GitHub Copilot Review (July 2026): Vision GA, Kimi K2.7, and Whether the Credits Math Changed

Copilot shipped six general-availability launches in a single week, and none of them raised the sticker price. Vision went GA, an open-weight coding model landed in the picker, browser tools reached VS Code, and org admins got per-team credit pools, all between July 1 and July 7 (GitHub changelog, pulled 2026-07-08). Our June 27 review rated Copilot 4.2 and predated every one of these. The re-call: 4.4 out of 5.

Here is the one thing to leave with. The July wave is a value increase at a flat price, not a repricing. Free, Pro at $10, Pro+ at $39, and Max at $100 all hold, and each now carries vision, Claude Sonnet 5, and a cheaper Kimi coding option inside its existing credit budget (GitHub Copilot plans, pulled 2026-07-08). The tenth-of-a-point math and the plan pick by persona are below, but the headline is simple: if you were on the fence in June, more shipped and the price did not move.

What changed since the June review

The June piece covered the June 1 move to AI credits, the 1M context window, and the model picker up through Claude Fable 5's brief suspension. Everything below landed after it. Each is dated and sourced.

• Vision is GA on every plan, including Free. You can attach images and PDFs (JPEG, PNG, GIF, WebP, PDF) to a chat prompt and Copilot reasons about them next to your code, in VS Code chat, on github.com, and in the CLI (GitHub changelog, July 1). No admin toggle. Paste a screenshot of a broken render or a PDF of an API spec and ask about it. That is the single most useful July change for a day-to-day developer.
• Kimi K2.7 Code is in the picker. Moonshot AI's model went GA July 1 for Pro, Pro+, and Max, the first open-weight model Copilot offers as a selectable option (GitHub changelog, July 1). GitHub frames it plainly as "a lower-cost option for your coding workflows," and the pricing backs that up: $0.95 per million input tokens and $4.00 per million output, versus $2.00 / $10.00 for Claude Sonnet 5 (GitHub models and pricing, 2026-07-08). On July 7 it opened to Business and Enterprise, though off by default there.
• Cost centers pool AI credits per team. An org admin can now cap how much of the shared monthly credit pool each cost center draws, so a team stays inside what its own licenses fund (GitHub changelog, July 2). Per-user budgets reached the billing UI on July 7. This is the admin story of the month.
• Claude Sonnet 5 is the new default-grade model. It went GA around June 30 for all paid plans and is at promotional pricing through August 31 (GitHub models and pricing). Anthropic calls it their "most agentic Sonnet yet"; on one agentic coding benchmark it scored 63.2% against Opus 4.8's 69.2% (AlphaSignal). Close enough to Opus for routine work at less than half the output token rate.
• Two Gemini models are on the clock. Gemini 2.5 Pro and Gemini 3 Flash are deprecated July 31; GitHub says migrate to Gemini 3.1 Pro and Gemini 3.5 Flash respectively, across chat, edits, agent mode, and completions (GitHub changelog, July 2). If your team pinned either, fix it this sprint.

Rounding out the week: browser tools for VS Code and the Copilot app both reached GA, and auto model selection in the CLI now routes by task (GitHub changelog). Useful, not decision-moving.

How the credit math looks now

The billing model is unchanged from June. One AI credit equals $0.01, code completions stay unlimited and never draw credits, and everything agentic (chat, CLI, cloud agent, third-party agents) meters against your plan's monthly budget (GitHub models and pricing, 2026-07-08). What July added is model choice inside that budget, and the choice matters because a routine task on Kimi K2.7 Code costs a fraction of the same task on a frontier model.

Plan	Price/mo	Monthly credit budget	July features included	Where the budget goes fast
Free	$0	2,000 completions, 50 chat requests	Vision, auto model selection	No premium models, no picker
Pro	$10	$15 in credits	Vision, Sonnet 5, Kimi K2.7, picker	Frontier models on 1M context
Pro+	$39	$70 in credits	All Pro plus premium models, audit logs	Sustained all-day agent runs
Max	$100	$200 in credits	All Pro+ at 2.9x the budget	Designed not to; heavy team-scale use
Source	plans	plans	changelog	pricing docs

All prices and budgets pulled from the Copilot plans page on 2026-07-08; new Pro, Pro+, and Max signups are still being gradually enabled, so a fresh paid seat may not be instant.

