That is not because benchmarks are useless. They are useful. But many of them are still optimized around narrow tasks, coding-heavy environments, academic-style reasoning problems, or synthetic prompts that do not look much like how knowledge workers actually use AI systems. The more practical question is different:

If I connect an AI agent to my company’s data, give it tools, and ask it to produce a useful piece of work, how close does it get to something I can actually use?

That is the question behind the Diligence Stack Agent Bench.

The benchmark is built around a view we have had for a while: outside of coding, one of the most common enterprise agent patterns will not be “let the model roam the open internet and hope for the best.” It will be something more structured. Companies will create MCP servers, internal tools, and knowledge systems that connect their proprietary data to applications like ChatGPT, Claude, and their own internal agents.

That means agent performance is not just about the model. It is also about the surrounding infrastructure: the retrieval system, the quality of the knowledge base, the tools available to the agent, the artifact-generation environment, and the cost structure of the workflow.

The model matters enormously. But the system around the model increasingly matters just as much.

This exercise mirrors the evaluation process we believe many organizations are undertaking, and many more will undertake, as enterprise AI moves from experimentation to adoption. The decision is not simply which model scores highest. Organizations must decide which models to adopt, which platforms or agent systems will route work among them, and how to manage token costs without compromising output quality. Evaluating models inside a realistic knowledge-work environment is therefore not an academic exercise; it is a practical proxy for how enterprises will select, deploy, and govern AI systems.

My Experience With the GPT-5.6 Series

I have been tasking GPT-5.6 Sol with demanding work for about a month, and it is incredible. It has an exceptional understanding of how to use and control subagents, it is remarkably intelligent, and, most importantly, it keeps working. When you give it a task, it continues until that task is complete instead of stopping at a plausible first draft.

I have only had access to GPT-5.6 Terra and Luna for a few days, but the early results are equally compelling. For knowledge work, Terra is undefeated on the combination of price, speed, and value.

Luna may be the best everyday model, period. It is better than the open-weight models we have tested while being significantly cheaper. On factuality-focused work, it can outperform Claude at a fraction of the price. The tradeoff is presentation: if you need a polished final artifact, you may have to do some additional work to make the output look better.

With Luna and Terra in the mix, frontier-level performance is finally affordable enough that you do not have to worry about the price on most tasks. Most of the daily work we ran through these models cost cents rather than dollars, while still producing usable, factually accurate results. That is not true of most models in the same price range, where lower cost usually comes with a clear drop in reliability. This is the first time token budgeting starts to look like a problem that can be solved by better, cheaper models rather than tighter limits: use capable models like Luna and Terra for the majority of work, and reserve the most expensive systems for the smaller set of tasks that truly need them.

If you are looking for quick factual answers and research, I have not found a better model. When that research needs to become a polished, actionable deliverable, you can feed it to GPT-5.6 Terra or Sol and have the stronger model synthesize and present it. That gives you the best of both worlds at a fraction of the price.

Overall, I would take the GPT-5.6 series over Claude for these workflows. Based on the pricing I have seen, I expect it can deliver equal or better work for roughly one-fifth to one-tenth of the price.

Why We Built the Benchmark

The Diligence Stack Agent Bench exists because we wanted to evaluate models and agents in a practical knowledge-work environment.

The core question is not whether a model can answer a trivia question, summarize a clean document, or solve a toy reasoning problem. The question is whether a model can take a messy, real-world knowledge-work task and produce a useful first draft.

For example:

If I ask an agent to build a financial model with a specific set of variables, source the assumptions from internal research, produce an Excel file, write a PDF report, and explain the confidence level behind the output, can it give me something I can actually use?

That is the bar.

In most knowledge-work environments, AI outputs are not replacing the entire workflow yet. They are usually the first step. A model or agent creates a draft, model, memo, deck, or report that a human reviews, edits, and improves. But that first draft still matters. The better it is, the less human cleanup is required. In some cases, if the draft is strong enough, it may be close to final.

That is what this benchmark tests: how close a model or agent can get to a complete, usable work product when given access to internal data and a computer-like tool environment.

The Two-Pronged Setup

When we started building the agent system that became the foundation of this benchmark, we began with a simple question: what does a knowledge-work agent actually need access to?

For Diligence Stack, the answer was three core data sources:

1. Historical Creative Strategies research and internal notes
2. Professional investment-grade research collected over time
3. Financial models and related structured files

There are a few obvious ways to give an agent access to this kind of data. You can upload files to Box, Google Drive, Dropbox, or another document system and let the connector handle search and retrieval. We tried versions of that. It works. But it is not ideal.

