Radar Trends to Watch: July 2026 – O’Reilly

Coauthored with Claude

The cleaning soap opera starring Anthropic and the US authorities looms within the background of this month’s Traits. It could be over by the point you learn this, or it could be headed for a 3rd act. OpenAI has been drawn in, and a spat between Alibaba and Anthropic could turn out to be a facet plot.  What is obvious is that governments that had been contemplating AI sovereignty are actually taking steps towards it. The open fashions are getting higher and higher, and fashions like Z.AI’s GLM, Xiaomi’s MiMo, and NVIDIA’s Nemotron are all there to fill the hole.

As of July 1, Fable 5 has been reopened to the general public, together with the brand new Sonnet 5, and Mythos is once more open to a restricted group of organizations. Has the curtain dropped on the opera’s last scene? Nobody is aware of, however I don’t assume so. Regardless, reverberations will proceed for a very long time.

AI Fashions

Open-weight fashions maintain narrowing the hole with closed-source frontiers, and the structure selections are widening: diffusion-based textual content technology, Mamba/MoE hybrids, on-device multimodal, and physical-world reasoning fashions. Deal with your prompts and abilities as moveable; the mannequin behind them will maintain altering, and the cost-versus-capability trade-offs are getting attention-grabbing once more.

• Anthropic has launched Claude Sonnet 5, which it claims has capabilities approaching Opus 4.8. Sonnet 5 focuses on agentic purposes and is considerably cheaper than Opus. Fable 5 is now accessible once more, though after July 7, it gained’t be accessible for subscription plans; it should solely be accessible by utilization credit.
• The US authorities has demanded that it approve customers of OpenAI’s latest mannequin, GPT-5.6, throughout its “evaluate interval.”
• Anthropic is demanding penalties towards Alibaba for allegedly utilizing distillation from Anthropic’s fashions to coach its Qwen mannequin.
• VibeThinker-3B is a small (3B parameter) model that’s aggressive with frontier reasoning fashions on benchmarks for math, code, and common reasoning.
• Z.AI’s open weight mannequin GLM-5.2 is the highest scoring open mannequin on the Synthetic Evaluation Intelligence Index, behind solely Claude Fable 5, Claude Opus 4.8, and GPT-5.5. It’s considerably smaller than its closed-source opponents.
• NVIDIA’s Nemotron 3 Ultra is a 550B token open-weight mannequin that mixes the Mamba structure, combination of consultants, and transformers. Its purpose is excessive efficiency on advanced, long-running duties.
• Hugging Face has launched the Fast Gemma Challenge: a contest to make use of brokers to make Gemma-4-E4B-it as quick as attainable. You provide the agent; it does the work; outcomes are posted stay to a leaderboard.
• The Open R1 challenge is making an attempt to construct a totally open supply clone of the DeepSeek-R1 mannequin, based mostly on DeepSeek’s tech report.
• Anthropic has launched Claude 5 Fable, a “Mythos-class mannequin” for common use. Fable and Mythos had been taken offline for a number of weeks as a result of the US authorities ordered Anthropic to ban entry by overseas nationals, however they’re again on-line as of July 1. Anthropic will require id verification for “a number of use circumstances” beginning July 8. This seems to be a response to US entry restrictions, though accounts may be revoked for underage utilization and violations of their acceptable use coverage. A completely predictable consequence is that many governments are questioning the knowledge of counting on AI fashions from the US.
• Ethan Mollick’s “What It Feels Like to Work with Mythos” is price studying to get a really feel for Fable’s capabilities. Fable burns numerous tokens however may also delegate elements of duties to cheaper Anthropic fashions. There are various guardrails that you could run into. Anthropic has additionally launched Mythos 5, which is identical mannequin with fewer guardrails, to a restricted group.
• Google has released DiffusionGemma, which stands out as the most attention-grabbing mannequin within the Gemma household. It’s an open weight 26B parameter mixture-of-experts mannequin that generates blocks of textual content in parallel utilizing a diffusion algorithm much like the algorithm continuously used for picture technology. It’s 4 instances sooner than equally sized fashions.
• Google has announced Gemini 3.5 Stay Translate, a real-time voice-to-voice translation service. It’s quick sufficient to maintain up with regular dialog and matches the speaker’s pacing and pitch.
• Xiaomi has released MiMo-V2.5-Professional-UltraSpeed, in collaboration with the TileRT challenge. At 1,000 tokens/second, UltraSpeed claims to be the quickest mannequin within the 1T class. Xiaomi claims that open weights are on the best way.
• Apple has officially announced its Apple Basis Fashions, which had been “co-developed with Google.” Maybe Siri will now be aggressive with different automated assistants?
• Cognition has introduced FrontierCode, a brand new benchmark for programming. It goes past earlier benchmarks, which solely examined outputs, to guage code high quality. Is the code maintainable? May the code be merged in a supply repository?
• Google provides to its Gemma 4 household with Gemma 4 12B, an open-weight multimodal mannequin that may run on laptops with 16 GB of RAM.
• Microsoft introduced MAI-Thinking-1, a frontier mannequin that it developed independently. MAI-Considering-1 is a mixture-of-experts mannequin with 35B lively parameters and roughly 1T whole parameters. The MAI family consists of fashions focusing on coding, transcription, and picture technology. The corporate additionally introduced an always-on autonomous agent based mostly on OpenClaw.
• NVIDIA has open-sourced its Cosmos 3 fashions, together with knowledge, coaching scripts, and associated instruments. Cosmos 3 is a set of frontier fashions for the bodily world: robots, autonomous autos, and different purposes that want to grasp how bodily objects behave.

