Home / Reports / The Frontier Reshuffle

Thank you for reading this post, don't forget to subscribe!

Industry Report · The AI Index

The Frontier Reshuffle: Gemini 3.5 Flash and Microsoft’s In-House MAI Models

In two weeks of mid-2026, the model race split into two fronts. Google launched Gemini 3.5 Flash — faster output and stronger agentic benchmarks — while Microsoft unveiled a family of seven in-house “MAI” models built to cut cost and reduce its reliance on OpenAI. Here is what each announcement actually claims, and why the two moves point in different directions.

By The AI Index · Updated Jun 25, 2026 · 6 min read

Google: Gemini 3.5 Flash and the speed front

On May 19, 2026, at Google I/O 2026, Google introduced Gemini 3.5 Flash as the first model in the new Gemini 3.5 family. Google positioned Flash — historically the lighter, faster tier — as a model that now outperforms last year’s Gemini 3.1 Pro on coding and agentic benchmarks, a sign of how quickly the cheaper tiers are closing on the frontier.

According to Google, Flash posts 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas (both agentic), a 1656 Elo on GDPval-AA, and 84.2% on CharXiv Reasoning for multimodal understanding. Just as notable as the scores is the throughput: Google says Flash produces output tokens roughly 4× faster than other frontier models, recasting “Flash” as a speed-and-agentic play rather than a budget compromise.

On price, the API lists $1.50 per million input tokens and $9.00 per million output tokens. Google also said a heavier Gemini 3.5 Pro would follow the next month, in June 2026 — signaling that Flash was the leading edge of a broader family rollout.

The numbers in full

Benchmark	Measures	Gemini 3.5 Flash
Terminal-Bench 2.1	Agentic / coding	76.2%
MCP Atlas	Agentic	83.6%
GDPval-AA	General capability (Elo)	1656
CharXiv Reasoning	Multimodal	84.2%

Source: Google. Figures as reported by Google; benchmarks reflect outperformance of Gemini 3.1 Pro on coding/agentic tasks.

Microsoft: seven in-house MAI models

Two weeks later, at Build 2026 in San Francisco on June 2, Microsoft pointed in a different direction. As reported by CNBC and Windows Central, the company unveiled a family of seven in-house “MAI” models, framing the launch as a bid for self-sufficiency and reduced reliance on outside providers — including OpenAI.

Named models include MAI-Code-1-Flash, which turns written descriptions into source code for apps and websites, and MAI-Thinking-1, a reasoning model. Microsoft AI CEO Mustafa Suleiman said the models surpassed OpenAI’s GPT-5.5 on McKinsey’s benchmarks while achieving a roughly 10× reduction in cost — and that hosting the models on Azure avoids royalty payments to partners.

The move did not come out of nowhere. In April 2026, Microsoft renegotiated its OpenAI deal, ending exclusivity and revenue-sharing — creating room to build models in-house.

“The MAI models surpassed OpenAI’s GPT-5.5 on McKinsey’s benchmarks while achieving a roughly 10× reduction in cost.”

— Mustafa Suleiman, Microsoft AI CEO, as reported by CNBC

What it means: two shifts at once

Read together, the two announcements describe a market splitting along two axes. On one, frontier labs like Google are racing on speed and agentic capability — pushing the lighter Flash tier past last year’s Pro while quadrupling throughput. On the other, hyperscalers like Microsoft are building in-house models to cut cost and dependence, using ownership of the stack (and Azure hosting) to sidestep royalties and reduce reliance on a single outside provider.

Neither move is purely about raw capability. Google’s pitch leans on price-performance and throughput; Microsoft’s leans on cost structure and strategic independence, enabled by the April 2026 renegotiation that ended its exclusivity and revenue-sharing arrangement with OpenAI. The benchmark and cost figures here are as reported by the companies and outlets cited; independent verification is still developing.

Frequently asked

What is Gemini 3.5 Flash and when was it released?

Gemini 3.5 Flash is the first model in Google’s new Gemini 3.5 family, released May 19, 2026 at Google I/O 2026. Google says it outperforms last year’s Gemini 3.1 Pro on coding and agentic benchmarks (Terminal-Bench 2.1 at 76.2%, MCP Atlas at 83.6%) while producing output tokens roughly 4× faster than other frontier models. API pricing is $1.50 per million input tokens and $9.00 per million output tokens.

What are Microsoft’s MAI models?

At Build 2026 (June 2, San Francisco), Microsoft unveiled a family of seven in-house “MAI” models, including MAI-Code-1-Flash, which turns written descriptions into source code, and MAI-Thinking-1, a reasoning model. They are framed as a bid for self-sufficiency and reduced reliance on outside providers, including OpenAI.

Why is Microsoft building its own models?

Microsoft AI CEO Mustafa Suleiman said the MAI models surpassed OpenAI’s GPT-5.5 on McKinsey’s benchmarks while achieving a roughly 10× reduction in cost, and that hosting on Azure avoids royalty payments to partners. In April 2026, Microsoft renegotiated its OpenAI deal, ending exclusivity and revenue-sharing — creating room to build in-house.

Related: AI model benchmarks 2026 · LLM market statistics 2026 · LLM token price index