Hi everyone 👋
Welcome back to AI Agent Weekly. This week, Anthropic launched its newest top model, OpenAI opened cyber tools to more defenders, and Mozilla announced a new open AI client you can run yourself. Let's get into it.
ANTHROPIC LAUNCHES CLAUDE OPUS 4.7: BETTER CODING, SHARPER VISION, SMARTER WORK
What's Happening: Anthropic released Claude Opus 4.7, its newest high-end model. It is now available to everyone. The model is better at hard coding tasks, sees images in higher detail, and follows instructions more carefully.
Report Includes:
Better at complex code: Opus 4.7 solves more tough software tasks than Opus 4.6, with gains on benchmarks like SWE-bench and Terminal-Bench.
Sharper vision: It can read images up to 2,576 pixels wide, helping with diagrams, screenshots, and technical drawings.
Smarter memory: It remembers context across long sessions, so you spend less time repeating yourself.
Safer by design: Cyber features are limited on purpose. Security pros can apply for special access through the Cyber Verification Program.
Same price: $5 per million input tokens, $25 per million output tokens.
Why It Matters: Opus 4.7 is a direct upgrade for developers who need a model that can handle long, complex tasks with less supervision. The better vision and memory mean it can work more like a real teammate. The careful cyber safeguards show Anthropic is testing safety steps before releasing even more powerful models.
OPENAI EXPANDS TRUSTED ACCESS FOR CYBER DEFENSE
What's Happening: OpenAI is opening its Trusted Access for Cyber program to more verified defenders. Approved users can now use GPT-5.4-Cyber, a version of GPT-5.4 with fewer restrictions for legitimate security work.
Report Includes:
GPT-5.4-Cyber: A special model fine-tuned for defensive security tasks like reverse engineering and vulnerability research.
Three guiding ideas: Democratized access (clear rules for who gets in), iterative deployment (learn and improve over time), and ecosystem resilience (support the defender community).
Easy sign-up: Individuals verify at chatgpt.com/cyber. Teams contact their OpenAI rep.
Built on real results: Codex Security has already helped fix over 3,000 high-severity bugs.
Why It Matters: Cyber threats are growing fast. This program lets skilled defenders use powerful AI tools while keeping safeguards in place. It shows OpenAI is balancing broad access with responsible controls - a model other labs may follow.
QWEN RELEASES QWEN3.6-35B-A3B:AN OPEN-SOURCE FOR ALL YOUR NEEDS
What's Happening: Alibaba's Qwen team open-sourced Qwen3.6-35B-A3B, a lightweight MoE model with 35 billion total parameters but only 3 billion active at a time. It is built for agentic coding and repository-scale work, designed for teams that need control, efficiency, and results.
Report Includes:
Open weights for full control: Free to download, use, and modify. Teams can host on their own servers or in private clouds to meet data governance and compliance needs.
Strong coding for real workflows: Matches larger models on agentic coding benchmarks and repo-level tasks, helping teams automate code reviews, refactoring, and legacy modernization.
Multimodal support for technical docs: Includes a vision encoder to process diagrams, architecture charts, and UI mockups, enabling richer context for development pipelines.
Easy integration: Compatible with common MLOps tools and identity systems, reducing time to production.
Why It Matters: Teams need AI that fits their security, cost, and workflow requirements. Qwen3.6-35B-A3B gives groups a high-performance, open model they can customize, audit, and deploy on their own infrastructure. This reduces vendor risk, lowers long-term costs, and accelerates adoption of agentic AI in regulated or sensitive environments. For engineering leaders evaluating build versus buy, this model offers a credible path to own the stack while staying competitive.
OPENAI UPDATES CODEX: NOW WORKS ACROSS YOUR COMPUTER AND TOOLS
What's Happening: OpenAI released a major Codex update. It can now use apps on your computer, work in a built-in browser, generate images, remember your preferences, and handle repeat tasks over time.
Report Includes:
Computer use: Codex can see your screen, click, and type - helpful for testing apps or working in tools without APIs.
In-app browser: Comment on web pages to give Codex precise instructions for frontend work.
