GitHub Copilot retires 6 models on 1 September 2026: the migration map admins need

Summary. On 31 July 2026 GitHub published a deprecation notice removing six models from every Copilot surface on 1 September 2026: Gemini 3.1 Pro, Claude Opus 4.5, Claude Opus 4.6, Claude Sonnet 4.5, Claude Sonnet 4.6 and Raptor mini. The suggested replacements are Gemini 3.6 Flash, Claude Opus 4.7 or 4.8 or Opus 5, Claude Sonnet 5, and MAI-Code-1-Flash. None of these swaps is price-neutral. Gemini 3.6 Flash lists at $1.50 per million input tokens and $7.50 per million output tokens, while Claude Opus 4.8 lists at $5 and $25. Claude Sonnet 5 moves from its introductory $2 and $10 to standard $3 and $15 on the same day the deprecation lands. Since Copilot moved to usage-based billing on 1 June 2026, those list rates are what your GitHub AI Credits are spent against, so a model swap is now a billing change. Copilot Business remains $19 per user per month and Copilot Enterprise $39 per user per month, and Enterprise admins may have to enable the replacement model in Copilot policies before anyone can select it. There are 28 days left.

This is the second Copilot deprecation notice in a week. On the same day, 31 July 2026, GitHub also deprecated Gemini 2.5 Pro and Gemini 3 Flash. Teams that already worked through that change should not assume this one is the same event with a longer list.

What GitHub is retiring on 1 September 2026

The changelog entry is short and specific. The models go from all Copilot experiences at once, which GitHub enumerates as Copilot Chat, inline edits, ask and agent modes, and code completions.

Model retiring	Retirement date	GitHub's suggested alternative
Gemini 3.1 Pro	1 September 2026	Gemini 3.6 Flash
Claude Opus 4.5	1 September 2026	Claude Opus 4.7, Claude Opus 4.8, or Claude Opus 5
Claude Opus 4.6	1 September 2026	Claude Opus 4.7, Claude Opus 4.8, or Claude Opus 5
Claude Sonnet 4.5	1 September 2026	Claude Sonnet 5
Claude Sonnet 4.6	1 September 2026	Claude Sonnet 5
Raptor Mini	1 September 2026	MAI-Code-1-Flash

Two details in that notice matter more than the list itself.

The first is a carve-out. GitHub states that Claude Sonnet 4.6 stays available to individual Copilot subscribers on annual plans, so those customers retain a Sonnet offering, and the deprecation of Sonnet 4.6 does not apply to them. Everyone else loses it.

The second is an action item that is easy to miss because it sits in the last paragraph. GitHub writes that Copilot Enterprise administrators may need to enable access to the alternative model through their model policies in Copilot settings, and that an admin can verify availability by checking individual Copilot settings and confirming the policy is enabled for that specific model. In other words, the replacement does not necessarily appear in the model picker on 1 September just because the old one vanished. If nobody enables it, your developers open VS Code that morning and find a shorter list.

No action is required to remove the deprecated models once the date passes. GitHub handles that.

The replacements are not like-for-like

Every suggested alternative differs from the model it replaces on price, on capability, or on both. Under the credits system these differences land on the bill directly, because credits are consumed against published API rates for each model.

Retiring model	Suggested replacement	Replacement list price per 1M tokens (input / output)	What actually changes
Gemini 3.1 Pro	Gemini 3.6 Flash	$1.50 / $7.50	A Pro-tier model is replaced by a Flash-tier one; Google reports 17% lower output token usage than 3.5 Flash
Claude Opus 4.5	Claude Opus 4.8	$5 / $25	Same family, newer generation; Opus 4.7 and Opus 5 are also listed as options
Claude Opus 4.6	Claude Opus 4.8	$5 / $25	Opus-class pricing is the reference point for the credit burn
Claude Sonnet 4.5	Claude Sonnet 5	$2 / $10 to 31 Aug, then $3 / $15	The price step-up lands the same day as the deprecation
Claude Sonnet 4.6	Claude Sonnet 5	$2 / $10 to 31 Aug, then $3 / $15	Annual individual subscribers keep Sonnet 4.6
Raptor Mini	MAI-Code-1-Flash	Small-tier Microsoft model	Microsoft-built rather than a partner model

