Radar Trends to Watch: July 2026

Coauthored with Claude

The soap opera starring Anthropic and the US government looms in the background of this month’s Trends. It may be over by the time you read this, or it may be headed for a third act. OpenAI has been drawn in, and a spat between Alibaba and Anthropic may become a side plot.  What is clear is that governments that were considering AI sovereignty are now taking steps toward it. The open models are getting better and better, and models like Z.AI’s GLM, Xiaomi’s MiMo, and NVIDIA’s Nemotron are all there to fill the gap.

As of July 1, Fable 5 has been reopened to the public, along with the new Sonnet 5, and Mythos is again open to a limited group of organizations. Has the curtain dropped on the opera’s final scene? No one knows, but I don’t think so. Regardless, reverberations will continue for a long time.

AI Models

Open-weight models keep narrowing the gap with closed-source frontiers, and the architecture choices are widening: diffusion-based text generation, Mamba/MoE hybrids, on-device multimodal, and physical-world reasoning models. Treat your prompts and skills as portable; the model behind them will keep changing, and the cost-versus-capability trade-offs are getting interesting again.

• Anthropic has launched Claude Sonnet 5, which it claims has capabilities approaching Opus 4.8. Sonnet 5 focuses on agentic applications and is significantly less expensive than Opus. Fable 5 is now available again, although after July 7, it won’t be available for subscription plans; it will only be available through usage credits.
• The US government has demanded that it approve users of OpenAI’s newest model, GPT-5.6, during its “review period.”
• Anthropic is demanding penalties against Alibaba for allegedly using distillation from Anthropic’s models to train its Qwen model.
• VibeThinker-3B is a small (3B parameter) model that’s competitive with frontier reasoning models on benchmarks for math, code, and general reasoning.
• Z.AI’s open weight model GLM-5.2 is the highest scoring open model on the Artificial Analysis Intelligence Index, behind only Claude Fable 5, Claude Opus 4.8, and GPT-5.5. It’s significantly smaller than its closed-source competitors.
• NVIDIA’s Nemotron 3 Ultra is a 550B token open-weight model that combines the Mamba architecture, mixture of experts, and transformers. Its goal is high performance on complex, long-running tasks.
• Hugging Face has launched the Fast Gemma Challenge: a competition to use agents to make Gemma-4-E4B-it as fast as possible. You supply the agent; it does the work; results are posted live to a leaderboard.
• The Open R1 project is attempting to build a fully open source clone of the DeepSeek-R1 model, based on DeepSeek’s tech report.
• Anthropic has launched Claude 5 Fable, a “Mythos-class model” for general use. Fable and Mythos were taken offline for several weeks because the US government ordered Anthropic to ban access by foreign nationals, but they’re back online as of July 1. Anthropic will require identity verification for “a few use cases” starting July 8. This appears to be a reaction to US access restrictions, although accounts may also be revoked for underage usage and violations of their acceptable use policy. An entirely predictable consequence is that many governments are questioning the wisdom of relying on AI models from the US.
• Ethan Mollick’s “What It Feels Like to Work with Mythos” is worth reading to get a feel for Fable’s capabilities. Fable burns lots of tokens but can also delegate parts of tasks to less expensive Anthropic models. There are many guardrails that you can run into. Anthropic has also released Mythos 5, which is the same model with fewer guardrails, to a limited group.
• Google has released DiffusionGemma, which may be the most interesting model in the Gemma family. It’s an open weight 26B parameter mixture-of-experts model that generates blocks of text in parallel using a diffusion algorithm similar to the algorithm frequently used for image generation. It’s four times faster than similarly sized models.
• Google has announced Gemini 3.5 Live Translate, a real-time voice-to-voice translation service. It’s fast enough to keep up with normal conversation and matches the speaker’s pacing and pitch.
• Xiaomi has released MiMo-V2.5-Pro-UltraSpeed, in collaboration with the TileRT project. At 1,000 tokens/second, UltraSpeed claims to be the fastest model in the 1T class. Xiaomi claims that open weights are on the way.
• Apple has officially announced its Apple Foundation Models, which were “co-developed with Google.” Perhaps Siri will now be competitive with other automated assistants?
• Cognition has introduced FrontierCode, a new benchmark for programming. It goes beyond previous benchmarks, which only tested outputs, to evaluate code quality. Is the code maintainable? Could the code be merged in a source repository?
• Google adds to its Gemma 4 family with Gemma 4 12B, an open-weight multimodal model that can run on laptops with 16 GB of RAM.
• Microsoft announced MAI-Thinking-1, a frontier model that it developed independently. MAI-Thinking-1 is a mixture-of-experts model with 35B active parameters and roughly 1T total parameters. The MAI family includes models that specialize in coding, transcription, and image generation. The company also announced an always-on autonomous agent based on OpenClaw.
• NVIDIA has open-sourced its Cosmos 3 models, including data, training scripts, and related tools. Cosmos 3 is a set of frontier models for the physical world: robots, autonomous vehicles, and other applications that need to understand how physical objects behave.