The lever the picker gives you is model tiering. Route routine completions-adjacent chat and boilerplate to Kimi K2.7 Code, reach for Sonnet 5 on real reasoning, and save Opus for the hard architectural calls. That stretches a fixed budget, and here is the shape of it as a planning sketch:

Example: routing on Copilot Pro ($15/mo in credits, 1 credit = $0.01)
Per-1M-token output rates from GitHub's models-and-pricing docs (2026-07-08):
  Kimi K2.7 Code : $4.00      Claude Sonnet 5 : $10.00      Claude Opus : $25.00

Same 1M output tokens of agent work, three ways:
  All Opus         -> $25.00  (already over a Pro month of credits)
  All Sonnet 5     -> $10.00  (two-thirds of the Pro budget)
  All Kimi K2.7    -> $4.00   (about a quarter of the Pro budget)

Illustrative only; real cost depends on input + cached + output tokens per call.
The point is the ratio: routing routine work to Kimi frees credits for the hard tasks.

The shape is what matters, not the exact figure. Vision costs tokens too, so a day spent pasting PDFs and screenshots into chat draws the same budget down faster. The default advice from June still holds: keep the standard context window on for everyday work and reach for 1M context only on genuine multi-file problems.

Plan picks by persona

Three buyers, three answers. The July wave moves two of them.

Solo developer. Pro+ at $39 is still the working tier, and it now buys more at the same price. Vision is in, Sonnet 5 is a strong default, and Kimi K2.7 Code lets you burn cheap credits on routine work and save the budget for hard tasks. Pro at $10 is fine only if your day is mostly inline completion (unlimited, free of credits) with a few agent runs (GitHub Copilot plans, 2026-07-08). The flip condition is unchanged from June: if you live in all-day multi-file agent editing, Cursor's Composer still edits cleaner and its flat usage removes the credit meter entirely. For that developer, Cursor is the pick, and our July Cursor review walks the current tiers.

Team admin. This is where July changes the buy. Before this month, per-team Copilot spend was hard to bound; cost centers now pool AI credits so each group is held to what its own licenses fund, and per-user budgets landed in the billing UI (GitHub changelog, July 2 and July 7). If your reason to standardize on Copilot was governance rather than raw editing speed, the control surface just got materially better. Enable the Kimi K2.7 Code policy if you want your team on the cheap model, because it is off by default for Business and Enterprise (GitHub changelog).

Enterprise. Copilot was already the control-plane buy for GitHub-native orgs, and the July wave widens that lead rather than changing it. Issue-to-PR, PR review, and automation run inside GitHub's own permission and audit boundary, and now finance can enforce chargeback boundaries per cost center without a side spreadsheet. The one action item is the Gemini deprecation: verify no team pinned Gemini 2.5 Pro or Gemini 3 Flash before July 31, and confirm the replacement models are enabled in Copilot settings policies (GitHub changelog, July 2).

Rating

Copilot earns 4.4 out of 5 in July 2026, up from 4.2 in June. The two tenths are specific, not vibes.

One tenth goes to vision reaching GA on every plan. Attaching a screenshot or a PDF spec to a prompt is a real workflow unlock, and GitHub shipped it to Free as well as paid, so it lifts the whole ladder (GitHub changelog, July 1). The other tenth is the model-picker economics: Kimi K2.7 Code at roughly a quarter of Sonnet 5's output rate turns the fixed credit budget from a hard ceiling into something you can route around (GitHub models and pricing, 2026-07-08). Cost centers earn part of that same tenth for admins, since per-team credit pools are the governance feature enterprise buyers kept asking for.

What holds it below a 4.5 is the same edge that held in June. Agent mode still trails Cursor Composer on whole-repo multi-file refactors, and the credit model keeps monthly cost variable in a way the old flat plan never was. The July churn adds two small frictions of its own: Kimi is off by default for orgs, so a team has to opt in, and the Gemini deprecation forces a migration on anyone who pinned those models.

The picks, restated by who you are. Enterprise and team admins should buy Copilot on Pro+ or the org plan: the cost-center pools and per-user budgets are the reason to standardize here over an editor-side rival. A solo developer who wants breadth and the GitHub-native control plane gets the seat at Pro+ for $39 (GitHub Copilot plans), a better deal this month than last at the same price. If your whole day is multi-file agent editing and you hate watching a usage meter, Cursor still edges it. And if you are only sampling AI coding, Free is the best it has been, because vision reached it too. Re-check the live plans page before committing a team; with token-metered billing your monthly number depends on how you route models, not just which box you tick.