The Gemini row is the one worth reading twice. Gemini 3.1 Pro is a Pro-class model; Gemini 3.6 Flash is a Flash-class one. Google positions 3.6 Flash as a workhorse rather than a downgrade, and the numbers in its launch post of 21 July 2026 back a real capability jump over 3.5 Flash: DeepSWE at 49% against 37%, MLE Bench at 63.9% against 49.7%, and OSWorld-Verified at 83.0% against 78.4%. Google also reports 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, and up to 65% fewer on DeepSWE. Priced at $1.50 and $7.50 per million tokens, it is cheaper than 3.5 Flash on output.

What Google's post does not tell you is how 3.6 Flash compares with Gemini 3.1 Pro, the model your team is actually losing. Every published comparison is against 3.5 Flash. If your team standardised on 3.1 Pro for long refactors, benchmark the replacement on your own repositories rather than on Google's chart. The model reference page confirms the practical envelope: a 1,048,576-token input limit, 65,536-token output limit, caching, code execution, function calling, structured outputs and thinking all supported, with computer use in preview.

MAI-Code-1-Flash is the other substitution that changes the character of the tool rather than its version number. Microsoft introduced it on 2 June 2026 as a coding model trained from the ground up on clean, traceable, enterprise-grade data without distillation from third-party models, and trained directly against the GitHub Copilot harnesses used in production. Microsoft reports it beating Claude Haiku 4.5 on all four coding benchmarks tested, including a 16-point lead on SWE-Bench Pro at 51.2% against 35.2%, while solving harder problems with up to 60% fewer tokens on SWE-Bench Verified. Those are vendor-run numbers on a vendor-chosen comparator, so treat them as a reason to test rather than a result. For how these families line up against each other outside the Copilot picker, see our frontier model comparison of Gemini 3.5 Pro, GPT-5.6 and Claude Fable 5.

The credit maths, and why this is a billing change

Copilot stopped counting premium requests on 1 June 2026. Mario Rodriguez, Chief Product Officer at GitHub, set out the reasoning in the announcement on 27 April 2026: "Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount." Under the replacement system, usage is calculated on token consumption, including input, output and cached tokens, using the listed API rates for each model.

Three consequences follow for anyone planning around 1 September.

Plan prices did not change, but headroom did

Copilot Pro stays at $10 per month with $10 in monthly AI Credits, Pro+ at $39 with $39, Business at $19 per user per month with $19, and Enterprise at $39 per user per month with $39. Code completions and next edit suggestions stay included and consume no credits. Existing Business and Enterprise customers received promotional included usage for June, July and August 2026 of $30 and $70 in monthly AI Credits respectively. That promotion covers August. It does not obviously cover September, which is the month the more expensive replacement models arrive. Check your own billing page rather than assuming.

Fallback is gone

Under the old model, a user who exhausted premium requests could drop to a lower-cost model and keep working. GitHub removed that behaviour: usage is now governed by available credits and admin budget controls. A team that burns its pool on Opus 4.8 at $5 and $25 does not quietly degrade to something cheaper. It stops, or it spends.

Pooling and budgets are the lever

GitHub introduced pooled included usage across a business so unused credits are not stranded per user, plus budgets settable at enterprise, cost centre and user level. When the pool is exhausted, an organisation chooses between allowing additional usage at published rates or capping spend. If you have not set those budgets, 1 September is a good forcing function. We cover the mechanics in more depth in our guide to Copilot AI credit pools and enterprise cost control.

Copilot code review also consumes GitHub Actions minutes in addition to AI Credits, billed at the same per-minute rates as other Actions workflows. Teams that turned on Copilot code review with agent skills and MCP after it went generally available on 29 July 2026 are now paying on two meters for the same pull request.

