# Opus 4.7 Lands as Codex Expands and OpenAI Moves Into Life Sciences

*By AI High Signal Digest • April 17, 2026*

Anthropic’s Opus 4.7 and OpenAI’s expanded Codex defined the day, while GPT-Rosalind signaled a sharper move toward domain-specific frontier models. The brief also covers Qwen’s new open model, Perplexity’s local-compute push, and a notable Pentagon AI policy development.

## Top Stories

*Why it matters:* The biggest shift today is toward more capable work agents and more specialized frontier models.

1. **Anthropic released Claude Opus 4.7.** Anthropic says it is its most capable Opus model yet, handling long-running tasks with more rigor, following instructions more precisely, and verifying its own outputs before reporting back [^1]. Pricing stays at **$5 / $25 per million tokens**, with availability across the API, Bedrock, Vertex AI, and Microsoft Foundry [^2]. Third-party measurements highlighted stronger coding and agentic performance, including **70% on CursorBench vs. 58% for 4.6** and **1753 on GDPval-AA** at max effort [^3][^4]. The tradeoff: Anthropic says 4.6-tuned prompts may need rework because 4.7 interprets instructions more literally, and the new tokenizer can raise token counts [^2].

2. **OpenAI expanded Codex far beyond coding.** Codex can now use apps on a Mac, open an in-app browser, generate images, remember preferences, and take on ongoing or repeatable tasks [^5][^6]. OpenAI also added **90+ plugins**, GitHub review-comment handling, and remote SSH connections to devboxes [^7][^8][^9].

> "Codex for (almost) everything." [^5]

This matters because OpenAI is turning Codex into a broader computer-use agent, not just a coding assistant [^10].

3. **OpenAI launched GPT-Rosalind for life sciences.** The new model is built for **biology, drug discovery, and translational medicine** [^11][^12]. OpenAI says it is optimized for scientific workflows, with stronger performance in **protein and chemical reasoning, genomics analysis, biochemistry knowledge, and scientific tool use** [^13]. Access starts as a trusted-access research preview for qualified customers including **Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific**, and OpenAI is also releasing a **Life Sciences plugin for Codex** [^14][^12].

## Research & Innovation

*Why it matters:* The research signal is two-sided: open models are getting stronger, but new benchmarks still show large capability gaps.

- **Qwen3.6-35B-A3B** is a new open sparse MoE model with **35B total parameters and 3B active**, released under Apache 2.0 [^15]. Alibaba says it reaches agentic coding performance on par with models **10x its active size** [^15], while third-party summaries highlighted **73.4 on SWE-Bench Verified**, **51.5 on Terminal-Bench 2.0**, and strong vision scores despite the small active parameter count [^16].
- **LongCoT** introduced **2,500 expert-designed** long-horizon reasoning problems across chemistry, math, computer science, chess, and logic [^17]. At launch, the best frontier models were still **below 10% accuracy**, underscoring how far long-context systems remain from robust long-horizon reasoning [^17].
- Google’s **Auto-Diagnose** shows what applied LLM tooling can look like in production: inside Google’s Critique code review system, it analyzes failure logs, summarizes relevant lines, and suggests root causes [^18]. Google reported **90.14%** root-cause diagnosis accuracy on 71 real-world failures and usage across **52,635** distinct failing tests after deployment [^18].

## Products & Launches

*Why it matters:* New launches are increasingly about giving agents durable access to local software, media workflows, and voice interfaces.*

- **Perplexity Personal Computer** integrates with the Mac app for secure orchestration across local files, browser workflows, and native apps like **iMessage, Apple Mail, and Calendar** [^19][^20]. It can also run in the background on a Mac mini and be triggered from iPhone [^21].
- **Gemini 3.1 Flash TTS** landed in AI Studio, with tag-based control over vocal delivery such as pace and accent, plus a composer view for iteration and code export [^22].
- **VOID**, now live on fal.ai, brings Netflix’s video object removal model to developers, including correction of physical interactions and support for **multiple objects, fast motion, and complex backgrounds** [^23].

