# Qwen3.5, Claude Sonnet 4.6 and Gemini 3.1 Pro in one week

> Week of 16 Feb 2026: Alibaba opens Qwen3.5, Anthropic and Google update their mid and top models, and Anthropic previews Claude Code Security.

Published 13 Oct 2026. Coverage: 16 Feb 2026 to 22 Feb 2026. By Ali Asif.

## Items

1. **Claude Sonnet 4.6** (Anthropic, Frontier): The mid-tier Claude gets a major upgrade, bringing near-flagship coding and computer-use skill to the cheaper model. [Primary source](https://www.anthropic.com/news/claude-sonnet-4-6)
2. **Gemini 3.1 Pro** (Google DeepMind, Frontier): Google's updated top model, with a model card and a stronger showing on reasoning benchmarks. [Primary source](https://deepmind.google/models/model-cards/gemini-3-1-pro/)
3. **Qwen3.5-397B-A17B** (Alibaba Qwen, Open weights): Alibaba's new open flagship is natively multimodal and activates only 17B of its 397B parameters per token. [Primary source](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)
4. **Tiny Aya** (Cohere Labs, Open weights): A small multilingual open model family that includes Arabic among many languages. [Primary source](https://huggingface.co/CohereLabs/tiny-aya-global)
5. **Claude Code Security** (Anthropic, Agents and coding): A limited preview that scans codebases for vulnerabilities and suggests patches for human review. [Primary source](https://www.anthropic.com/news/claude-code-security)

Three labs pushed updates inside five days, and the pattern is easy to read: the mid-priced tier is catching the top.

## Sonnet 4.6

Anthropic released Claude Sonnet 4.6 on 17 February. The pitch is that the model most people actually use now handles work that needed Opus not long ago, especially coding and operating a computer. It stays at Sonnet pricing.

I care about this more than the flagship launches. Frontier models set the headlines, but the mid-tier is what a company runs in volume. When a cheaper model clears the bar, workflows that did not pencil out last quarter suddenly do. That is how adoption actually spreads: through a price line crossing a threshold.

## Gemini 3.1 Pro and Qwen3.5

Google shipped Gemini 3.1 Pro on 19 February, with a model card that lists its reasoning results. On the open side, Alibaba released Qwen3.5-397B-A17B on 16 February. It is natively multimodal, covers a long list of languages, and uses a sparse design so only 17B parameters fire per token. That ratio is why big open models keep getting cheaper to serve.

Neither of these is a surprise by now. What stands out is the cadence. Each lab ships a visible improvement every few months, and my agent setup has to be built so I can swap models without rewriting everything.

## Smaller releases

Cohere Labs released Tiny Aya, a compact multilingual family that includes Arabic. Anthropic previewed Claude Code Security, which scans a codebase for vulnerabilities and proposes fixes that a person approves. Security review is a natural job for an agent, because it is tedious and exhaustive, and it is the kind of work that benefits from a human making the final call.

The security preview is worth another moment. Code review at scale is boring, which is precisely why people skip it, and why an agent can add real value there. I would not let it patch anything unreviewed. The workable pattern is that the agent finds and drafts, and a person with accountability signs off. That split between machine effort and human accountability is how I think about every agent I run, whether it reads code or supplier invoices.

**What I'd watch next:** how Sonnet 4.6 and Qwen3.5 compare to the top tier on real agent tasks, and whether Tiny Aya holds up on Gulf Arabic.

## My take

Sonnet 4.6 is the item that changes my bill. If the cheaper tier does most of what the top tier did a quarter ago, I can run more agents for the same money. And Tiny Aya listing Arabic is the kind of detail I check first, because a multilingual claim means little if the Gulf dialects are an afterthought.

## Sources

- https://www.anthropic.com/news/claude-sonnet-4-6
- https://deepmind.google/models/model-cards/gemini-3-1-pro/
- https://huggingface.co/Qwen/Qwen3.5-397B-A17B
- https://huggingface.co/CohereLabs/tiny-aya-global
- https://www.anthropic.com/news/claude-code-security

Canonical: https://heyaliasif.com/ai/2026-02-16-week-in-ai