The annual legacy plan is the awkward case

Copilot Pro and Pro+ subscribers on an existing annual plan stayed on legacy premium request-based billing after 1 June 2026. For them, model multipliers still apply, and GitHub's multiplier reference is the number that matters.

Model	Multiplier on legacy annual plans	Status on 1 September 2026
Claude Opus 4.5	15	Retiring
Claude Opus 4.6	27	Retiring
Claude Opus 4.7	27	Stays
Claude Opus 4.8	27	Stays
Claude Sonnet 4.5	6	Retiring
Claude Sonnet 4.6	9	Retiring, except for annual individual subscribers
Gemini 3.1 Pro	6	Retiring
Gemini 3.5 Flash	14	Stays
MAI-Code-1-Flash	0.33 (promotional rate)	Stays
Raptor mini	0.33	Retiring

Read down the right-hand column and the shape of the problem appears. An annual subscriber losing Claude Opus 4.5 at a multiplier of 15 has, on that same table, only 27-multiplier Opus options left. That is roughly 1.8 times the allowance burn for the same number of requests, before any change in how many tokens the newer model uses.

There is a second gap. As of 4 August 2026, GitHub's legacy multiplier table lists no entry for Claude Opus 5, Claude Sonnet 5 or Gemini 3.6 Flash, which are three of the four suggested replacements. GitHub notes the table is subject to change and that multipliers do not apply to usage-based billing at all. Annual subscribers should therefore watch that page in the last week of August rather than plan against today's version of it. The one mitigation GitHub does publish: auto model selection in Copilot Chat, Copilot CLI, the GitHub Copilot app and Copilot cloud agent earns a 10% discount, so a 1x model bills at 0.9x.

The honest read is that the annual legacy plan is a transitional artefact. It is being drained, not maintained. If your organisation still has annual Pro or Pro+ seats attached to engineering work, price the monthly usage-based equivalent now rather than in the week the models disappear. We worked through the general shape of that decision in our comparison of AI coding agent harnesses across Claude Code, Codex and Copilot CLI.

The 1 September collision nobody scheduled

Two independent price events land on the same date, and they compound.

GitHub removes Claude Sonnet 4.5 and 4.6 on 1 September 2026 and points those users at Claude Sonnet 5. Anthropic's Sonnet 5 launch post of 30 June 2026 states that introductory pricing of $2 per million input tokens and $10 per million output tokens runs through 31 August 2026, after which the model is priced at $3 and $15. That is a 50% increase in list rate, effective the same morning the migration is forced.

There is a third factor that is easy to overlook, and Anthropic states it in a footnote rather than the headline. Sonnet 5 uses an updated tokenizer, and the same input can map to roughly 1.0 to 1.35 times as many tokens depending on content type. Anthropic set the introductory pricing so the transition would be roughly cost-neutral. Once the introductory window closes, that cushion goes with it.

Stack the three together for a team moving off Sonnet 4.6 on 1 September: a forced model change, a 50% list-price step-up, and up to 35% more tokens for identical input. The arithmetic on a worst-case content mix is not a 50% increase. It is closer to double. Our detailed treatment of that tokenizer effect sits in the Claude Sonnet 5 September price cliff.

The mitigation is unglamorous and works: measure your real input mix against the new tokenizer during August, while the introductory rate is still live and the old models still exist. A week of parallel running in August costs less than a month of surprise in September.

Reasoning level is a new dial on the same bill

On 3 August 2026 GitHub shipped a change that interacts directly with this migration. You can now set the reasoning level for Copilot cloud agent when you delegate a task, for models that support it. GitHub's own framing is blunt about the trade: a higher level can improve answers to complex problems, but it consumes more tokens, and therefore more credits. The setting is available on all paid plans that include Copilot cloud agent, which GitHub lists as Pro, Pro+, Business, Enterprise and Max.

For a migration month this is useful rather than incidental. When you move a team from Opus 4.6 to Opus 5, the naive comparison holds reasoning level constant and reports that the new model costs more. The useful comparison asks whether the newer model at a lower reasoning level matches the older one at a higher level. Anthropic's own cost-performance curves for Sonnet 5 make that argument explicitly: effort level is the dial that decides where a model sits between cheap and capable.

