# Gemini 3.5 Flash at I/O, Qwen3.7-Max, Cursor Composer 2.5

> Google opens the Gemini 3.5 family at I/O with Flash first, Alibaba announces Qwen3.7-Max, and Cursor ships its own coding model.

Published 13 Oct 2026. Coverage: 18 May 2026 to 24 May 2026. By Ali Asif.

## Items

1. **Gemini 3.5 Flash** (Google DeepMind, Frontier): First model of the 3.5 family, reported at 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas, and the new default in the Gemini app and AI Mode. [Primary source](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/)
2. **Gemini Omni** (Google DeepMind, Image, video, world): A model that creates content from mixed inputs, starting with video. [Primary source](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/)
3. **Qwen3.7-Max** (Alibaba Qwen, Frontier): Announced May 20 with claims of 35-hour agent runs and over 1,000 tool calls without degradation. [Primary source](https://www.alibabacloud.com/blog/qwen3-7-the-agent-frontier_603154)
4. **Composer 2.5** (Cursor, Agents and coding): Cursor's in-house coding model, released May 18 at $0.50 in and $2.50 out per million tokens, built on Moonshot's Kimi K2.5. [Primary source](https://cursor.com/blog/composer-2-5)
5. **Project Genie with Street View** (Google DeepMind, Image, video, world): Experimental interactive 3D simulations of real places, running in the browser. [Primary source](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie-expands/)
6. **Self-hosted sandboxes and MCP tunnels for Managed Agents** (Anthropic, Agents and coding): May 19 let agents run in your own sandboxes and reach private MCP servers through a tunnel, in research preview. [Primary source](https://claude.com/blog/claude-managed-agents-updates)
7. **Codex Appshots and non-experimental Goal mode** (OpenAI, Agents and coding): The May 21 Codex update made Goal mode a regular feature and added Appshots on macOS. [Primary source](https://developers.openai.com/codex/appshots)
8. **Remote agents in Vibe, powered by Mistral Medium 3.5** (Mistral AI, Agents and coding): Mistral's coding agents move to the cloud and run tasks in parallel, with a Work mode preview in Le Chat. [Primary source](https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/)
9. **AlphaEvolve updates** (Google, Science): Real-world uses reported in supply chains, chip design, molecular simulation and power grids. [Primary source](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/alphaevolve-updates/)

Google I/O took most of the oxygen this week, and the rest of the field answered inside five days.

## Gemini 3.5 Flash

Google introduced Gemini 3.5 Flash on May 19 as the first model in the 3.5 family, and called it its strongest agentic and coding model so far. The reported numbers are 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, 84.2% on CharXiv Reasoning and 1656 Elo on GDPval-AA, all Google's. It says the model beats Gemini 3.1 Pro on the coding and agentic benchmarks it lists, runs about four times faster in output speed than other frontier models, and often costs less than half as much on long agentic tasks.

It became the default in the Gemini app and AI Mode in Search, and developers get it in Antigravity, the Gemini API in AI Studio, and Android Studio. Gemini Spark, a personal agent that runs in the background, is on 3.5 Flash and in trusted-tester testing. Google also said Gemini 3.5 Pro is in internal use and should roll out next month.

Alongside it came Gemini Omni, a model that creates content from mixed inputs and starts with video, and Project Genie with Street View, which turns real places into interactive 3D simulations in the browser.

## Everyone else, same week

Alibaba announced Qwen3.7-Max on May 20 at its cloud summit. It was not yet usable at announcement. The headline claim is agent runs of up to 35 hours and more than 1,000 tool calls without degrading. Cursor released Composer 2.5 on May 18, priced at $0.50 and $2.50 per million tokens. Cursor's own FAQ says it is built on Moonshot's Kimi K2.5. That tells you how far open weights now travel inside commercial products.

On the tooling side, Anthropic added self-hosted sandboxes and MCP tunnels for Managed Agents on May 19. OpenAI made Codex Goal mode a regular feature on May 21. Mistral moved its Vibe coding agents to the cloud on May 22.

## Science

Google reported AlphaEvolve uses in supply chains, chip design, molecular simulation and power grids.

What I'd watch next: whether Gemini 3.5 Pro shows up in June as promised, and what the Pro price looks like next to Flash.

## My take

Google said Gemini 3.5 Pro is coming next month, so I am treating Flash as the preview of the real thing. The number I care about in the Flash post is about half the cost of other frontier models on long agentic tasks, because my agents loop. Cheap loops are what let a small team run a procurement desk that never sleeps.

## Sources

- https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/
- https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/
- https://www.alibabacloud.com/blog/qwen3-7-the-agent-frontier_603154
- https://cursor.com/blog/composer-2-5
- https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie-expands/
- https://claude.com/blog/claude-managed-agents-updates
- https://developers.openai.com/codex/appshots
- https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/
- https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/alphaevolve-updates/

Canonical: https://heyaliasif.com/ai/2026-05-18-week-in-ai