Meta Wants Your AI Glasses to Act on What You See

FEATURED INSIGHT💡

GPT-6 Sol cuts API prices in half

OpenAI released GPT-6 Sol and Luna on September 22, bringing cheaper options to its newest model family. Sol costs $2 per million input tokens and $10 per million output tokens, half the rates for GPT-5.6 Sol under its promotional pricing. Luna costs $0.10 and $0.50, respectively. Those standard API rates apply to prompts of up to 272,000 input tokens; longer prompts and other processing tiers require their own cost comparison.

The savings also depend on how much context an agent can reuse. OpenAI says its improved caching gives a 90% discount on cached input reads and now preserves that reuse when developers change reasoning effort or tool availability. That matters for an agent repeatedly working through the same repository or documents. OpenAI also reports better factuality and coding results, but its factuality test deliberately uses conversations that previously produced errors. It should not be read as a typical customer's error rate.

For teams paying for repeated agent runs, this is a reason to revisit which model handles each task. Compare the cost of a completed, accepted result, including retries and review, before moving an entire workflow to the cheaper tier. Work and Codex received the new models, while the release notes distinguish them from the models offered in Chat. Check the product and model name before assuming a familiar subscription now runs GPT-6.

Some teams never seem to stop moving. They're on Attio, the agentic CRM.

Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.

With Attio, you’ll get:

• Leads automatically prioritised and routed to the right rep

• Expansion and risk signals caught the moment they land

• Follow-ups written in your voice, already there when you arrive

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Get started now

ON THE HORIZON 🌅

Meta wants Muse to act on what your glasses see

Meta plans to bring Muse to its AI glasses in the coming months. At Connect, it described an agent that can act on a product, a wall flier, or a school-supply list in the wearer's view. The new angle is access to the scene around you: a request could begin with what you're looking at, without first photographing it or typing a description. Meta also described voice conversations that continue while Muse handles work in the background.

This remains a planned glasses integration, so the demonstration does not establish everyday reliability. The useful test will be whether Muse identifies the intended object, asks when the request is ambiguous, and completes the right action through a connected service. A purchase based on the wrong package is still the wrong purchase, however convenient the interface feels. For developers, the opportunity is to connect visual context to a narrow, verifiable task, with a clear point for the wearer to approve consequential actions.

LATEST IMPORTANT NEWS 📰

Claude Opus 5.5 promises lower bills for long jobs

Anthropic launched Claude Opus 5.5 on September 22, with standard input and output prices of $4 and $20 per million tokens, each 20% below Opus 5. Cheaper cache reads and reduced token use bring typical token-billed workload costs down about 40%, the company says; that is a workload estimate, not a universal discount. AWS confirms availability through Bedrock and Claude Platform on AWS. Teams running long coding jobs should compare actual token use and accepted results before budgeting around the larger saving.

Microsoft’s Copilot Autopilot will work between prompts

Microsoft announced Copilot Home, Code, and Autopilot on September 25. Home combines conversational and delegated work with editable Office files; Code builds small applications; Autopilot is designed to follow channels and carry recurring work forward without a fresh prompt. Home and Code will begin rolling out through Frontier, while Autopilot expands to private preview at month's end. These are staged releases. Microsoft says the longer-running agent capabilities use usage-based billing, making spending limits and task ownership part of evaluating them.

AlphaFold opens a protein library for 2,800 viruses

NVIDIA and research partners released predicted protein-complex structures for more than 2,800 viruses through the AlphaFold Database. The September 24 announcement describes AlphaFold2 inference accelerated with NVIDIA's BioNeMo software, alongside an open workflow for predicting additional structures. The dataset gives researchers a starting point for investigating how viral proteins interact, including targets with little existing structural information. Predictions carry confidence labels and still need experimental checking; the release supplies hypotheses for laboratory work, rather than evidence that a treatment or vaccine will succeed.

Can Robotics Make Regenerative Farming Scalable?

Greenfield Robotics is on a mission to give farmers alternatives to herbicide-dependent weed control. BOTONY is the beginning of a broader robotic farming system designed for a wider range of applications in the field.

The Reg A+ offering is now live for investors who want to be part of what comes next.

Invest in Robots Replacing Herbicides

^{This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. Please read the Offering Circular and related disclosures before investing. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment.}

FOR THE TECHNICALLY INCLINED 🛠️

RRSI improves AI agents without retraining the model

A new RRSI preprint studies how an agent can improve the software around a fixed language model: its prompts, tools, memory, and control flow. Unrestricted editing can produce a system that memorizes the tasks used to select changes. RRSI limits how many edits a candidate can bundle, records earlier hypotheses and their results, and screens proposals for benchmark-specific shortcuts. It also removes changes that add too little value or too much cost. This makes improvement a sequence of testable engineering decisions.

The released implementation uses Claude Opus 4.8 as the fixed underlying model. Researchers report gains of up to 4.7 percentage points on benchmarks outside the optimization set, plus 30% fewer model tokens than unregularized evolution. These are research results, not proof of unlimited self-improvement. For builders, the useful distinction is whether an edit helps on tasks it never saw during selection. Keep a separate evaluation set, record the cost of each accepted change, and check whether the gain survives beyond the tasks used to develop it.

AI TOOL OF THE DAY 🚀

JetBrains Air lets developers run AI agents inside JetBrains IDEs, inspect proposed changes in context, and leave comments on specific lines before accepting the work.

Your agents work while you sleep

Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.

Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.

Make your first agent