Set a default reasoning level per repository class before 1 September. Long refactors on a legacy service justify a high setting. Dependency bumps do not.

What admins have to do before 1 September

GitHub has shipped three admin-side changes in the past week that together decide whether this migration is uneventful.

Enable the replacement models in policy

This is the step named in the deprecation notice. Copilot Enterprise administrators may need to enable the alternative model in Copilot settings model policies. Verify by opening your own Copilot settings and confirming the policy is on for that specific model; once enabled, it appears in the Copilot Chat model selector in VS Code and on github.com. Do this for every replacement your teams will touch, not only the one you personally use.

Check what your default enablement actually does

GitHub changed default model enablement for Copilot Business and Enterprise on 29 July 2026. If your organisation was relying on defaults rather than an explicit policy, confirm which models that default now covers before the deprecation, not after.

Use policy targeting instead of a single global switch

Enterprise teams model policy targeting entered public preview on 31 July 2026, and enterprise team specialization for managed settings followed on 3 August 2026. Together they let you give an expensive model to the teams that need it without opening it to everyone. That is the difference between granting Opus 5 to twelve platform engineers and granting it to four hundred seats sharing one credit pool.

The failure mode we expect to see most often on 1 September is not technical. It is an admin who enabled one replacement, assumed the rest followed, and finds out from a support ticket.

India-specific considerations

For Indian engineering organisations and global capability centres, three points sharpen the decision.

Copilot seat pricing is set in US dollars, so the rupee cost of a Business seat at $19 per user per month moves with the exchange rate on top of any usage change. Budget the credit overage separately from the seat line, because the seat line is predictable and the overage is not.

Second, the cheaper replacements deserve a fair trial here rather than a reflexive upgrade to Opus. A large offshore delivery team running routine tickets, test scaffolding and dependency work is exactly the workload MAI-Code-1-Flash and Gemini 3.6 Flash are priced for. Reserving Opus-class models for a small set of hard problems is a better cost structure than buying frontier capability for every seat. The real cost is usually the token mix, not the seat count.

Third, if your Copilot usage touches personal data, the Digital Personal Data Protection Act 2023 applies to what goes into the prompt, not to the model choice. Model deprecation is a good moment to re-check which repositories are in scope for Copilot at all, and whether prompt content crossing a border was ever reviewed. Changing models does not change that assessment, but nobody else is going to schedule it.

A 14-day migration plan

Twenty-eight days remain. Half of that is enough if the work is sequenced.

Days 1 to 3: inventory. Pull Copilot usage by model from your billing page and identify which of the six retiring models your teams actually use. Most organisations find two of the six carry almost all the traffic, and the other four are noise.

Days 4 to 7: enable and target. Turn on the replacement models in Copilot policy, then use enterprise team policy targeting to restrict the expensive ones to the teams with a case for them. Set budgets at the enterprise and cost centre level while you are in that screen.

Days 8 to 11: benchmark on your own code. Run the same twenty real tasks through the old model and its replacement, at a fixed reasoning level, and record both pass rate and token consumption. Vendor benchmarks compare against models you were not using; your repository does not.

Days 12 to 14: set defaults and tell people. Publish a one-page note stating which model is now the default for which kind of work, what the reasoning-level convention is, and what happens when the credit pool runs low. Do this before 1 September, because the alternative is doing it during an incident.

Then leave the last two weeks of August as slack. GitHub has shipped Copilot changes on 29, 30 and 31 July and again on 3 August 2026. The cadence is not slowing down before September.

FAQ

Which models does GitHub Copilot retire on 1 September 2026?

Six models leave every Copilot surface: Gemini 3.1 Pro, Claude Opus 4.5, Claude Opus 4.6, Claude Sonnet 4.5, Claude Sonnet 4.6 and Raptor Mini. GitHub published the list on 31 July 2026 and applies it to Copilot Chat, inline edits, ask and agent modes, and code completions. No user action is needed to remove them.

