Welcome back. There's a new ceiling for agentic performance. Anthropic just dropped Opus 4.7 and it's also the stepping stone to Mythos, the more powerful model they're still keeping behind closed doors for safety reasons.

Also: How to build the right AI stack in 2026, six Claude Code team tips for mastering Opus 4.7 and find out why data from dead startups is now a goldmine for AI labs.

Anthropic unveils Claude Opus 4.7 with powerful new tools for devs: The AI lab just dropped its latest flagship model, designed to tackle complex engineering tasks with minimal oversight. Opus 4.7 is more rigorous with long-running workflows, follows instructions with more accuracy, and self-verifies its work before finishing. Vision capabilities also got a major upgrade, now supporting images at over three times the previous resolution.

OpenAI arms devs with its most capable Codex upgrade: The team behind ChatGPT just gave its coding agent the ability to see, click, and type across your Mac, allowing multiple agents to run in parallel. A new in-app browser lets you leave comments directly on pages to streamline frontend development. This update also introduces image generation via gpt-image-1.5, persistent memory that remembers your preferences, and over 90 new plugins.

Alibaba drops a lean open model for coding agents: The Chinese tech giant released Qwen3.6-35B-A3B, a mixture-of-experts model that packs 35B total parameters but only uses 3B for inference. Even with its small footprint, it outperforms the dense Qwen3.5-27B in coding and rivals Claude Sonnet 4.5 in multimodal tasks. You can find the weights on Hugging Face and ModelScope, and it integrates directly with Claude Code, OpenClaw, and Qwen Code.

For 10+ years, AI Council has been gathering the world's top AI infrastructure minds to share what’s working (and where we’re headed next). 

You’re invited for 3 days of high-quality discourse with 1,200+ technical experts, including office hours, small groups, and zero marketing keynotes. 

Speakers include:

…and more. Takes place May 12-14 in San Francisco.

See the full speaker lineup here and use code THECODE20 for 20% off.

The coordination gap. As AI agents handle more complex tasks, the question has shifted from whether you should use them to how you should orchestrate them. Anthropic recently published a guide breaking down five coordination patterns (see image above), and the pick depends on one thing: how much your agents need each other.

Sophistication backfires. Many teams over-engineer their setups by choosing patterns based on complexity rather than fit. Anthropic's advice is to start with the simplest option that could work and only add complexity when it breaks. For most use cases, that means the first two patterns on the list (see image above).

Proof is in the kernel. Cursor and NVIDIA recently showed what the right pattern choice looks like in practice. They pointed a multi-agent system at CUDA kernel optimizations for Blackwell GPUs, the kind of assembly-level tuning that normally takes specialized engineers months. A planner agent distributed and rebalanced work across autonomous workers, coordinated by a single markdown file. Three weeks later, the system delivered a 38% speedup.

Before you build. Ask whether your agents need each other's work while it's still in progress. If so, shared state. Otherwise, keep them independent. Anthropic's full guide walks through when to transition between patterns.

API costs can really spiral when you're dealing with high-volume tasks. Google DevRel engineer Patrick Loeber recently shared a simple, one-line fix that cuts Gemini pricing in half for any jobs that don't need an immediate response.

Just add service_tier: "flex" to the config of any “generate_content” call:

Flex cuts your Gemini costs by 50% by using Google’s spare capacity. You'll trade a 1 to 15 minute delay for massive savings, like Flash-Lite at just $0.125 per million input tokens.

It's ideal for background tasks, batch jobs, or nightly data processing. Stick to the standard tier for real-time needs and offload everything else to Flex. Check the full breakdown in the docs.

How to use Hooks effectively in coding agents: This tutorial covers how to use hooks in AI coding agents like Copilot or Claude Code. You'll learn how to intercept the agent's lifecycle, inject custom context, and use pre-tool hooks to ensure the AI writes clean, linted code before any changes are saved.

HyperFrames (by HeyGen): A new framework for generating AI videos straight from code. It allows AI agents to whip up high-quality MP4s using standard HTML, CSS, and JavaScript.

Designing synthetic datasets for the real world (by Google): Specialized AI struggles because real-world training data is often scarce, private, and expensive. Simula solves this by using reasoning models to systematically design entire synthetic datasets, proving that structured, high-quality data generation heavily outperforms simpler baselines.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 240K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team