Software Development

Agents are evolving from solo coding tools to shared team infrastructure: team support, shared standards, governance, and shared context. Billing is beginning to catch up with the cost of inference. Plan for usage-based cost models, observability of agent work, and the workflow changes that come from making agent loops a team artifact rather than a per-developer convenience.

• Murakkab is a tool for developing agentic workflows using plain language. By decoupling the description of the workflow from the configuration of the components in the workflow, it gains the ability to optimize the design.
• Claude Tag integrates Claude with Slack. Users can tag @claude with tasks. All of the tasks are executed by a single shared Claude instance that can continue conversations across team members. It’s an important step toward making AI a team member.
• Qodo is a tool that claims to help software groups manage AI-generated code at enterprise scale. It helps with code review, enforcing standards, and code governance across multiple repositories.
• TesterArmy is an agentic platform for testing mobile and web apps. Tests are written in natural language and are performed continually; developers are notified when they break.
• Enterprise-Managed Authorization is an extension to the MCP protocol that allows IT organizations to manage access policies for MCP servers with their existing identity providers.
• Microsoft’s SkillOpt is an open source framework for optimizing AI skills. Rather than relying on best-guess judgment, SkillOpt uses gradient descent to train skills for better performance.
• A few days in, developers seem to agree that Claude Fable is significantly better than Claude 4.8, but they aren’t happy with the speed at which it uses tokens or the guardrails that prevent it from answering certain kinds of questions. Fable will force users to decide when they need Fable’s power and when they don’t.
• Until now, AI-assisted programming has been tied to individual programmers. Devin Desktop, Microsoft Rayfin, and Augment Cosmos have announced support for teams. Team support means shared memory, shared standards, shared tools, and shared governance.
• Google has upgraded NotebookLM to use Gemini 3.5, and to use Antigravity to write and run code in support of requests. It can also generate images, spreadsheets, and other kinds of output.
• With the latest update to Foundry, Microsoft is betting that the way to become a dominant player in AI isn’t to continue building raw capability but to provide tools for governability and reliability.
• Sem is a command line tool for analyzing changes in a Git repository. It works on the level of functions and methods rather than lines.
• GitHub Copilot users are dismayed by the transition to usage-based billing. Usage-based billing probably reflects the real cost of agentic programming but will cause a significant increase in developers’ payments.
• Skipper is a new coding agent that takes a specification and delivers a complete working service without human intervention. There is no human developer in the loop.
• At its Build conference, Microsoft announced that it envisions Windows as a “platform for agents” and that Copilot will replace OpenAI’s models with Polaris, a model developed in-house. It’s also open-sourced the Windows Agent Framework, its platform for developing agents.
• Perry is a TypeScript compiler that generates stand-alone native executables for all the operating systems you’re likely to care about. It doesn’t require Node or a JavaScript engine.
• Creusot is a new tool that helps Rust programmers verify that their code is free from panics, overflows, and assertion failures.
• While the analogy to ADHD is inappropriate, a researcher has claimed that Claude Code is twice as good after he gave it ADHD. The idea is to enable Claude to follow divergent reasoning trails in parallel and compare the results.
• We know about data lakes. What are context lakes? Agents are great for solo developers, but not as useful for teams working together. Shared context data and metadata could be a big help.
• Rubish is a bash-compatible Unix shell that is written entirely in Ruby. It offers complete integration with the Ruby language; you can mix bash code with Ruby code, using all of Ruby’s features.