Do I lose Claude Sonnet 4.6 completely?

Not in every case. GitHub states that Claude Sonnet 4.6 remains available to individual Copilot subscribers on annual plans, so those customers keep a Sonnet offering, and the deprecation does not apply to them. Every other Copilot customer, including Business and Enterprise seats, loses access to Sonnet 4.6 on 1 September 2026.

What do administrators have to do before the deadline?

Copilot Enterprise administrators may need to enable the alternative model through model policies in Copilot settings. GitHub says an admin can verify availability by opening individual Copilot settings and confirming the policy is enabled for that model. Once enabled, it appears in the Copilot Chat model selector in VS Code and on github.com.

How much more will the replacement models cost?

That depends on the swap. Gemini 3.6 Flash lists at $1.50 and $7.50 per million input and output tokens. Claude Opus 4.8 lists at $5 and $25. Claude Sonnet 5 runs at $2 and $10 until 31 August 2026, then $3 and $15 from 1 September.

Why does Claude Sonnet 5 get more expensive on the same day?

The two events are unrelated but simultaneous. Anthropic set Sonnet 5 introductory pricing at $2 and $10 per million tokens through 31 August 2026, moving to $3 and $15 afterwards. GitHub independently chose 1 September 2026 for its deprecation, so teams forced onto Sonnet 5 arrive at standard rates rather than introductory ones.

Does the tokenizer change affect my bill?

Yes. Anthropic notes that Sonnet 5 uses an updated tokenizer and the same input can map to roughly 1.0 to 1.35 times as many tokens depending on content type. Introductory pricing was set so the move from Sonnet 4.6 was roughly cost-neutral. After 31 August that offset ends.

What happens if my credits run out mid-month?

Fallback to a cheaper model no longer exists. GitHub removed it when Copilot moved to usage-based billing on 1 June 2026, and usage is governed by available credits and admin budget controls instead. Organisations choose whether to allow extra usage at published rates or cap spend when the pooled allowance is exhausted.

Are model multipliers still relevant?

Only for Copilot Pro and Pro+ subscribers on an existing annual plan who stayed on legacy premium request-based billing. Multipliers do not apply to usage-based billing. On that legacy table Claude Opus 4.5 sits at 15 while the remaining Opus options sit at 27, which raises allowance burn for the same work.

How eCorpIT can help

eCorpIT is a senior-led engineering organisation in Gurugram, certified for CMMI Level 5 and ISO 27001:2022, and we run AI tooling migrations for engineering teams that cannot afford a month of unplanned spend. We benchmark replacement models against your own repositories rather than vendor charts, configure Copilot policy targeting and credit budgets per team, and hand back a written default-model convention your engineers will actually follow. If the 1 September deadline is sitting on someone's list without an owner, talk to us and we will scope the migration in a week.

References

1. Upcoming August 2026 model deprecations in GitHub Copilot, GitHub Changelog, 31 July 2026
2. Gemini 2.5 Pro and Gemini 3 Flash deprecated, GitHub Changelog, 31 July 2026
3. GitHub Copilot is moving to usage-based billing, Mario Rodriguez, GitHub, 27 April 2026
4. Model multipliers for annual plans on request-based billing (legacy), GitHub Docs
5. Customize the reasoning level for Copilot cloud agent, GitHub Changelog, 3 August 2026
6. Enterprise team specialization for managed settings, GitHub Changelog, 3 August 2026
7. Enterprise teams model policy targeting in public preview, GitHub Changelog, 31 July 2026
8. Default model enablement for Copilot Business and Enterprise, GitHub Changelog, 29 July 2026
9. Copilot code review: Agent skills and MCP now generally available, GitHub Changelog, 29 July 2026
10. Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, Google, 21 July 2026
11. Gemini 3.6 Flash model reference, Google AI for Developers
12. Introducing MAI-Code-1-Flash, Microsoft AI, 2 June 2026
13. Introducing Claude Sonnet 5, Anthropic, 30 June 2026

Last updated: 4 August 2026.