GitHub Copilot now offers Gemini 4 Hydrogen, while six older models are due to leave Copilot on 14 December 2026. A small team should not respond by changing every workflow at once. First identify where a named model is selected, test one repeated task against an available replacement and keep a clear review and rollback path.
This matters beyond software companies. A food brand, property agency or creative studio may rely on a developer, contractor or internal automation to maintain forms, websites and customer handovers. A model change is useful only when the resulting work remains accurate, reviewable and safe to ship.
What GitHub actually released
GitHub announced Gemini 4 Hydrogen as generally available in Copilot on 13 November. It is available to Copilot Pro, Pro+, Max, Business and Enterprise users across supported Copilot surfaces, including Visual Studio Code, the command line, the coding agent and GitHub.com. GitHub says rollout is gradual, so an eligible user may not see it immediately.
For Business and Enterprise accounts, administrators can manage access through model policies. GitHub also states that unconfigured generally available models can be enabled through the organisation's default model policy. The practical point is that a new option may become available without every team having made an explicit task-by-task decision.
Availability is not the same as suitability. GitHub describes Hydrogen as a fast model for agentic coding, tool use and multistep work, and bills it at provider list pricing under usage-based billing. Those product statements do not establish the quality, cost or risk of a specific business workflow.
Clear now: find where model choice affects delivery
List the workflows that name a model or depend on Copilot behaviour: code generation, reviews, tests, documentation, deployment scripts and any agent that can edit a repository. For each one, record the owner, the current model, the allowed tools and the evidence required before a change reaches production.
GitHub also announced that six Copilot models will be deprecated on 14 December, including Gemini 3.8 Flash and Gemini 4 Argon. Its suggested alternatives include Gemini 4 Hydrogen and Gemini 4 Carbon. A workflow with a hard-coded model, a policy restriction or an untested fallback deserves attention before that date; a casual chat with no production dependency is lower priority.
Build now: run one controlled comparison
Choose a representative task that already has a known good result. Give the current model and the proposed replacement the same repository context, instructions, tool permissions and definition of done. Review the output rather than relying on the model's own completion message.
For a restaurant website, the task might be adding an allergen note without changing menu prices or the order route. For a property workflow, it might be updating a viewing-request form without exposing private applicant data or converting a request into a confirmed appointment. These are illustrative tests, not reported client results.
- Correctness — did the change meet the stated acceptance criteria?
- Scope — did it alter unrelated files, offers or customer journeys?
- Review effort — how many corrections were needed before approval?
- Cost and time — what did the complete reviewed task consume?
- Recovery — can the team restore the previous working state?
Launch now: change the policy with an owner and a fallback
If the comparison is acceptable, enable the model for a small group or one bounded workflow first. Keep human approval on production changes, record the model and policy used, and retain the last working configuration. If the model is not yet visible, the documented gradual rollout is a reason to wait, not to bypass account controls.
Teams approaching the 14 December deadline should also test the suggested replacement for any retiring model they actually use. Do not assume that automatic enablement means existing prompts, integrations or review standards will behave identically.
Proof loop: keep the model only if the system improves
Compare equal tasks using the same measures: accepted output, correction rounds, review time, failed tests, usage cost and production incidents. Separate a faster first response from a faster approved result. A new model is commercially useful when the whole workflow becomes more dependable or efficient, not when the demo looks impressive.
The immediate founder action is simple: inventory named-model dependencies, prioritise anything affected by the 14 December retirements, test one real task and assign the final production decision to a person. That turns a product update into a controlled operating improvement.
Sources & further reading
Klarnow editorial analysis of GitHub's 13 November 2026 release and deprecation notices, verified on 15 November 2026. Availability depends on Copilot plan, client, rollout and account policy. Food and property examples are illustrative. No first-hand model test, cost saving or business outcome is claimed.

