Copilot retires 2 Gemini models on 31 July 2026: the admin migration checklist

Summary. GitHub announced on 2 July 2026 that Gemini 2.5 Pro and Gemini 3 Flash will be deprecated across all GitHub Copilot experiences on 31 July 2026, covering Copilot Chat, inline edits, ask and agent modes, and code completions. The named replacements are Gemini 3.1 Pro for Gemini 2.5 Pro and Gemini 3.5 Flash for Gemini 3 Flash. The switch is not automatic for every organisation: GitHub states that Copilot Enterprise administrators may need to enable access to the alternative models through their model policies in Copilot settings. The replacements are not priced the same as what they replace. Gemini 3.1 Pro Preview lists at $2 per million input tokens and $12 per million output tokens with a 1M-token context window, and Gemini 3.5 Flash at $1.50 and $9 with the same context size. Since Copilot moved to usage-based billing on 1 June 2026, with AI Credits consumed at the published API rates for each model, a model swap is now a billing change. Copilot Business includes $19 in monthly credits and Copilot Enterprise $39, matching their per-user seat prices.

Eleven days remain as this is published. The work is small if you start now and awkward if you find out on 1 August, because the failure mode is not an error message, it is a model policy that was never enabled.

What GitHub actually announced

The changelog entry is short and the table inside it is the whole instruction set.

Three sentences in that entry carry the operational weight. GitHub says to update workflows and integrations to use the supported models before the date. It says Copilot Enterprise administrators may need to enable access to the alternative models through their model policies in Copilot settings. And it says no action is required to remove the older models once they have been deprecated.

The verification path GitHub gives is worth following exactly: as an administrator, check your individual Copilot settings and confirm the policy is enabled for the specific model. Once enabled, the model appears in the Copilot Chat model selector in VS Code and on github.com. If it is not in the selector, the policy is not on, whatever the settings page seems to say.

The part that breaks quietly

Model deprecations in Copilot do not throw a loud error at the user. A developer opens the model picker, the model they used yesterday is gone, and they pick whatever is at the top of the list. That is the good case.

The bad case has three shapes, and each one is a different team.

The first is an enterprise where model policies were configured tightly a year ago and nobody has touched them since. Gemini 3.1 Pro and Gemini 3.5 Flash may sit behind a policy toggle that was never enabled because those models did not exist when the policy was written. Developers lose a Gemini option and silently move to a different vendor's model, which changes both output behaviour and credit burn.

The second is a team with automation. Any script, GitHub Action, Copilot CLI invocation or SDK call that names gemini-2.5-pro or gemini-3-flash as a string is a hard-coded reference to something that stops existing on 31 July. GitHub's instruction to update workflows and integrations is aimed squarely at this group.

The third is the newest and the least obvious. Enterprise managed-settings.json reached general availability on 1 July 2026 with 5 supported keys, and one of them is model. An enterprise that pinned a default model in that file has centralised a value that is about to become invalid. The file is fetched on every authentication, held in memory and refreshed hourly, and it takes precedence over user configuration for the supported keys. A stale pin there is worse than a stale pin on one laptop, because it applies to everyone holding a Copilot Business or Copilot Enterprise licence from that enterprise.

Check that file before anything else:

            gh api repos/YOUR-SOURCE-ORG/.github-private/contents/copilot/managed-settings.json \
  --jq '.content' | base64 -d
          

If the model value names either deprecated model, you have found your highest-priority fix.

What the replacements cost

This is where a deprecation stops being housekeeping. Copilot usage now consumes GitHub AI Credits, calculated on token consumption including input, output and cached tokens, using the listed API rates for each model. Swapping models changes the rate applied to every session.

Two implications follow, and they point in opposite directions.

Teams moving from Gemini 3 Flash to Gemini 3.5 Flash are moving to a model Google positions as bringing near-Pro level coding and reasoning at Flash-tier cost. Gemini 3.5 Flash defaults to medium thinking effort, with support for minimal, low, medium and high thinking levels. Thinking effort is a cost dial. A team that leaves everything on high will not recognise its own credit line.

Teams moving from Gemini 2.5 Pro to Gemini 3.1 Pro are moving up a generation of a reasoning model. OpenRouter's listing describes 3.1 Pro as delivering enhanced software engineering performance, improved agentic reliability and more efficient token usage across complex workflows, with a new medium thinking level to balance cost, speed and performance. More efficient token usage at a per-token rate is not automatically cheaper per task, and it is not automatically more expensive either. It depends on your workload. Measure it in the first week rather than arguing about it.

