On May 19, Gemini 3.5 Flash went generally available. Google announced it at I/O. It landed in GitHub Copilot. The benchmark numbers were respectable: 76.2% on Terminal-Bench 2.1, competitive pricing at $1.50/$9.00 per million tokens.
On May 20, GitHub removed it. Along with every other Gemini model.
The GitHub Changelog entry from that day is brief: “We’ve made updates to the models available in Copilot on the web. To streamline the experience and prioritize reliability, we’ve removed Gemini 3.5 Flash, Gemini 2.0 Flash, GPT-5.2-Codex, and GPT-5.4 nano from Copilot Chat.” The listed reason is reliability and simplification.
That’s the entire official statement. It does not explain why Gemini 3.5 Flash was added May 19 and removed May 20.
What the Copilot Model Roster Looks Like Now#
Before the removal, Copilot’s web interface offered a mix of OpenAI, Anthropic, and Google models. After it:
What stayed:
- GPT-5.4 (OpenAI)
- GPT-5.5 “Spud” (OpenAI)
- Claude Haiku 4.5 (Anthropic)
- Claude Sonnet 4.6 (Anthropic)
- Claude Opus 4.7 (Anthropic)
- o3-mini (OpenAI)
What’s gone:
- Gemini 3.5 Flash (Google) — removed one day after GA
- Gemini 2.0 Flash (Google)
- Any other Gemini model that had been accessible
- GPT-5.2-Codex (OpenAI — older generation)
- GPT-5.4 nano (OpenAI — low-end tier)
The result is that Copilot’s model choice on the web is now exclusively split between two AI labs: OpenAI, which Microsoft has billions invested in, and Anthropic, which Amazon and Google have both invested in. Google’s own models no longer appear in the product.
“Reliability” Is Doing a Lot of Work Here#
GitHub’s stated reason is reliability and simplification. This could be accurate — consumer-facing multi-model routing is genuinely complicated, and a model that’s day-zero GA may have capacity issues. Google’s APIs are not infallible.
But the timing creates a credibility problem for that framing. You don’t ship a model integration to users, announce it publicly at your biggest developer conference of the year, and then pull it 24 hours later for reliability reasons without something going materially wrong — or without a decision being made above the integration layer.
Microsoft and Google are, structurally, competitors. Microsoft invested $13 billion in OpenAI. Google has a competing AI stack (Gemini, Vertex, Antigravity) that it has been building out aggressively. GitHub Copilot competes directly with Google’s Gemini Code Assist, which went GA at Google I/O the same week. The two companies are in the same market, and Microsoft’s platform is surfacing Google’s models.
That’s not inherently untenable — both companies have also invested in Anthropic — but it creates alignment pressure that “reliability” alone doesn’t fully explain.
The simpler read is that Google’s models in Copilot created an internal inconsistency Microsoft wasn’t willing to maintain. A GitHub user on Copilot can now choose between OpenAI and Anthropic. If they want Google, they go to Gemini Code Assist.
Why the Timing Matters Even More This Week#
The model removal happened May 20. GitHub Copilot’s billing transition to AI Credits goes live June 1 — eleven days later. That billing change already caused a stir when preview bills showed one developer’s $39/month Opus usage becoming $902 under the new system.
The Gemini removal reduces the model roster complexity at precisely the moment GitHub needs to simplify its billing surface. Fewer models means fewer multipliers, fewer edge cases in the pricing table, and a cleaner story to tell enterprise buyers who are evaluating the June 1 transition.
That may be the most honest version of “reliability and simplification”: simplifying the commercial surface, not just the technical one.
Six Days Later: Enterprise Gets Model Assignment Controls#
On May 26, GitHub shipped a separate but related change: targeted model rules. Enterprise and Business plan admins can now assign specific models to specific organizations within their account, rather than applying a single model policy across the entire enterprise.
This matters a lot in the context of June 1 billing. If a team of senior engineers runs Opus 4.7 agentic sessions daily, that’s the right model for the job — but at a 27x credit multiplier, it’s also expensive. A team of junior developers doing quick code completions and occasional Sonnet chat doesn’t need Opus. Under the old model policy, you had one knob for the whole enterprise. Now you have per-org granularity.
For enterprise admins, this is a meaningful cost control lever. You can pin cost-sensitive organizations to Sonnet 4.6 (1x multiplier) or Haiku 4.5 (0.33x) and reserve Opus access for teams where the productivity lift justifies the spend. Given that preview bills have shown 23x increases for heavy Opus users, this granularity could be the difference between a manageable billing transition and a department-level budget crisis.
| Model | Multiplier | Best For |
|---|---|---|
| Claude Haiku 4.5 | 0.33x | High-volume completions, lightweight chat |
| Claude Sonnet 4.6 | 1x | General-purpose coding, most teams |
| GPT-5.4 | 6x | Teams preferring OpenAI’s code style |
| Claude Opus 4.7 | 27x | Complex agentic sessions, senior engineers |
What This Means If You’re Evaluating Your Platform Choices#
The Gemini removal and the model rules launch in the same week send a coherent signal: GitHub Copilot is consolidating around a curated model set and giving enterprises more control over cost, not more model choice.
If you’re a developer who chose Copilot partly because it offered multiple frontier models including Google’s — that’s no longer the value proposition on the web interface. The differentiation is now GitHub integration depth, enterprise controls, and the quality of the OpenAI and Anthropic models on offer.
For teams actively choosing between platforms right now:
Stay on Copilot if: GitHub integration (PR review, Actions, Issues, code search) is core to your workflow, and your model preference is OpenAI or Claude. The targeted model rules and June 1 billing controls give enterprise admins the tools they need to manage cost.
Look at alternatives if: You were using Copilot specifically for model diversity, or you’re on a Pro+ plan and your preview bill was a shock. Claude Code Max at $200/month offers flat-rate Opus access. Direct API access gives you every model without platform markup.
One thing to do before June 1: Pull your preview bill in GitHub Billing Overview. If the number is uncomfortable, the targeted model rules are now available to change what your team defaults to. Set org-level rules, configure spending caps, and audit which teams actually need Opus vs. Sonnet before Monday.
The signal from this week is that Copilot is becoming a more controlled, more curated product — fewer models, more enterprise governance, usage-based billing with per-org cost assignment. Whether that’s a better product depends on whether your team’s needs match what’s left on the roster.
Google’s models are not on that roster anymore. At least for now.
Sources:
- Updates to available models in Copilot on the web — GitHub Changelog, May 20, 2026
- Target Copilot models to organizations with model rules (public preview) — GitHub Changelog, May 26, 2026
- GitHub Copilot Drops Gemini Models: Strategic Move or Technical Glitch? — Digital Applied
- GitHub Copilot removes Gemini AI models days after launch — IT Pro