most of the day. Max at $100/mo carries 20,000 credits and "2.9x+ included usage than Pro+," aimed at sustained, high-volume agent runs where even Pro+ would run dry.

The 1M context window in practice

This is the real capability gain since May, and it earns the rating bump. A one-million-token window means an agent can hold an entire large codebase, a long spec, and a multi-file change in working memory without you hand-feeding it the relevant files (GitHub changelog, June 4 2026). For whole-repo refactors and "why does this break across these eight files" debugging, that is a different class of task than what fit before.

The catch is in the billing. Credits are consumed on token use: input, output, and cached tokens, priced per model (GitHub docs). A 1M-context interaction sends vastly more input tokens than a default-window one, so the per-interaction cost climbs with the window you choose. GitHub says as much and recommends the default window for everyday tasks, reaching for extended context only on genuinely complex multi-file problems (GitHub changelog, June 4 2026). That advice is also a budgeting instruction. Leave 1M context on by default and you are paying frontier-model input rates on every keystroke-adjacent chat.

To make the trade concrete, here is a worked budget. It is illustrative, not a measurement: GitHub publishes per-model token rates but not a fixed credits-per-1M-context-call figure, so treat the numbers as a planning sketch, not a benchmark.

Hypothetical: a solo dev on Copilot Pro, 1,500 credits/mo (1 credit = $0.01)
Assume an agent task with extended context + reasoning costs ~5 credits.
(Illustrative only; actual cost depends on model + tokens consumed.)

  1,500 credits / 5 credits per task = ~300 heavy agent tasks/month
  Across ~22 working days = ~13-14 heavy tasks/day before the meter runs out.

  Push to 20 heavy 1M-context tasks/day and you exhaust Pro
  in roughly 15 working days, then it is upgrade or pay overage.

The shape is the point, not the exact figure. If your day is mostly inline completion (free, uncapped) with a few agent runs, Pro holds. If your day is agent-mode-with-1M-context as the default mode of work, Pro is the wrong tier and Pro+ at 7,000 credits is the honest pick. One more lever worth knowing: paid plans get a 10% discount on model costs while using auto model selection in Chat, the CLI, or the cloud agent (GitHub docs). Picking the model by hand forfeits that discount.

The model picker today, and the Fable 5 lesson

Model breadth remains a genuine Copilot strength. On Free you get Haiku 4.5, GPT-5 mini, and similar lightweight models through auto selection only. On Pro and above you get the full picker; Pro+ unlocks premium models including Opus (GitHub Copilot plans, 2026-06-27). One assistant, multiple frontier labs, switchable in the editor, with no need to commit to a single vendor's model. That is still a real reason to pick Copilot over a single-model tool.

The Fable 5 episode is the asterisk. Anthropic's first Mythos-class model went GA in the Copilot picker on June 9, pitched at long-horizon autonomous coding. Three days later it was gone: an editor's note dated June 12 confirmed access "suspended across all GitHub Copilot experiences" following Anthropic's announcement, with every other Claude model unaffected (GitHub changelog, Fable 5). The cause was outside GitHub's control, an export-control order after a jailbreak, which is precisely what makes it instructive. A multi-model menu is also a multi-policy-surface menu. If part of why you picked Copilot was access to a specific cutting-edge model, that access can vanish on a regulator's timeline, not a product one. The fix is to not architect your workflow around one model in the picker. Treat model breadth as resilience, not as a lock on the single best model of the month.

Copilot vs Cursor, now with credit math

The comparison most readers actually want. The two tools are not chasing the same buyer, and the June pricing change sharpens the line rather than blurring it.

Cursor's editing model is still ahead. The Tab completions and Composer's multi-file editing remain a step beyond Copilot's agent mode on whole-repo refactors, and Cursor's flat usage tiers mean a heavy agent day does not leave you watching a credit meter. For a solo developer who lives all day in one editor and runs agents constantly, that predictability is worth real money: Cursor's included usage covers a full working day, where Copilot Pro's 1,500 credits do not. If you are that developer, Cursor is the pick, and our Cursor review walks the tiers.

Copilot's case is narrower and still solid. It is the control plane. Issue-to-PR, PR summaries, and review suggestions execute inside GitHub's own permission and audit boundary, so a Copilot-generated PR inherits the same branch protections, required reviews, and audit-log entries as a human one. No editor-side competitor can replicate that, because none of them own the platform the work lands in. Pro+ adds the audit logs a regulated team needs. For a team already on a GitHub org with compliance requirements, that is the whole ballgame, and the model picker is a bonus on top.

The credit model narrows the cost gap in Cursor's favor for heavy solo users and widens Copilot's structural lead for org-bound teams. Same change, opposite effect, depending on which buyer you are.

The verdict, by who you are

Copilot earns a 4.2 in June 2026. The 1M context window is a real capability gain, and model breadth across GPT, Claude, and Gemini is a durable strength. It stays below Cursor's 4.5 for two reasons that have not moved: agent mode still trails Composer on multi-file refactors, and the new credit model adds cost uncertainty the old flat plan did not have, with Pro's budget tight enough that heavy users effectively must pay more.

The picks, by situation:

• Team on a GitHub org with compliance needs. Copilot, on Pro+. The PR automation and review inside GitHub's permission and audit model is the buy, the audit logs come with Pro+ at $39/mo, and model breadth is the bonus (GitHub Copilot plans, 2026-06-27). Nothing else competes on the control plane.
• Solo developer running agents and 1M context all day. Not Pro. The 1,500-credit budget runs out mid-month under that load. Go Pro+ (7,000 credits) or Max (20,000), or switch to Cursor, whose flat usage covers a full agent day and whose Composer edits multi-file refactors more cleanly.
• Developer who mostly accepts inline completions. Pro at $10/mo, comfortably. Completions are unlimited and never touch credits, and 1,500 credits cover the occasional chat or agent run with room to spare. Skip the 1M-context default unless you need it.
• Just trying AI coding. Free. Two thousand completions a month and auto model selection is a real on-ramp; upgrade when you hit the model-picker wall, not before.

Pick your constraint first. If it is reach, governance, and one assistant across many IDEs, Copilot's control plane is the product and Pro+ is the seat. If it is peak capability per dollar in a single editor with predictable monthly cost, Cursor still out-earns it. Re-check the live plans page before committing a team, because with token-metered billing your monthly number now depends on how you work, not just which box you tick.