# Claude Opus 4.7, Seedance 2.0 report, Gemini Robotics-ER 1.6

> Opus 4.7 ships at an unchanged price, Codex gets a browser and computer use, and DeepMind and Physical Intelligence both push robot models forward.

Published 13 Oct 2026. Coverage: 13 Apr 2026 to 19 Apr 2026. By Ali Asif.

## Items

1. **Claude Opus 4.7** (Anthropic, Frontier): Better long-running coding, images up to 2,576 pixels on the long edge, a new xhigh effort level and the same $5 and $25 per million tokens. [Primary source](https://www.anthropic.com/news/claude-opus-4-7)
2. **Codex app 26.415: in-app browser and computer use on macOS** (OpenAI, Agents and coding): The April 16 Codex update adds an early in-app browser and computer use on macOS, not offered in the EEA, UK or Switzerland at launch. [Primary source](https://developers.openai.com/codex/browser)
3. **Seedance 2.0: Advancing Video Generation for World Complexity** (ByteDance Seed, Image, video, world): The technical report for ByteDance's video model was the top paper on Hugging Face this week. [Primary source](https://arxiv.org/abs/2604.14148)
4. **Gemini Robotics-ER 1.6** (Google DeepMind, Robotics): Better spatial reasoning and multi-view understanding, plus gauge reading developed with Boston Dynamics, available in the Gemini API and AI Studio. [Primary source](https://deepmind.google/blog/gemini-robotics-er-1-6/)
5. **Claude Design** (Anthropic Labs, Agents and coding): A research-preview product launched April 17 for designs, prototypes and slides, powered by Opus 4.7. [Primary source](https://www.anthropic.com/news/claude-design-anthropic-labs)
6. **Agents SDK update** (OpenAI, Agents and coding): The April 15 update adds sandboxed runs, memory controls and more room to customize the agent loop. [Primary source](https://developers.openai.com/api/docs/guides/agents)
7. **HY-World 2.0** (arXiv, Image, video, world): A multimodal world model for reconstructing, generating and simulating 3D worlds. [Primary source](https://arxiv.org/abs/2604.14268)

Anthropic and OpenAI both shipped on the same Thursday, and I spent most of the week reading release notes.

## Opus 4.7

Anthropic released Claude Opus 4.7 on April 16. The pitch is advanced software engineering that holds up on long tasks. Vision now accepts images up to 2,576 pixels on the long edge, which is over three times what earlier Claude models took, and there is a new xhigh effort level between high and max that Claude Code now uses by default. Task budgets arrive in public beta on the API. Anthropic's quoted partner numbers include CursorBench at 70% against 58% for Opus 4.6. Those come from the partners and the lab, so read them as claims.

Pricing stays at $5 per million input tokens and $25 per million output. The catch is the new tokenizer, which Anthropic says maps the same input to roughly 1.0 to 1.35 times more tokens depending on content. Opus 4.7 is on Bedrock, Vertex AI and Microsoft Foundry as well as the Claude API.

## Codex gets a browser

OpenAI's April 16 Codex app update adds an early in-app browser and computer use on macOS. It was not offered in the EEA, UK or Switzerland at launch. A day earlier the Agents SDK picked up sandboxed runs and memory controls. The direction is clear: coding agents are turning into general desktop workers. Anthropic made a similar move inside Claude Code with a new /ultrareview command, and gave Pro and Max users three free reviews to try it.

## Robots and video

Google DeepMind shipped Gemini Robotics-ER 1.6 on April 14. It is the reasoning layer for robots, with better spatial understanding, multi-view reasoning and a gauge-reading skill built with Boston Dynamics. That last piece interests me most, since reading instruments is a real part of inspection work at airfields. The ByteDance Seedance 2.0 report topped Hugging Face's papers, and an HY-World 2.0 paper pushed 3D world models the same week.

Anthropic also launched Claude Design on April 17 as a research preview for designs, prototypes and slides.

What I'd watch next: whether the gauge-reading result generalizes beyond Boston Dynamics' own robots, and how fast the token-count change shows up on real invoices.

## My take

I run most of my agents on Claude, so the line in the Opus 4.7 post I care about is the one about the tokenizer: the same text can map to roughly 1.0 to 1.35 times more tokens. Same sticker price, possibly a bigger bill. I would meter a week of real tender work before calling it a free upgrade.

## Sources

- https://www.anthropic.com/news/claude-opus-4-7
- https://developers.openai.com/codex/browser
- https://arxiv.org/abs/2604.14148
- https://deepmind.google/blog/gemini-robotics-er-1-6/
- https://www.anthropic.com/news/claude-design-anthropic-labs
- https://developers.openai.com/api/docs/guides/agents
- https://arxiv.org/abs/2604.14268

Canonical: https://heyaliasif.com/ai/2026-04-13-week-in-ai