Google stopped selling Gemini Code Assist. The migration plan is the product now

THE SIGNAL IN ONE SENTENCE

The first deadline has already passed. Google stopped selling new Gemini Code Assist Standard and Enterprise subscriptions on October 9. Existing subscriptions still work, and Google says they remain supported until the end of their contractual agreements. This is not a sudden service shutdown. It is the start of a commercial off-ramp with three different clocks and a replacement that asks more of the organization around it. Teams can add licenses to an existing subscription until January 31, 2027. Annual subscriptions can continue to renew until that same date. Monthly subscriptions have a longer runway and can renew until December 31, 2027. After those cutoffs, Google tells customers to move agentic coding work to Antigravity through eligible Gemini Enterprise editions. That sounds tidy on a deprecation page. It is not a one-click migration plan. Gemini Code Assist is licensed as a coding assistant for supported development environments. Antigravity is part of a broader family of autonomous developer tools inside Gemini Enterprise, including Antigravity 2.0, a command-line interface and IDE extensions. Google describes the platform as capable of multi-step code review, debugging, feature development, refactoring and test generation. The destination therefore changes more than the badge in the corner of VS Code. It can change the contract, billing account, entitlement model, administrative roles, enabled APIs, agent permissions, model menu, telemetry, logging, data route, regional constraints and compliance evidence. A team that waits until the last renewal invoice to discover those differences has already chosen its migration strategy. It is called panic. The plain signal is this: when a coding assistant becomes an agent platform, the migration plan is part of the product evaluation. Start with the calendar because Google has made it wonderfully easy to misread. October 9 closed new Gemini Code Assist Standard and Enterprise purchases. January 31, 2027 closes additional license purchases and annual auto-renewal. December 31, 2027 closes monthly auto-renewal. A contract may end on its own date. The useful date for any organization is therefore not the final date on Google's page. It is the earliest date when that organization's growth, renewal or contractual position stops matching its staffing plan. A company hiring twenty engineers in February 2027 cannot assume it will expand its existing pool. A team on an annual plan should not read the monthly deadline as extra time. A business buying through a reseller or negotiated agreement should verify its own paper rather than borrowing a date from somebody else's console. Make one ledger with every subscription, billing account, contract owner, renewal type, seat count, unassigned license, business unit and actual end date. Add the people who depend on the service and the repositories or workflows where it matters. This is the boring spreadsheet that prevents an expensive surprise. Then inventory the current behavior, not just the product name. Which engineers use inline completion? Which use chat, code transformation, agent mode or command-line automation? Is private-code customization part of the workflow? Are developers relying on IDE policy, project context, stored configuration, memory, Model Context Protocol connections or a specific model? Which features are merely available, and which ones are woven into release work? Usage data should answer that question before procurement does. A tool installed on 500 laptops may be important to 40 people and ignored by the rest. Conversely, a small command-line integration can sit inside a release pipeline that nobody remembers until it fails. Google's destination has its own prerequisites. The Gemini Enterprise developer-tools documentation says the project needs an eligible Standard, Plus, Standard Emerging Market or pay-as-you-go edition, depending on eligibility. It also needs an invoiced Cloud Billing account that receives an active monthly invoice. Users require the Gemini Enterprise User role even if they already have the Admin role. New subscriptions may enable the developer tools by default, while existing subscriptions require an administrator to turn them on for each location. Those are not decoration. They determine who can sign in on migration morning. The security defaults deserve an equally close look. Google documents controls for file access outside the working folder, terminal execution, sandboxing and browser access. Outside-folder access defaults to deny. Terminal auto-execution defaults to always proceed, while sandbox mode is disabled. Browser access can be unrestricted unless an administrator configures allowed URLs. Those defaults may be sensible for one team and unacceptable for another. The job is not to copy old policy labels into a new console. It is to rebuild the permission boundary around what the new agent can actually do. A completion tool suggests a line. An agent may inspect files, run terminal commands, follow links, use connected tools and carry a task across several steps. The blast radius changes