Image generation: Uses gpt-image-1.5 to create and edit visuals inside coding workflows.
90+ new plugins: Connect to Jira, GitLab, Microsoft Suite, and more for richer context.
Memory and automations: Codex can remember your preferences and schedule future work.
Why It Matters: Codex is moving beyond writing code to becoming a full development partner. By working across your whole workflow - code, browser, apps, and docs - it reduces context switching and helps developers ship faster.
GOOGLE LAUNCHES GEMINI 3.1 FLASH TTS: EXPRESSIVE AI SPEECH WITH FINE CONTROL
What's Happening: Google released Gemini 3.1 Flash TTS, a new text-to-speech model in public preview. It supports 70+ languages and lets you control tone, style, and pacing using simple audio tags.
Report Includes:
200+ audio tags: Add tags like [whispers], [fast], or [excitement] directly in your text to shape the output.
High quality: Ranks first on public TTS leaderboards with Elo score of 1211.
SynthID watermark: All output includes a hidden mark to identify AI-generated audio.
Easy start: Available now in Google AI Studio and Vertex AI.
Why It Matters: Great voice output is key for accessible apps, audiobooks, and enterprise alerts. Gemini 3.1 Flash TTS gives developers fine-grained control without complex setup - a big step toward natural, customizable AI speech.
ANTHROPIC ADDS ROUTINES TO CLAUDE CODE: AUTOMATE WORK WHILE YOU SLEEP
What's Happening: Anthropic launched Routines in Claude Code (research preview). A routine is a saved automation - prompt, repo, and connectors - that runs on a schedule, via API, or in response to events like GitHub PRs.
Report Includes:
Three trigger types: Scheduled (hourly, nightly), API-triggered, or GitHub webhook-based.
Runs in the cloud: No need to keep your laptop open.
Common uses: Triage bugs nightly, verify deploys, auto-review PRs, or port code between languages.
Usage limits: Pro users get 5 routines/day, Max gets 15, Team/Enterprise gets 25.
Why It Matters: Automation is powerful, but setting up cron jobs and infrastructure is tedious. Routines package Claude Code automations into simple, reusable workflows. This lowers the barrier to building reliable, always-on agent systems.
OPENAI INTRODUCES GPT-ROSALIND FOR LIFE SCIENCES RESEARCH
What's Happening: OpenAI launched GPT-Rosalind, a model built for biology, drug discovery, and medical research. It helps scientists review literature, plan experiments, and analyze complex data.
Report Includes:
Built for science: Better at chemistry, protein design, genomics, and using scientific tools in multi-step workflows.
Strong results: Leads on BixBench, a real-world bioinformatics benchmark. Outperforms GPT-5.4 on 6 of 11 LABBench2 research tasks.
Trusted access: Available to qualified U.S. enterprise customers through a safety review process.
Free plugin: A Life Sciences Research Plugin for Codex connects to 50+ scientific databases and tools.
Why It Matters: Drug discovery takes 10-15 years. AI that helps scientists move faster through early research could speed up breakthroughs. GPT-Rosalind shows how domain-specific models can add real value in high-stakes fields.
MOZILLA ANNOUNCES THUNDERBOLT: OPEN-SOURCE, SELF-HOSTED AI CLIENT
What's Happening: Mozilla's MZLA subsidiary announced Thunderbolt, an open-source AI client you can self-host. It gives organizations full control over their data, models, and AI infrastructure.
Report Includes:
Bring your own models: Connect to commercial APIs, open-source models, or local deployments.
Enterprise integrations: Works with deepset's Haystack for agent orchestration and RAG pipelines.
Cross-platform: Native apps for Windows, macOS, Linux, iOS, and Android.
Security first: Self-hosted option, optional end-to-end encryption, and device-level controls.
Why It Matters: Many organizations want AI benefits without giving up control. Thunderbolt offers a sovereign alternative to cloud-only AI clients. By being open-source and self-hostable, it lets teams build AI systems that fit their security and compliance needs.
Thanks for reading.
See you next week with more AI agent updates.
— Rakesh's Newsletter
