Gemini CLI stopped serving requests on Thursday. If you had it wired into your terminal workflow, your CI scripts, or your project configurations, it stopped working 48 hours ago with 30 days' notice — less time than many sprint cycles.
This is the third forced migration from a major AI coding tool in under two months. GitHub switched to usage-based Credits pricing on June 1. Amazon blocked new Q Developer signups on May 15 and announced at AWS Summit last Wednesday that the tool reaches end-of-life April 30, 2027. And now Gemini CLI, retired exactly 30 days after Google launched Antigravity 2.0 as its successor.
None of these migration costs appear in any productivity research on AI coding tools.
What "30 Days" Means in Practice
Google announced Antigravity 2.0 on May 19. Gemini CLI stopped serving requests on June 18. The migration window was 30 days.
Thirty days is shorter than most quarterly planning cycles. For teams with Gemini CLI in CI/CD pipelines, script libraries, or project-level hook configurations, 30 days is the window between "we should figure this out" and "the build is broken." Not because the migration is technically hard — Antigravity CLI replaces gemini with agy in your shell and most functionality carries over — but because migration requires finding every place the old tool existed in your infrastructure and updating it. That work has to be scheduled. Someone has to own it. It has to not collide with the sprint in progress.
For a lot of teams, it did collide. The developers who discovered the Gemini CLI sunset on Thursday by watching their terminal return an error message are not outliers.
The Research Gap
The productivity literature on AI coding tools measures developers in steady-state. METR's controlled study, Faros.ai's two-year telemetry analysis, LinearB's benchmark report covering 8.1 million pull requests — all of them measure what happens when AI tools are working, adopted, and integrated into an established workflow. None of them measure what happens during months when the tool is changing underneath you.
This creates a selection effect in the data. If your team adopted Copilot in January 2025, started measuring productivity in February, and ran that measurement through March, your data doesn't include what happened when Copilot had twelve major incidents between November and April. It doesn't include the GitHub Credits migration in June. It doesn't include however many hours your senior engineers spent evaluating whether to stay on Copilot or switch to something else entirely.
The measurement windows in most AI productivity research are 12 to 24 weeks. The AI tooling landscape is changing faster than that.
Three Migrations in 50 Days
What's different about Q2 2026 isn't that any one migration is unusually disruptive. It's the frequency.
May 15: Amazon blocks new Q Developer signups. Teams evaluating the tool have to pivot.
June 1: GitHub switches to usage-based Credits pricing for Copilot. Teams running agentic sessions at scale see their cost model change. The actual invoice was the first signal most admins got.
June 17 to 18: AWS Summit announces Kiro as the mandatory Q Developer replacement, with an April 2027 deadline. The next day, Gemini CLI stops serving requests.
If you're a team that uses multiple AI tools — Copilot for daily completions, Gemini CLI for terminal scripting, Q Developer for AWS workflow tasks — you have had three infrastructure-layer changes in 50 days. The productivity research will not capture this.
What the Migration Actually Costs
The migration cost isn't primarily a learning curve. That's the part that gets described in migration guides, and it's usually not the problem.
Learning Antigravity CLI takes a few hours. The documentation is reasonable. The commands are sensible. That is not the cost.
The cost is finding every place gemini appears in your team's scripts, CI configs, and project hook files. Updating them. Testing that everything still works. Communicating to the team what changed. Handling the edge cases where Antigravity CLI behaves differently in subtle ways. That's a workday per affected developer, maybe two, depending on how deeply the old tool was integrated — before accounting for the cognitive overhead of figuring out what broke and why, and the senior engineer time spent evaluating whether to migrate or switch to something else entirely.
This is real work. It is not coding. It doesn't show up as commits or PR cycles. If you track your time at the OS level — across every application, not just your editor — you'll see migration weeks as a pattern: total coding time drops, browser time spikes (documentation, comparisons, error investigation), and context-switching looks more fragmented than usual. That's the signature of an infrastructure migration. Most productivity dashboards will call it a bad week without identifying the cause.
The Pattern Isn't New, But the Frequency Is
Developer tools have always changed. Package managers, build systems, test frameworks — all of them evolve and deprecate. Developers are used to occasional migration costs.
What's different now is the feedback loop velocity. AI tools are iterating at the pace of model improvement, roughly every three to six months for significant capability changes. But the workflow integrations developers build around these tools aren't designed for three-to-six-month deprecation cycles. A team that built a serious Gemini CLI workflow in January had five months before it stopped working. That's a capital investment with a five-month payback window, after which it became technical debt overnight.
The teams that are handling this period well are the ones that stayed shallow deliberately. They use AI tools through official integrations rather than building custom pipelines. They treat any workflow they build around a specific tool as having a lifespan of one major tool revision, not three years. They have a strong preference for thin harnesses over deep integrations.
This is the opposite of how teams were taught to think about developer tooling for the last decade. Deep integration was the signal of commitment. For AI tools in 2026, deep integration is a liability.
What This Does to ROI Calculations
The ROI conversation about AI coding tools is almost entirely conducted in terms of steady-state productivity gains. Developers go faster. Fewer bugs. More PRs shipped. Calculate the gain, subtract the licensing cost, call the remainder your ROI.
This framework misses migration costs entirely. Not because they're small, but because they're irregular and hard to attribute. A team that measured a 25% productivity gain from Copilot in Q3 2025 may have experienced a meaningful productivity dip in Q2 2026 due to the Credits pricing evaluation, Kiro assessment, and Gemini CLI emergency recovery — and their annual productivity data will show neither number clearly, because the loss gets diluted across the rest of the year.
The AI coding tool market is moving at this velocity now. Probably faster. The developers who build measurement practices now — who know what their steady-state productivity looks like at the OS level and can recognize when a tool change disrupted it — will have something the published research doesn't: a personal baseline that tells them what each migration actually cost, not what the announcement said it would.
Gemini CLI is gone. Antigravity is the future, apparently. If you noticed your workflow breaking Thursday and it took most of Friday to sort out, that's not just a bad day. It's a migration cost. Write it down somewhere.