Google ships Gemini 3.6 Flash to GitHub Copilot as its flagship coding model stalls in testing

Gemini 3.6 Flash Ships to GitHub Copilot While Google's Flagship Model Remains in Testing

The model Google shipped to GitHub Copilot today is not the one it planned to lead with. Gemini 3.6 Flash went generally available across Copilot's Pro, Pro+, Max, Business, and Enterprise tiers on July 21, delivering higher task-completion rates and better token efficiency than its predecessor on coding and agentic workflows 1. On the Artificial Analysis Index, it burns 17 percent fewer output tokens than Gemini 3.5 Flash 2. The gains are real. But the flagship Gemini 3.5 Pro, unveiled at Google I/O 2026 in May, has yet to ship. According to ByteIota, 3.5 Pro was expected to arrive by June but remains in internal testing after falling short on coding performance, hallucination rates, and multi-step reasoning during evaluation 2.

That gap between Flash shipping and Pro sitting out tells you where Google's model pipeline actually stands. And the fact that it is playing out inside GitHub Copilot tells you how the AI coding market is restructuring.

The Flash numbers hold up on inspection. On DeepSWE, which measures an agent's ability to resolve real software engineering tasks end-to-end, scores jump from 37 to 49 percent 2. MLE Bench, testing autonomous ML research, climbs from 49.7 to 63.9 percent 2. On OSWorld-Verified, a computer-use benchmark, Gemini 3.6 Flash posts 83.0 percent, the highest score in AlphaSignal's comparison, ahead of GPT 5.6 Luna and Grok 4.5 3. Output pricing dropped from $9 to $7.50 per million tokens, while input holds at $1.50 per million 2.

These are the results of a model doing genuine work. But they are the results of Google's lighter tier. The model Google put on stage at I/O, the one positioned to compete head-to-head with the best from OpenAI and Anthropic, has not made it to production.

ByteIota also reports that the same announcement covering 3.6 Flash included a mention of Gemini 4, described as "the most ambitious pre-training run yet," with no specs and no timeline 2. The picture from the outside: Flash is carrying the developer lineup, Pro is on hold, and the next major bet is already being signaled.

None of this is happening in isolation. GitHub Copilot has become a model marketplace where developers can pick from models built by OpenAI, Google, and Anthropic in the same model picker 1. Enterprise and Business administrators must enable the Gemini 3.6 Flash Preview policy before their teams can access it 1. Billing for this model runs on usage-based pricing at provider list rates, not the flat subscription many Copilot subscribers expect 1.

For developers, that multi-model picker is the practical end of single-vendor lock-in inside Copilot. You can run Gemini 3.6 Flash for its token efficiency, switch to an Anthropic model when you need deeper reasoning, or stay on OpenAI's stack, all from the same dropdown. The trade-off is cost complexity: usage-based billing means your monthly spend now depends on which model you pick and how aggressively it runs.

For investors, the signal is in the gap. A flagship model that misses its shipping window while its lighter sibling carries the product line points to a pipeline problem, not a routine product cycle. Google has the compute, the talent, and the distribution through Copilot and its own Gemini API. What it does not have, as of today, is its top-tier model in production. The Flash line is strong. But Flash was never meant to be the headline act.