The blunt version: the deprecation is free, the model choice is not, and nobody will notice the difference until the credit pool report lands.

The 8-step checklist

Work through this in order. Steps 1 to 4 are the ones that must land before 31 July 2026.

1. Confirm which models your organisation actually uses. Do not assume. Pull Copilot usage metrics before you plan anything, since GitHub shipped improved accuracy and coverage in Copilot usage metrics reports on 2 July 2026 and the current numbers are more trustworthy than a memory of last quarter.

2. Enable the replacement models in your model policies. In Copilot settings, enable Gemini 3.1 Pro and Gemini 3.5 Flash. Then verify through your own individual Copilot settings that the policy is on for each specific model, and confirm each appears in the Copilot Chat model selector in VS Code and on github.com.

3. Grep your automation for hard-coded model names. Actions workflows, Copilot CLI scripts, SDK calls and internal tooling. A single search across your organisation is faster than waiting for a broken pipeline:

            gh search code --owner YOUR-ORG "gemini-2.5-pro OR gemini-3-flash" --limit 200
          

4. Fix the `managed-settings.json` model pin. Either move it to a supported model or set it to auto. Enterprise administrators can set model to auto in managed-settings.json so that Copilot auto model selection becomes the default for new conversations, with users still able to switch model per conversation.

            {
  "model": "auto",
  "disableBypassPermissionsMode": true
}
          

Note the version floor: the model permission is available in VS Code 1.126 and above. An estate on older VS Code builds will not pick this up.

5. Decide your fallback posture before you need one. Under usage-based billing, the old behaviour where users who exhausted premium requests fell back to a lower-cost model no longer exists. Usage is governed by available credits and admin budget controls instead. A model migration that pushes people to a costlier model and a credit pool that runs dry are the same incident.

6. Set or re-check budgets and session limits. Admins can set budgets at the enterprise, cost center and user levels, and choose whether to allow additional usage at published rates or cap spend when the included pool is exhausted. GitHub also shipped AI credit session limits in Copilot CLI and the SDK on 1 July 2026, which is the right control for long agent runs.

7. Tell developers what changed, in one message. Two models disappear from the picker on 31 July, here are the two that replace them, here is which one to pick for which job, raise a ticket if either is missing from your selector. Anything longer will not be read.

8. Measure credit burn in the first week of August. Compare against July. If the delta is ugly, the lever is model choice and thinking effort, not a policy argument about whether developers should use agents.

Auto model selection, and when not to use it

Setting model to auto is tempting because it makes this class of problem disappear. GitHub also shipped Copilot CLI auto model selection that routes based on task on 1 July 2026, and enterprises can now default new conversations to auto.

Auto is the right default for most teams. It survives the next deprecation without an admin touching anything, and routing a quick question and a multi-hour agent session to the same expensive model is exactly the waste that usage-based billing now bills you for.

Auto is the wrong default in three cases. If you are running evaluations and need a fixed model to compare against, pin it. If a regulated workload requires a named model recorded in an assessment, pin it. If your team has tuned prompts or custom agents against one model's behaviour, pin it and schedule the retuning rather than discovering the drift in a code review.

There is a middle option most teams miss: set auto centrally in managed-settings.json so new conversations start there, and let the teams that need determinism switch per conversation. GitHub explicitly preserves that per-conversation override.

How to validate the swap before 31 July

A model migration is a behaviour change, and the cheapest way to find out what changed is to run the same work twice.

Pick 10 tasks your team actually does, not benchmark prompts. A bug fix with a reproduction, a refactor across three files, a test-writing task, a code review comment, a schema migration, a flaky test diagnosis. Run each on the outgoing model and on its replacement, in the same repository, with the same prompt.

Record four things per run: whether the output was usable without rework, how many turns it took, wall-clock time, and credits consumed. The first two matter more than the last two. A model that costs 20% more per million tokens and needs one fewer round trip is cheaper in practice, and a per-token comparison will tell you the opposite.

Two behaviours specific to these replacements are worth probing directly. Gemini 3.5 Flash defaults to medium thinking effort and supports minimal, low, medium and high levels, so run at least one task at two different levels and see whether the higher setting earns its cost on your work. Gemini 3.1 Pro introduces a medium thinking level of its own and is positioned around agentic reliability and long-horizon stability, so test it on a multi-step agent task rather than a single question, which is where the difference shows.