Software program Growth

Brokers are evolving from solo coding instruments to shared workforce infrastructure: workforce help, shared requirements, governance, and shared context. Billing is starting to meet up with the price of inference. Plan for usage-based price fashions, observability of agent work, and the workflow adjustments that come from making agent loops a workforce artifact fairly than a per-developer comfort.

• Murakkab is a device for growing agentic workflows utilizing plain language. By decoupling the outline of the workflow from the configuration of the parts within the workflow, it positive factors the power to optimize the design.
• Claude Tag integrates Claude with Slack. Customers can tag @claude with duties. The entire duties are executed by a single shared Claude occasion that may proceed conversations throughout workforce members. It’s an necessary step towards making AI a workforce member.
• Qodo is a device that claims to assist software program teams handle AI-generated code at enterprise scale. It helps with code evaluate, imposing requirements, and code governance throughout a number of repositories.
• TesterArmy is an agentic platform for testing cellular and internet apps. Checks are written in pure language and are carried out frequently; builders are notified once they break.
• Enterprise-Managed Authorization is an extension to the MCP protocol that permits IT organizations to handle entry insurance policies for MCP servers with their current id suppliers.
• Microsoft’s SkillOpt is an open source framework for optimizing AI abilities. Reasonably than counting on best-guess judgment, SkillOpt makes use of gradient descent to coach abilities for higher efficiency.
• A number of days in, developers seem to agree that Claude Fable is considerably higher than Claude 4.8, however they aren’t proud of the velocity at which it makes use of tokens or the guardrails that forestall it from answering sure sorts of questions. Fable will drive customers to resolve once they want Fable’s energy and once they don’t.
• Till now, AI-assisted programming has been tied to particular person programmers. Devin Desktop, Microsoft Rayfin, and Increase Cosmos have introduced support for teams. Crew help means shared reminiscence, shared requirements, shared instruments, and shared governance.
• Google has upgraded NotebookLM to make use of Gemini 3.5, and to make use of Antigravity to jot down and run code in help of requests. It could additionally generate photographs, spreadsheets, and different kinds of output.
• With the most recent replace to Foundry, Microsoft is betting that the best way to turn out to be a dominant participant in AI isn’t to proceed constructing uncooked functionality however to supply instruments for governability and reliability.
• Sem is a command line device for analyzing adjustments in a Git repository. It really works on the extent of features and strategies fairly than traces.
• GitHub Copilot customers are dismayed by the transition to usage-based billing. Utilization-based billing most likely displays the actual price of agentic programming however will trigger a major enhance in builders’ funds.
• Skipper is a brand new coding agent that takes a specification and delivers an entire working service with out human intervention. There isn’t a human developer within the loop.
• At its Build conference, Microsoft introduced that it envisions Home windows as a “platform for brokers” and that Copilot will change OpenAI’s fashions with Polaris, a mannequin developed in-house. It’s additionally open-sourced the Home windows Agent Framework, its platform for growing brokers.
• Perry is a TypeScript compiler that generates stand-alone native executables for all of the working techniques you’re more likely to care about. It doesn’t require Node or a JavaScript engine.
• Creusot is a brand new device that helps Rust programmers confirm that their code is free from panics, overflows, and assertion failures.
• Whereas the analogy to ADHD is inappropriate, a researcher has claimed that Claude Code is twice pretty much as good after he gave it ADHD. The thought is to allow Claude to comply with divergent reasoning trails in parallel and examine the outcomes.
• We learn about knowledge lakes. What are context lakes? Brokers are nice for solo builders, however not as helpful for groups working collectively. Shared context knowledge and metadata could possibly be an enormous assist.
• Rubish is a bash-compatible Unix shell that’s written totally in Ruby. It provides full integration with the Ruby language; you may combine bash code with Ruby code, utilizing all of Ruby’s options.