The problem is that most enterprise data was not designed for agents. PDFs have headers, footers, watermarks, irrelevant boilerplate, broken tables, embedded images, and inconsistent formatting. Presentations are even messier. Audio and video are generally worse. A basic connector can search over that content, but it is rarely optimized for high-quality agent retrieval.

So we built a better pipeline.

The Diligence Stack Knowledge Base Pipeline

The knowledge base pipeline is designed to structure unstructured data so agents can use it efficiently.

At a high level, it takes uploaded files, cleans them, splits them where appropriate, removes junk, extracts useful structure, and makes the content easier for agents to search and reason over. That means removing things like watermarks, repeated headers, useless page artifacts, and other noise that makes retrieval worse.

When images appear in files, the system generates detailed descriptions of those images so the agent can reason over charts, diagrams, screenshots, and visual content. The original files remain available, but the agent is no longer forced to treat them as opaque blobs.

The system is designed to handle essentially every major file format, including PDFs, documents, spreadsheets, slides, images, audio, and video.

In practice, that means you could upload a four-hour Zoom recording, 200 PDFs of old emails, a folder of financial models, an audio recording of your dog barking, and a pile of research notes, and the agent should still be able to find and use the relevant information.

Under the hood, the system uses Gemini embedding models from Google for multimodal embedding, Convex for database storage, Gemini 3.1 Flash Lite for data processing, and a pipeline architecture designed by GPT-5.6 Sol to be robust across different downstream models.

The point is not just to make one model perform better. The point is to make the whole agent environment better.

Why This Matters

This matters because it lets us give any agent access to high-quality internal knowledge through MCP or through our own agent environment.

Instead of asking a model to search raw files in a generic document store, we can provide a purpose-built knowledge base that has already been cleaned, structured, embedded, and optimized for agent retrieval.

That changes the economics and performance of the workflow.

In our testing, searching through the Diligence Stack knowledge system instead of using the more generic agent search pattern saved roughly five minutes on simple requests. We do not yet have fully audited cost numbers across every configuration, but the likely savings on large-model requests are meaningful, potentially in the range of several dollars per request depending on the model, search pattern, and context size.

That adds up quickly.

Faster retrieval means lower latency. Better retrieval means fewer wasted tool calls. Cleaner context means lower token usage. Better source selection means stronger outputs. The economics are not subtle: better infrastructure makes agents cheaper, faster, and more useful.

The Diligence Stack Agent

We then built this knowledge system into our own agent.

The Diligence Stack agent has access to its own computer-like environment, tools to create documents, tools to create spreadsheets and financial models, access to the full knowledge bases, and the ability to download and inspect original files. It is designed for longer-horizon work, not just short chat responses.

That distinction matters.

A normal chatbot can answer a question. A useful knowledge-work agent needs to investigate, retrieve, synthesize, model, format, check its work, and produce artifacts. It needs to behave more like a junior analyst with tools than a text box with a memory.

In the Diligence Stack Agent Bench, this setup performed well. In one benchmark run, our internal agent outperformed Claude Cowork running Claude 5 Fable on the same task, while also coming in at a lower price. Creative name for the benchmark? Maybe not. Useful signal? Absolutely.

Why the Harness Matters

One of the clearest lessons from this testing was that the model is only one part of the result. The harness—the tools, retrieval system, context management, execution environment, and revision loop wrapped around the model—can materially change both output quality and cost.

We tested this directly with Claude 5 Fable. To make the comparison as controlled as possible, we gave Cowork access to identical tools through MCP and the exact same skills available to our Diligence Stack harness. In this benchmark run, Claude 5 Fable still produced a better result inside our harness than it did inside Claude Cowork. The Cowork run scored worse and cost more, while our harness used the same model more efficiently and produced a stronger deliverable.

That does not mean Cowork will be worse for every task, or that one run settles the question. It does mean model comparisons need to account for the environment in which the model is operating. A strong model inside an inefficient harness can waste tool calls, consume more tokens, retrieve weaker context, and still produce a worse final output. A well-designed harness can make the same model both cheaper and more useful.

What the Benchmark Measures

The benchmark is designed around practical output quality.

We are not asking, “Which model sounds smartest?” We are asking, “Which model produces the most useful work product when connected to internal data and tools?”

The grading framework evaluates several dimensions:

Prompt and deliverable completion. Did the model actually do what was asked? Did it produce the requested files, sections, analysis, and supporting materials?

