GitHub Copilot Drops Gemini: Model Shakeup May 2026

This post anchors to GitHub's Copilot Chat changelog entry shipped today, May 20, 2026 — the day GitHub removed every Gemini model from web chat, 30 days after Opus 4.7 left the Pro tier and 12 days before usage-based billing begins. The net effect: Microsoft's OpenAI-derived stack and its Anthropic partner stack now hold every seat in the Copilot model picker.

The timing compounds the editorial whiplash. Yesterday, May 19, GitHub published a changelog celebrating Gemini 3.5 Flash as "generally available for GitHub Copilot." Less than 24 hours later, the entire Gemini line was gone from the web interface. For developers who track Copilot's model picker as a buying signal, today's cull — combined with April's Opus reshuffle and May 17's GPT-5.3-Codex base-model swap — marks the end of the multi-vendor experiment on the web surface.

This guide documents what changed, when, and why — with the full Copilot Tier Matrix assembled from GitHub's changelog, the authoritative premium-request multiplier table, and the live pricing page. It also covers what Copilot Pro users who wanted Opus access should do now, and what June 1 usage-based billing means for planning.

On May 20, 2026, GitHub shipped a single changelog entry titled "Updates to available models in Copilot on web." The key sentence: "All Gemini models and several other models (e.g., GPT-5.2 Codex and GPT-5.4 nano) have been removed from Copilot Chat on the web."

GitHub framed the cut as a reliability and simplification move, not a commercial or contractual decision. The changelog reads: "Moving forward, web chat will support a more limited set of new model rollouts as we work to ensure optimal performance." The company confirmed that OpenAI and Claude models across price points remain available across Copilot plans.

Two important scope constraints apply. First, the removal is explicitly "Copilot Chat on the web" — the announcement does not state that Gemini has been removed from VS Code, JetBrains, or the Copilot CLI. Developers using Gemini models through those surfaces should verify current availability directly in the IDE picker. Second, GPT-5.2 Codex and GPT-5.4 nano were also removed alongside Gemini — both had been pre-announced for deprecation in a May 1 changelog entry.

The strategic framing — that GitHub (Microsoft) is consolidating on its own OpenAI-derived stack and the Anthropic partner stack — is consistent with observed behavior but is Digital Applied analysis, not a GitHub-stated reason. Officially, the stated reason is reliability and simplification. What GitHub chose to keep is consistent with that analysis: GPT-5.5, GPT-5.4, GPT-5.3-Codex, Claude Opus 4.7, Sonnet 4.6, and Haiku 4.5 all survived the cull.

Today's Gemini cull did not happen in isolation. One month ago, on April 20, 2026, GitHub VP of Product Joe Binder authored a major restructuring announcement for Copilot individual plans. Two changes mattered most: Opus-family models were removed from the $10/mo Pro tier, and GitHub paused new sign-ups for Pro, Pro+, and Student plans.

Binder's stated rationale was direct: "Opus-family models have been removed from Copilot Pro. Rate limits that would be required to keep Opus sustainable on Pro would result in a worse overall experience." The underlying driver he identified was agentic workflows: "Long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support."

The sign-up pause is still in effect as of today. The live github.com/features/copilot/plans page shows "Upgrades are paused as we roll out a flexible billing experience" on both the Pro and Pro+ cards. The April 20 announcement also granted existing Pro and Pro+ subscribers the option to cancel and request a prorated refund — with a deadline of today, May 20, 2026. That refund window closes at the same moment as today's Gemini removal.

The developer community's reaction to the April 20 announcement centered on cost. A Hacker News thread with 130+ comments included commenter claims that GitHub was "burning money because of Opus." That framing is community sentiment, not a GitHub-stated reason — but it aligns with the broader Anthropic inference cost story we covered in our Anthropic cost analysis. The 15× multiplier on Opus 4.7 is GitHub's mechanism for making that cost visible to individual developers before June 1 usage-based billing takes effect.

The table below is the first cross-tier model availability matrix combining the May 20 Gemini removal, the April 20 Opus-from-Pro reshuffle, the May 17 GPT-5.3-Codex base-model swap on Business and Enterprise, and the live premium-request multipliers in one grid. Sources: the authoritative multiplier table, the plans pricing page, and the May 20 changelog.

The multiplier column is the key cost lever for Pro users. At 300 premium requests per month, Opus 4.7 at 15× means 20 effective Opus prompts per month on the Pro plan — roughly one substantive agentic session. Pro+ at 1,500 requests and the same 15× multiplier yields 100 effective Opus prompts. This math is what drove the April 20 Pro removal: the original plan structure was not priced for the compute demand that agentic Opus sessions actually generate. Verify current figures at docs.github.com on publish day — multipliers change on GitHub's approximately two-week cadence.

No other outlet has pulled the receipt on this. The April 16 changelog announcing Claude Opus 4.7 on Copilot included a specific line: "This model is launching with a 7.5× premium request multiplier as part of promotional pricing until April 30th." Hacker News commenters flagged this as "introductory" pricing and predicted the post-promo rate would be higher. They were correct in 14 days: today's rate is 15×, exactly double the April 16 launch price.

The doubling is not a surprise price hike — GitHub published the promo end date at launch. What makes the trajectory editorially significant is that most coverage still cites Opus 4.7 as a "7.5× multiplier" model without acknowledging that rate expired on April 30. Any cost estimate for agentic Opus use in June and beyond should use the 15× figure. For the full context of why Anthropic inference costs at frontier scale push hosts toward this kind of pricing, see our Anthropic cost problem analysis and the Claude Opus 4.7 complete guide.