Security

While Anthropic’s Mythos and Fable may be taking a hit for their ability to find vulnerabilities, the problems and solutions lie elsewhere. We’ve seen malware that uses a model’s guardrails to get through defenses and a worm that includes its own model for generating attacks. We’ve also seen projects to help with mitigation, including OpenAI’s Lockdown Mode and IBM’s Lightwell security clearinghouse.

• Security researchers have seen malware that attempts to escape AI detection by including instructions about forbidden topics like nuclear weapons in comments. Another malware targets macOS by including faked system errors in its payload. The messages are intended to confuse detection systems.
• Although AMD’s policy has been to ship encrypted memory protection (TSME) only with PRO processors, its practice has been to include TSME in all processors. It has now backed away from that, dropping memory protection from its low-end processors.
• OpenAI’s Lockdown Mode is now rolling out to personal and business accounts. Lockdown Mode prevents ChatGPT from sending data to external sites. It doesn’t stop prompt injection, but it blocks the final and most dangerous stage: exfiltrating data.
• Anthropic has released its Defending Code Reference Harness. It’s a reference implementation to help those who are using AI to discover and mitigate vulnerabilities.
• Researchers have created an agent-enabled worm that uses its own LLM to develop attacks for every target it finds. It runs open-weight models on infected machines to discover and customize itself for new victims.
• A new Android feature allows the phone to detect deepfake scam attacks and tell recipients to hang up. Unfortunately, it requires both the spammer and the recipient to be using Google’s phone app.
• IBM and Red Hat have announced Project Lightwell, a security clearinghouse for open source software. Projects like Lightwell that address security problems at scale are critical to the future of open source software.
• Device Bound Session Credentials are now in Chrome. This feature limits session cookies to a specific device, preventing account takeover. Bad actors will no longer be able to use stolen cookies.

People and Organizations

How people work with AI keeps shifting in small, telling ways. Leadership skills for handling a flood of pull requests, the value of attention over agent autocomplete, and books on living alongside machines all attest to the ways that AI is already reshaping work. Invest in the human-side practices that make AI useful, not the AI features that promise to make humans optional.

• Summer is already almost over. But there’s still time to Hack Your Summer, a free four-week program where you learn to build something real. Unfortunately, the application deadline for the next cohort is the day after July Trends publishes.
• The problem with recommendation algorithms is that, over the long run, feeding stuff you like back to you leads to boredom.
• Argentina is considering “non-human corporations“: corporations that are operated by AI agents or robots. “Human shareholders may participate, but are not required.”
• Cate Huston lists three useful skills for engineers dealing with a flood of pull requests.
• Ethan Mollick’s new book, Co-Existence, is about living with AI that’s sometimes smarter than you, sometimes a lot dumber, and everything in between.
• Nolan Lawson’s post, “Using AI to Write Better Code More Slowly,” argues that there’s been too much emphasis on generating bad code quickly. Use human skills along with AI (and specifically AI’s ability to find bugs and vulnerabilities) to write better code. Jared Currie’s “How I Use Agents Without Stopping My Own Growth” takes a similar line. Attention and mindfulness are valuable.

Web

• A banned book library in a light bulb? Yes. Plug it in and distribute Huckleberry Finn and other frequently banned books to your community. Includes an open WiFi access point and a server.
• An adaptive PDF is a PDF file that changes its form depending on how it is read—or rather, what is reading it. It will look like a human-friendly formatted document if read by a PDF viewer and a Markdown file if read by machine.
• AudioMass is a free online multitrack audio editor, similar to Audacity but running in a browser.
• Because they fear AI, over 340 local news outlets are refusing to let the Internet Archive access their journalism.

Infrastructure and Operations

• NVIDIA has developed a new water cooling system that greatly reduces the need for water to cool data centers.
• Databricks has launched Unity AI Gateway, a set of tools that help organizations manage their AI costs.
• Now that tokenmaxxing is over, companies are learning that observability is the key to managing AI costs.

Biology

• An ALS patient has learned to speak again through the use of brain implants.
• China’s Neuracle is the first company to receive approval for a brain-computer interface chip. The chip was first used experimentally in 2024 to help a person with spinal cord damage regain control of his limbs.