# Claude Sonnet 4.5, Sora 2, and DeepSeek's sparse attention

> Week of 29 Sept 2025: Claude Sonnet 4.5 and Agent SDK, Sora 2, DeepSeek V3.2-Exp, GLM-4.6, Granite 4.0, Tinker.

Published 13 Oct 2026. Coverage: 29 Sept 2025 to 5 Oct 2025. By Ali Asif.

## Items

1. **Claude Sonnet 4.5** (Anthropic, Frontier): Anthropic's new coding and agent model, which it says can hold focus on complex tasks for more than 30 hours. [Primary source](https://www.anthropic.com/news/claude-sonnet-4-5)
2. **Building agents with the Claude Agent SDK** (Anthropic, Agents and coding): The Claude Code SDK is renamed and opened up as a general toolkit for building agents. [Primary source](https://www.anthropic.com/engineering/building-agents-with-the-claude-agent-sdk)
3. **Sora 2 system card** (OpenAI, Image, video, world): Video with synchronized sound and better physics, launched with a social app and a consent system for likenesses. [Primary source](https://cdn.openai.com/pdf/50d5973c-c4ff-4c2d-986f-c72b5d0ff069/sora_2_system_card.pdf)
4. **DeepSeek-V3.2-Exp** (DeepSeek, Open weights): Introduces sparse attention to cut long-context cost, and DeepSeek halved its API prices alongside it. [Primary source](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp)
5. **GLM-4.6** (Z.ai, Open weights): Context grows to 200K and agentic coding improves in an open-weights model. [Primary source](https://huggingface.co/zai-org/GLM-4.6)
6. **Granite 4.0 H Small** (IBM, Open weights): A hybrid Mamba and transformer family under Apache 2.0 with a lower memory footprint. [Primary source](https://huggingface.co/ibm-granite/granite-4.0-h-small)
7. **Announcing Tinker** (Thinking Machines Lab, Agents and coding): An API for fine-tuning open models where you control the training loop and the lab runs the infrastructure. [Primary source](https://thinkingmachines.ai/blog/announcing-tinker/)

Anthropic released Claude Sonnet 4.5 on Monday, and for a company that works through AI tooling this was the week's headline.

## Sonnet 4.5

Anthropic calls it its best coding model and says it can keep working on a complex task for more than 30 hours. I treat any single-number claim like that with caution until I've watched it on my own work. What I can say is that the direction fits how I use agents: long jobs with many steps, where the model has to remember what it decided an hour ago.

The same day they renamed the Claude Code SDK to the Claude Agent SDK. That detail is worth more than it sounds. It means the loop that powers Claude Code, with its tools, permissions, context management and subagents, is now offered as a building block for any agent, not only coding ones. If you want an agent that reads RFQs, drafts a supplier email and files the result, you don't have to invent the plumbing.

## Sora 2

OpenAI launched Sora 2, with video that comes with synchronized dialogue and sound effects and noticeably better physics, plus a social app built around letting people use their own likeness through a cameo system. I link the system card because OpenAI's launch page blocked my link check. For a content operator the two things to watch are quality and consent controls. Both decide whether a tool like this is usable in a brand or a government context.

## Open weights keep arriving

DeepSeek-V3.2-Exp added sparse attention to reduce the cost of long inputs and the company cut its API prices by more than half. Z.ai shipped GLM-4.6, now with a 200K window. IBM released Granite 4.0, a hybrid Mamba and transformer family under Apache 2.0, and it is certified to ISO 42001, which procurement people in regulated industries actually ask about.

## A different kind of tool

Thinking Machines Lab launched Tinker, an API for fine-tuning open models where you keep control of data and the training algorithm while they handle the GPUs. I haven't fine-tuned anything yet, but a service like this lowers the cost of trying.

## What I'd watch next

OpenAI's developer day is next week. I expect agent-building tools and an apps platform for ChatGPT.

## My take

Sonnet 4.5 plus the Agent SDK is the combination that matters to me. The model is one half. The other half is being able to wire it to my own tools with the same scaffolding Anthropic uses for Claude Code. That's what lets one operator run something that looks like a small department.

## Sources

- https://www.anthropic.com/news/claude-sonnet-4-5
- https://www.anthropic.com/engineering/building-agents-with-the-claude-agent-sdk
- https://cdn.openai.com/pdf/50d5973c-c4ff-4c2d-986f-c72b5d0ff069/sora_2_system_card.pdf
- https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
- https://huggingface.co/zai-org/GLM-4.6
- https://huggingface.co/ibm-granite/granite-4.0-h-small
- https://thinkingmachines.ai/blog/announcing-tinker/

Canonical: https://heyaliasif.com/ai/2025-09-29-week-in-ai