Three days before today's Gemini removal, GitHub shipped a quieter but structurally significant change: GPT-5.3-Codex replaced the previous default model for Copilot Business and Enterprise. The May 17 changelog entry reads directly: "GPT-5.3-Codex is now the base model for Copilot Business and Enterprise."

The practical impact for Business and Enterprise admins is a shift in what code completions and chat suggestions use by default when developers have not overridden the model picker. GPT-5.3-Codex carries a 1× multiplier, matching Claude Sonnet 4.6 and the prior GPT-5.2-Codex in cost terms. The switch does not change the available model roster for those tiers — it changes the default selection.

Taken with the April 20 Pro restructuring and today's Gemini cull, the May 17 swap completes a three-part repositioning of Copilot's default model stack: OpenAI Codex at the Business/Enterprise base, OpenAI and Claude models at the premium tier, and the Gemini line removed from the web surface entirely. The Microsoft-Anthropic partnership concentration is now visible in the default picker state, not just in the tier-matrix fine print.

For teams running Copilot Business or Enterprise and building AI-powered development workflows, our AI transformation service can help evaluate whether your current model mix and billing structure are optimized ahead of the June 1 usage-based billing transition.

On June 1, 2026, GitHub transitions Copilot from the current premium-request-based system to a usage-based "AI credits" billing model. The transition was confirmed via a May 12 changelog entry that released April 2026 usage reports for Business/Enterprise admins and individual Pro/Pro+ subscribers.

What GitHub has confirmed: the April reports translate existing usage into AI credit terms, giving organizations a directional signal before the June 1 switch. GitHub describes the April report as "useful for understanding cost shape, top consumers, and model usage" but frames it as "a directional signal, not a recalculated bill." Business and Enterprise admins should download these reports now and model their Opus vs Sonnet vs auto-selection split against the 15× and 1× multipliers respectively.

What GitHub has not confirmed as of May 20: the per-credit dollar pricing for AI credits after June 1. Developers building cost models for June and beyond should treat the current multiplier table as the last anchored reference point, and monitor GitHub's changelog for the credit-price announcement that is likely to drop between now and June 1.

Today is also the deadline to request a prorated refund for existing Pro and Pro+ subscriptions — the 30-day window from the April 20 announcement closes today, May 20, 2026.

Most coverage of today's Gemini removal reads it as a reliability story: Google couldn't keep up with Copilot's traffic demands, so GitHub simplified its model roster. That framing is consistent with GitHub's stated reason and may be entirely accurate. The strategic frame, however, is more interesting — and more important for teams choosing developer tooling.

Three changes in 30 days point in the same direction. April 20: Opus removed from Pro, concentrating Anthropic access on premium tiers where it generates more revenue per seat. May 17: GPT-5.3-Codex becomes the Business/Enterprise default, making OpenAI's Codex line the ground state for paid team accounts. May 20: Gemini removed entirely from web chat, eliminating the only non-Microsoft, non-Anthropic model family from the visible picker.

The result is that the Copilot model picker in May 2026 is an OpenAI-first, Anthropic-second environment — with Microsoft owning the first layer via its OpenAI partnership and the Anthropic stack providing the premium-tier differentiation. This is vertical integration arriving not through procurement announcements or press releases, but through two-week changelog cycles in a developer tool used by millions of engineers.

For developers who relied on Gemini-in-Copilot for its specific characteristics — particularly the 14× multiplier Gemini 3.5 Flash carried before removal, which positioned it between the 7.5× GPT-5.5 and 15× Opus 4.7 — the standalone path forward is Google's Gemini Code Assist product. See our Gemini 3.5 Flash vs GPT-5.5 vs Opus 4.7 head-to-head for what you're giving up technically, and our Copilot vs Cursor vs Windsurf comparison for the broader alternative landscape if today's changes prompt a tool review.

The broader market wave this fits into: from May 4 through May 21, 2026, the industry saw Anthropic's Blackstone JV, OpenAI's Deployment Co, SAP's Claude Joule integration, PwC's 30,000-seat Anthropic deal, KPMG's 276,000-seat expansion, and now EY's $1B Microsoft partnership. The Copilot model picker is the consumer surface of the same vertical-integration wave visible in those enterprise procurement headlines.

The editorial moment most rewrites of today's story will miss is the sequence on May 19 and 20. Yesterday, May 19, Google launched Gemini 3.5 Flash at Google I/O — arguably Google's biggest model announcement of 2026. On the same day, GitHub published a separate changelog entry: "Gemini 3.5 Flash is generally available for GitHub Copilot." Gemini 3.5 Flash on Copilot carried a 14× premium-request multiplier before removal — positioning it as a near-Opus-tier model for Copilot users who preferred the Gemini stack.

Less than 24 hours later, May 20, the entire Gemini line was removed from Copilot Chat on the web. Gemini 3.5 Flash moved from GA-on-Copilot to removed-from-Copilot within a single news cycle. The timing suggests the May 20 decision was in motion before the May 19 GA announcement shipped — or that the GA and the removal were coordinated as part of the same cleanup pass. GitHub's changelog offers no timeline for the decision.

For developers evaluating the Gemini line against Copilot's current roster, see our April 2026 AI coding assistant baseline for the pre-shakeup positioning, and the Cursor Composer 2.5 launch analysis — which shipped May 18, one day before the Gemini GA, and offers a concrete cost alternative for Pro users at $0.50/M input and $2.50/M output versus Copilot Pro's $10/mo with a 300-request premium cap.