## Industry Moves

*Why it matters:* Big companies are now reorganizing teams, distribution, and platforms around AI-native workflows.

- **Apple** is reportedly sending **close to 200 Siri staffers** to a multi-week coding bootcamp using AI tools such as **Claude Code** and **Codex** [^24]. The report says roughly **60 engineers** stay on core development and another **60** handle evaluations and safety checks [^24].
- **GenAI traffic is no longer a one-player market.** Similarweb data shows ChatGPT at **56.72%** share one month ago, Gemini at **25.46%**, and Claude at **6.02%**, versus **77.43%**, **6.00%**, and **1.40%** respectively 12 months earlier [^25].
- **Salesforce launched Headless 360,** exposing Salesforce, Agentforce, and Slack as **APIs, MCP, and CLI** so AI agents can access workflows and data directly without the browser being the primary interface [^26].

## Policy & Regulation

*Why it matters:* Government adoption is moving from theory to contract language, and the terms matter.*

- Reporting cited in analysis says **Google is negotiating a classified Pentagon AI agreement** to deploy Gemini in secure environments [^27]. The same analysis says proposed language would mirror OpenAI’s earlier Pentagon deal [^27][^28]. Critics argue that even where red lines are stated, broad "all lawful purposes" language may leave room for wider military or surveillance use [^28][^29].

## Quick Takes

*Why it matters:* These smaller updates help show where momentum is building next.*

- **Stanford HAI** released the **2026 AI Index**, a 400+ page report covering AI performance, investment, labor, policy, and public sentiment [^30].
- **Google DeepMind and Boston Dynamics** are powering **Spot** with Gemini Robotics embodied reasoning models for inspection-style tasks [^31][^32].
- A **Cursor/UChicago** study across 500 teams found developers tackled **68% more high-complexity tasks** as models improved, while overall AI usage rose **44%** [^33][^34].
- **PrismML** open-sourced **Ternary Bonsai**, a 1.58-bit model family it says is **9x smaller** than 16-bit counterparts [^35].

---

### Sources

[^1]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2044785261393977612)
[^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2044787072947601796)
[^3]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2044792017553645668)
[^4]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2044856740970402115)
[^5]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2044827705406062670)
[^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2044828148890812538)
[^7]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2044828378147311990)
[^8]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2044828391753551874)
[^9]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2044828473060139208)
[^10]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2044828214867202519)
[^11]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2044861690911850863)
[^12]: [𝕏 post by @kevinweil](https://x.com/kevinweil/status/2044862783947448611)
[^13]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2044861694216900901)
[^14]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2044861695911477643)
[^15]: [𝕏 post by @Alibaba_Qwen](https://x.com/Alibaba_Qwen/status/2044768734234243427)
[^16]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2044780695361290347)
[^17]: [𝕏 post by @iScienceLuvr](https://x.com/iScienceLuvr/status/2044685430323396709)
[^18]: [𝕏 post by @omarsar0](https://x.com/omarsar0/status/2044769798845079665)
[^19]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2044805973085454518)
[^20]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2044805998272196679)
[^21]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2044806021244497964)
[^22]: [𝕏 post by @GoogleAIStudio](https://x.com/GoogleAIStudio/status/2044852335848133113)
[^23]: [𝕏 post by @fal](https://x.com/fal/status/2044879796371349831)
[^24]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2044723995808076275)
[^25]: [𝕏 post by @Similarweb](https://x.com/Similarweb/status/2044682637860573534)
[^26]: [𝕏 post by @Benioff](https://x.com/Benioff/status/2044981547267395620)
[^27]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2044782394214175156)
[^28]: [𝕏 post by @BlackHC](https://x.com/BlackHC/status/2044836601285726663)
[^29]: [𝕏 post by @BlackHC](https://x.com/BlackHC/status/2027892902328971383)
[^30]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2044838334342524965)
[^31]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2044763625680765408)
[^32]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2044763631858909269)
[^33]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044841478913130930)
[^34]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2044841481077465236)
[^35]: [𝕏 post by @PrismML](https://x.com/PrismML/status/2044833023682896134)