Keep the results. The next deprecation notice will arrive, and a team that already has a comparison harness treats it as a morning of work instead of a fire drill.

What is not changing

Base plan pricing is unchanged. Copilot Business remains $19 per user per month and Copilot Enterprise $39, each including AI Credits equal to the seat price. Code completions and next edit suggestions remain included in all plans and do not consume AI Credits, which means the completions experience is unaffected by which chat model you standardise on.

The model catalogue around Gemini also kept moving through the same fortnight. Claude Sonnet 5 became generally available for GitHub Copilot on 30 June 2026 and Kimi K2.7 Code on 1 July 2026, so an organisation reworking its model policy has more to consider than a like-for-like Gemini swap. If you are re-evaluating rather than just replacing, our comparison of GPT-5.6 and Claude Sonnet 5 for enterprise agents and the Gemini 3.5 Pro, GPT-5.6 and Claude Fable 5 comparison cover the trade-offs in detail.

And existing Copilot Business and Copilot Enterprise customers received promotional included usage of $30 and $70 in monthly AI Credits for June, July and August 2026. That promotion ends after August. A model migration in late July whose cost impact is masked by promotional credits through August is a bill you meet in September.

India-specific considerations

For Indian product teams and GCCs, three things deserve attention.

Approval latency is the first. If your enterprise model policy changes need a security review, an 11-day window is not generous. Start the policy request now and do the workflow greps while it sits in the queue.

Model pinning for client commitments is the second. Delivery partners frequently name a model in a statement of work or a data processing schedule. Under the Digital Personal Data Protection Act 2023 the relevant question is not the model name but where processing happens and under whose instructions, so a substitution forced by a vendor deprecation should be notified to the client rather than absorbed silently. Write the notification once and reuse it.

Credit economics is the third. A 100-seat Copilot Business estate carries $1,900 of included credits a month at $19 per seat. Agent sessions consume credits at the published rate for the selected model, so standardising a large India delivery team on the Pro tier when the Flash tier would do is a decision with a visible monthly number attached. Pool the credits, then look at the distribution rather than the average.

FAQ

How eCorpIT can help

eCorpIT is a CMMI Level 5 certified technology consultancy in Gurugram, and our senior engineering teams run model migrations for client estates as a controlled change rather than a scramble: inventory first, policy second, measured credit impact third. We audit hard-coded model references across your workflows, set a defensible default in enterprise managed settings, and put a credit baseline in place before the promotional period ends in August 2026. If Copilot is deployed across more than a few teams, talk to us about a short migration review.

Related reading: our guides to controlling Copilot cost with AI credit pools, deciding whether to enable Kimi K2.7 in Copilot, enterprise inference cost for GPT-5.6 and the rollout controls for Copilot browser tools in VS Code.

References

1. Upcoming deprecation of Gemini 2.5 Pro and Gemini 3 Flash, GitHub Changelog, 2 July 2026.

1. Enterprise managed-settings.json is generally available, GitHub Changelog, 1 July 2026.

1. Enterprises can default to auto model selection, GitHub Changelog, 1 July 2026.

1. Copilot CLI auto model selection routes based on task, GitHub Changelog, 1 July 2026.

1. GitHub Copilot is moving to usage-based billing, Mario Rodriguez, GitHub Blog, 27 April 2026.

1. Gemini 3.1 Pro Preview pricing and specifications, OpenRouter, retrieved 20 July 2026.

1. Gemini 3.5 Flash pricing and specifications, OpenRouter, retrieved 20 July 2026.

1. Supported AI models in Copilot, GitHub Docs.

1. Configure enterprise managed settings, GitHub Docs.

1. Improved accuracy and coverage in Copilot usage metrics reports, GitHub Changelog, 2 July 2026.

1. Set AI credit session limits in Copilot CLI and SDK, GitHub Changelog, 1 July 2026.

1. Cost centers now support AI credit pools, GitHub Changelog, 2 July 2026.

1. Claude Sonnet 5 is generally available for GitHub Copilot, GitHub Changelog, 30 June 2026.

1. Kimi K2.7 Code is generally available in GitHub Copilot, GitHub Changelog, 1 July 2026.

Last updated: 20 July 2026.