when the workflow changes. Decide which repositories an agent can read, which paths it can edit, which commands require review, which network destinations it can reach and whether untrusted code must run in a sandbox. Then test those rules with an ordinary developer account. An admin screenshot is not evidence that the policy works at the edge. Logging is another migration choice. In the documented Gemini Enterprise settings, prompt-and-response logging and metadata logging are disabled by default. Administrators can enable them to create an audit trail and measure use. Google says the logs stay within the organization's project and are not used by Google for model training or human review. Turning logs on creates visibility. It also creates retained records of prompts, responses, code context or operational metadata that need access controls, retention rules and a legal basis appropriate to the organization. Leaving them off reduces that record but can make security investigations, adoption analysis and cost attribution harder. The correct answer is a deliberate policy, not whatever the default happened to be. Model access can shift too. Google currently lists several Gemini models as enabled for licensed users and allows administrators to opt into certain third-party models. Those third-party choices are off by default, require consent, use pay-as-you-go billing and can send the prompt and necessary code context to the selected provider under the platform's stated protections. Google also says Antigravity continues to use lightweight Gemini models for coordination even when a third-party model handles the primary reasoning and code generation. That matters for architecture review and cost controls. A model picker is also a data route and a billing switch. Do not assume the old and new product have feature parity because both can finish a function in an IDE. Google points to Antigravity extensions for VS Code and JetBrains, and it provides a separate guide for moving from Gemini CLI to Antigravity CLI. The existence of those routes is useful. It does not establish that code customization, quotas, response behavior, repository context, extension management, policy enforcement, regional processing or every current workflow carries over unchanged. Build a feature crosswalk with evidence. Put every required capability in a row. Record its current behavior, the proposed Antigravity behavior, administrative prerequisite, data path, cost rule, test owner and result. Mark unknown as unknown. A blank cell is more honest than a green check copied from a sales slide. The next step is a real pilot. Choose representative teams rather than only enthusiastic early adopters. Include a large monorepo, a regulated codebase, a remote developer, a restricted network, a JetBrains user, a VS Code user, a command-line-heavy workflow and somebody whose daily work is mostly maintenance. Give them production-shaped but contained tasks. Measure completion quality, accepted changes, test pass rate, time to review, security-policy prompts, unauthorized attempts, token consumption, latency and developer override behavior. Watch what the agent touches, not only what it produces. A fast patch is not a migration success if it reads the wrong directory or quietly doubles tool spending. Keep a fallback lane during cutover. The existing Code Assist subscription still works within its contract, so use the overlap. Avoid switching every team, extension, policy and pipeline on the same day. Move a cohort, preserve the old configuration, test incident response, and define the conditions that send the cohort back. A rollback plan needs more than reinstalling an extension. It should identify how to revoke new roles, disable APIs, restore editor settings, remove credentials, preserve necessary audit evidence, stop pay-as-you-go overages and return automation to the previous path. Give that plan an owner and rehearse it before the first important repository moves. Some organizations may decide Antigravity is the right destination. Others may use the deadline to compare competing coding tools or to reduce the number of agent platforms they support. The deprecation notice does not require loyalty. It requires a decision before the contract decides for you. Google's published compliance limits make that decision sharper for regulated teams. Its documentation says Antigravity in Gemini Enterprise does not support several controls and certifications, including Access Transparency, FedRAMP, IL4 and IL5, ISO 27001 and ISO 42001, ITAR, and SOC 1, 2 and 3. That is not a footnote to discover after a pilot reaches sensitive code. It belongs near the top of the eligibility screen. No migration checklist can guarantee equivalence because Google has not promised equivalence. The safe plan treats contracts, identity, policy, telemetry, data handling, cost and human workflow as one release. Map the deadline. Inventory the real use. Prove the destination. Move one cohort. Keep the return route open. Google stopped selling a subscription. What engineering teams are buying next is not simply another autocomplete box. It is an operating model for agents, and those are sold one permission at a time.