Safety

Whereas Anthropic’s Mythos and Fable could also be taking a success for his or her means to seek out vulnerabilities, the issues and options lie elsewhere. We’ve seen malware that makes use of a mannequin’s guardrails to get by defenses and a worm that features its personal mannequin for producing assaults. We’ve additionally seen initiatives to assist with mitigation, together with OpenAI’s Lockdown Mode and IBM’s Lightwell safety clearinghouse.

• Safety researchers have seen malware that makes an attempt to flee AI detection by together with directions about forbidden subjects like nuclear weapons in feedback. Another malware targets macOS by together with faked system errors in its payload. The messages are supposed to confuse detection techniques.
• Though AMD’s coverage has been to ship encrypted reminiscence safety (TSME) solely with PRO processors, its apply has been to incorporate TSME in all processors. It has now backed away from that, dropping reminiscence safety from its low-end processors.
• OpenAI’s Lockdown Mode is now rolling out to private and enterprise accounts. Lockdown Mode prevents ChatGPT from sending knowledge to exterior websites. It doesn’t cease immediate injection, however it blocks the ultimate and most harmful stage: exfiltrating knowledge.
• Anthropic has launched its Defending Code Reference Harness. It’s a reference implementation to assist those that are utilizing AI to find and mitigate vulnerabilities.
• Researchers have created an agent-enabled worm that makes use of its personal LLM to develop assaults for each goal it finds. It runs open-weight fashions on contaminated machines to find and customise itself for brand new victims.
• A brand new Android function permits the telephone to detect deepfake scam assaults and inform recipients to hold up. Sadly, it requires each the spammer and the recipient to be utilizing Google’s telephone app.
• IBM and Crimson Hat have announced Challenge Lightwell, a safety clearinghouse for open supply software program. Tasks like Lightwell that handle safety issues at scale are essential to the way forward for open supply software program.
• Machine Certain Session Credentials are now in Chrome. This function limits session cookies to a particular gadget, stopping account takeover. Dangerous actors will not be capable to use stolen cookies.

Folks and Organizations

How individuals work with AI retains shifting in small, telling methods. Management abilities for dealing with a flood of pull requests, the worth of consideration over agent autocomplete, and books on residing alongside machines all attest to the ways in which AI is already reshaping work. Spend money on the human-side practices that make AI helpful, not the AI options that promise to make people non-compulsory.

• Summer season is already virtually over. However there’s nonetheless time to Hack Your Summer, a free four-week program the place you be taught to construct one thing actual. Sadly, the appliance deadline for the subsequent cohort is the day after July Traits publishes.
• The problem with recommendation algorithms is that, over the long term, feeding stuff you want again to you results in boredom.
• Argentina is contemplating “non-human corporations“: companies which are operated by AI brokers or robots. “Human shareholders could take part, however are usually not required.”
• Cate Huston lists three helpful abilities for engineers coping with a flood of pull requests.
• Ethan Mollick’s new book, Co-Existence, is about residing with AI that’s typically smarter than you, typically lots dumber, and every part in between.
• Nolan Lawson’s submit, “Using AI to Write Better Code More Slowly,” argues that there’s been an excessive amount of emphasis on producing dangerous code shortly. Use human abilities together with AI (and particularly AI’s means to seek out bugs and vulnerabilities) to jot down higher code. Jared Currie’s “How I Use Agents Without Stopping My Own Growth” takes the same line. Consideration and mindfulness are priceless.

Internet

• A banned book library in a light bulb? Sure. Plug it in and distribute Huckleberry Finn and different continuously banned books to your group. Consists of an open WiFi entry level and a server.
• An adaptive PDF is a PDF file that changes its form relying on how it’s learn—or fairly, what’s studying it. It’s going to appear like a human-friendly formatted doc if learn by a PDF viewer and a Markdown file if learn by machine.
• AudioMass is a free on-line multitrack audio editor, much like Audacity however working in a browser.
• As a result of they worry AI, over 340 native information retailers are refusing to let the Web Archive entry their journalism.

Infrastructure and Operations

• NVIDIA has developed a brand new water cooling system that significantly reduces the necessity for water to chill knowledge facilities.
• Databricks has launched Unity AI Gateway, a set of instruments that assist organizations handle their AI prices.
• Now that tokenmaxxing is over, firms are learning that observability is the important thing to managing AI prices.

Biology

• An ALS affected person has learned to speak once more by using mind implants.
• China’s Neuracle is the primary firm to obtain approval for a brain-computer interface chip. The chip was first used experimentally in 2024 to assist an individual with spinal twine injury regain management of his limbs.