# Claude Opus 4.5, FLUX.2, and a wave of strong open releases

> Week of 24 Nov 2025: Claude Opus 4.5, FLUX.2, Z-Image, DeepSeekMath-V2, Fara-7B, INTELLECT-3, and Anthropic on tool use.

Published 13 Oct 2026. Coverage: 24 Nov 2025 to 30 Nov 2025. By Ali Asif.

## Items

1. **Claude Opus 4.5** (Anthropic, Frontier): Anthropic's new top model, positioned for coding, agents and computer use, at a lower price than earlier Opus. [Primary source](https://www.anthropic.com/news/claude-opus-4-5)
2. **Introducing advanced tool use** (Anthropic, Agents and coding): Tool search, programmatic tool calling and usage examples keep agents from drowning in tool definitions. [Primary source](https://www.anthropic.com/engineering/advanced-tool-use)
3. **FLUX.2** (Black Forest Labs, Image, video, world): A new image model family with multi-reference editing, better text and open weights for the dev variant. [Primary source](https://bfl.ai/blog/flux-2)
4. **Z-Image** (Alibaba Tongyi-MAI, Image, video, world): A compact open image model, with a fast Turbo variant that runs on modest hardware. [Primary source](https://github.com/Tongyi-MAI/Z-Image)
5. **DeepSeekMath-V2** (DeepSeek, Science): An open math model trained to verify its own proofs, aimed at theorem proving. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-Math-V2)
6. **Fara-7B** (Microsoft, Agents and coding): A 7B open model built to operate a computer, small enough to run locally. [Primary source](https://huggingface.co/microsoft/Fara-7B)
7. **INTELLECT-3** (Prime Intellect, Open weights): A large open model post-trained with reinforcement learning on a decentralized training stack. [Primary source](https://huggingface.co/PrimeIntellect/INTELLECT-3)

Anthropic released Claude Opus 4.5 on Monday, which closed out a month in which every major lab put a flagship on the table.

## Opus 4.5

Anthropic positions Opus 4.5 as its best model for coding, agents and computer use, and it lowered the price relative to earlier Opus models. That second fact deserves attention. For two years the pattern was that the best model cost the most and you used it sparingly. Cheaper top-tier models change what you can afford to automate. A task that wasn't worth running at the old price might clear the bar now.

I'm skeptical of any release-day benchmark and I'm running my own comparison on the work I actually do: reading specs, building comparison sheets, drafting supplier questions. I'll report what I find.

The companion piece I'd recommend is Anthropic's engineering post on advanced tool use. It describes letting the model search for tools on demand instead of loading hundreds of definitions up front, and calling tools from code to avoid stuffing intermediate results into the context. Anyone building agents that touch many systems will hit this problem sooner than they expect.

## Images and open weights

Black Forest Labs released FLUX.2, with multi-reference editing and stronger text rendering, and open weights for the dev version. Alibaba's Tongyi lab opened Z-Image, a compact model with a Turbo variant that runs fast on consumer cards. Together those mean decent image generation is no longer a cloud-only service.

DeepSeek published DeepSeekMath-V2, an open model trained to check its own proofs. Microsoft released Fara-7B, a small computer-use agent. Prime Intellect released INTELLECT-3, trained with reinforcement learning on a decentralized setup, which is more interesting as a training story than as a leaderboard entry.

The White House also issued an executive order launching a federal science effort called the Genesis Mission. I haven't read it closely enough to say more.

One more thing on the image side. Having FLUX.2 and Z-Image arrive in the same week means a small team can now generate product shots, diagrams and brand assets without a monthly bill. For a company that sells physical infrastructure, being able to visualize a helipad lighting layout in minutes changes how fast a proposal can go out.

## What I'd watch next

DeepSeek has gone quiet since September. A new release from them would not surprise me, and it could be a reasoning model.

## My take

Fara-7B is the sleeper for me. A computer-use model that runs on a local machine means a portal-filling agent where supplier credentials and documents never leave the device. I'd start by testing it on harmless forms, and keep any send behind a human click.

## Sources

- https://www.anthropic.com/news/claude-opus-4-5
- https://www.anthropic.com/engineering/advanced-tool-use
- https://bfl.ai/blog/flux-2
- https://github.com/Tongyi-MAI/Z-Image
- https://huggingface.co/deepseek-ai/DeepSeek-Math-V2
- https://huggingface.co/microsoft/Fara-7B
- https://huggingface.co/PrimeIntellect/INTELLECT-3

Canonical: https://heyaliasif.com/ai/2025-11-24-week-in-ai