GitHub Copilot App 0.2.28 Puts AI Credit Spend Where Developers Can Actually See It

GitHub Copilot app 0.2.28 is the sort of release that product teams like to call “quality of life.” That undersells it. The release puts token usage, context window, and AI credit spend into quick chats, which is exactly where agent cost visibility belongs: next to the developer making the decision, not buried in a billing report someone opens after the invoice already hurt.

This is the practical problem with agentic coding tools in 2026. The industry spent years teaching developers that AI assistance was a subscription feature. Now the product surface is shifting toward metered model usage, long-context sessions, background workers, cloud agents, scheduled automations, PR review, and MCP/tool calls. That is a different economic model. If the UI does not show cost at the moment of use, the workflow is lying by omission.

The GitHub Copilot app v0.2.28 release landed on June 8 at 23:47 UTC with 19 assets across Windows, macOS, and Linux. The github/app repository describes the product as a desktop application for agent-driven development with parallel workstreams, GitHub integration, and PR lifecycle management. During research, the repo had roughly 1,445 stars, 62 forks, 727 open issues, and a push timestamp less than a minute after the release. This is still a technical-preview surface, but it is becoming a serious workspace for agent operations.

The cost meter finally reaches casual chat

The headline feature is not conversation search, though search matters. It is that quick chats now show token usage, context window, and AI credit spend in the chat title menu, matching regular sessions. Quick chats are where developers are most likely to be casual: explain this file, summarize that thread, tell me why the test failed, draft a quick plan, compare these approaches. Casual is dangerous under token-priced billing because casual requests often drag in non-casual context.

A quick chat against a small snippet is one thing. A quick chat that carries repository context, previous conversation history, issue details, PR comments, and an expensive model is another. GitHub’s own Copilot pricing docs define 1 AI credit as $0.01 USD and price models by input, cached input, output, and, for some providers, cache writes. GPT-5.5 in Copilot is listed at $5 per million input tokens, $0.50 per million cached input tokens, and $30 per million output tokens in the default tier. Above the long-context threshold of 272K input tokens, GPT-5.5 rises to $10 per million input tokens and $45 per million output tokens. That is not a moral panic; it is a routing signal.

The UI change matters because developers do not optimize what they cannot see. If cost only appears in an admin dashboard, teams get folklore: “Copilot is expensive,” “agents are wasteful,” “never use the big model,” or “finance is overreacting.” If cost appears next to the interaction, developers can form accurate instincts. They can see when a prompt bloats output, when a context window is absurd, when a quick explanation should have been a local grep, or when a hard debugging task really did justify the expensive model.

Search and repository selection are cost controls wearing UX clothes

Version 0.2.28 also adds conversation search with Cmd+F or Ctrl+F, including match-case, whole-word matching, navigation between matches, and configurable dialog placement. That sounds like standard desktop app table stakes. In an agent workspace, it is more than that. Long-running agent sessions accumulate decisions, failed approaches, tool results, code references, and review notes. If developers cannot search them, they ask the model to summarize them. Summaries cost tokens and often flatten the parts that mattered. Search is cheaper, more precise, and less likely to invent a tidy narrative over messy work.

The new “New session in repository…” option when starting from an issue is another small feature with real economics. Issues do not always live in the repo where implementation belongs. They may sit in planning repos, umbrella projects, customer escalations, or cross-team boards. Starting an agent in the wrong repository burns tokens reading irrelevant code, produces bad plans, and creates human cleanup work. Letting the user choose the project repository at session start reduces wrong-context work before it begins.

That is the pattern across the release. Scheduled session automations now show a sidebar badge and the next scheduled run time in hover preview. PR and merge drawer comments render progressively while scrolling, improving behavior on PRs with many comments. Conversation timelines are always shown by default, with timeline navigation and bookmarks available when enough history exists. Unlimited plans no longer show incorrect exhausted-quota indicators, and session AI-credit spend is shown consistently in the usage gauge and workspace info popover. None of this is glamorous. All of it helps developers understand what work is active, where it runs, how much state it has, and what it costs.

Agent work needs instrumentation before policy

The broader Copilot context makes this release more important. GitHub has moved all Copilot plans onto usage-based billing, with included AI credits, optional additional spend, budget controls, and temporary limits for some individual-plan usage while safeguards are in place. Code completions and next edit suggestions remain unlimited for paid plans, but agentic features, code review, long-context requests, and model choice can all move the bill. Copilot code review is especially easy to misread because it consumes both AI credits and GitHub Actions minutes on private repositories.

That means teams need workflow guidance, not just budget caps. A hard budget without routing rules is just a surprise outage waiting for a developer at 4:30 p.m. on release day. The better model is explicit: use completions for small local edits; use lightweight models for narrow explanations, docs cleanup, fixture generation, and test scaffolding; reserve frontier models and large context for architecture changes, deep debugging, security-sensitive review, and ambiguous work where a wrong answer is expensive. Use cloud agents when the task has clear acceptance criteria and can be reviewed as a branch. Use scheduled automations only when someone owns the schedule and the output.

Copilot app 0.2.28 gives teams a useful feedback loop for that policy. Ask developers to look at the quick-chat spend. Capture examples in retros. Which prompts burned output tokens because they asked for essays? Which sessions pulled too much context? Which repository-selection mistakes were avoided? Which scheduled automations ran when nobody expected them? This is the same discipline teams eventually learned around CI minutes: the bill is less surprising when engineers can see which workflows are expensive.

There is also a competitive lesson here. OpenAI Codex is exposing token usage and sandbox/debug surfaces through its own runtime work. GitHub Copilot app is exposing AI credit spend, session timelines, scheduled automation state, and repository selection inside the desktop workflow. Cursor, Claude Code, Codex, Copilot, and local models are all converging on the same product truth: coding agents are not just model wrappers. They are operating environments. Operating environments win on observability as much as capability.

The release is not a full answer. Teams still need better org-level attribution, clearer model-routing defaults, and reports that tie spend to accepted work rather than raw tokens. But putting token usage, context window, and AI credit spend into quick chats is the right move because it trains the user at the point of behavior. If usage-based billing is going to be tolerable, the meter has to live where the work happens.

Sources: GitHub release — GitHub Copilot app v0.2.28, GitHub Copilot app repository, GitHub Docs — About the GitHub Copilot app, GitHub Docs — Models and pricing for GitHub Copilot, GitHub Community — usage-based billing FAQ