Factual accuracy and verification. Are the claims correct? Are the numbers internally consistent? Does the model avoid hallucinating facts, companies, timelines, or financial assumptions?

Source grounding and priority. Did the model use the right sources? Did it prioritize the Diligence Stack knowledge base and internal materials before reaching for weaker external context? Did it cite and trace its assumptions appropriately?

Analysis and actionability. Did the output merely summarize information, or did it synthesize it into something useful? Does the report help a decision-maker understand what matters, what is uncertain, and what to do next?

Artifact quality and auditability. Are the PDF, Excel file, model, or other deliverables usable? Are they readable, well-structured, and traceable? Can a human inspect the assumptions and understand how the output was created?

Tool strategy and robustness. Did the agent use tools intelligently? Did it search efficiently, inspect the right files, recover from missing information, and avoid wasting time on low-value steps?

Cost and token efficiency. How expensive was the run? How much did the model spend to reach the result? Cost is tracked separately from quality because the cheapest bad answer is still a bad answer, but price matters once output quality clears a usable threshold.

In the current scoring framework, the final quality score is weighted most heavily toward completion, factual accuracy, source grounding, and analytical usefulness. Design, auditability, tool behavior, and cost efficiency also matter, but the benchmark intentionally does not let a cheap model win if the output is not useful.

That is the right tradeoff. In real knowledge work, the highest-cost outcome is not paying a few extra dollars for a model. The highest-cost outcome is a polished but wrong deliverable that wastes a senior person’s time.

How We Think About Pricing

Pricing is part of the benchmark, but it is intentionally a smaller part of the weighted score.

The reason is simple: output quality matters most. A cheap run that produces a weak, incomplete, or inaccurate deliverable is not actually cheap. It just moves the cost downstream to the human who has to fix it. In knowledge work, the expensive failure mode is not spending a few more dollars on a model. The expensive failure mode is trusting a polished but wrong output, or getting a draft that still requires hours of senior cleanup.

So we treat price as an important constraint, not the primary objective.

The goal of the benchmark is to help users understand which models produce the best outputs, and then decide which of those outputs are worth the price for their specific workflow. For some tasks, the highest-scoring model may be worth paying for because it gets meaningfully closer to a usable final deliverable. For other tasks, a slightly lower-scoring model may be the better choice if it delivers 90% of the quality at a much lower cost.

That is why we look at both absolute quality and cost-adjusted performance. The right answer is not always “use the cheapest model” or “use the highest-scoring model.” The right answer is usually: find the best output quality available at the price you are willing to pay.

This is especially important for agentic workflows because costs can compound quickly. Long-running agents search, retrieve, reason, inspect files, generate artifacts, and revise outputs. A model that is slightly more expensive per token can still be the better economic choice if it uses tools more efficiently and produces a cleaner first draft. Likewise, a cheaper model can be attractive if the task is lower stakes, more repetitive, or easier for a human to review.

In other words, price matters. But price only matters after the output clears the bar of being useful.

Why We Didn’t Include Open Models

We also tested open models extensively, but did not include most of them in the current benchmark results.

The reason is straightforward: almost all of the open models we tested produced materially worse outputs while also costing more to run in this agent environment. That is a difficult combination to recommend. Lower quality can be acceptable when it comes with a meaningful cost advantage, but lower quality at a higher price does not offer users a useful tradeoff.

DeepSeek V4 Pro was the one clear outlier. It was far cheaper than the other open models we tested, although the quality tradeoff is visible in its benchmark score. For lower-stakes, high-volume, or highly reviewable tasks, that price difference may still make it interesting.

This is not a permanent exclusion. We will keep testing open models and add more of them as their performance and economics improve. For this version of the benchmark, however, including a longer list of models that were both weaker and more expensive would not have helped readers make better decisions.

Where GPT-5.6 Fits

The GPT-5.6 series performed strongly in this environment because the task rewarded more than raw fluency. It rewarded long-context handling, source discipline, synthesis, spreadsheet construction, report writing, and the ability to produce usable artifacts in a single run.

GPT-5.6 Sol and GPT-5.6 Terra were the strongest quality leaders in our testing, with GPT-5.6 Luna standing out as a strong value option. The key takeaway is not simply that one model “won.” It is that stronger models produced more complete first drafts, handled ambiguity better, and were more effective at turning messy internal data into structured outputs.

That is exactly what matters for agentic knowledge work.

There is also a broader point here: as models improve, the benchmark ceiling rises. Tasks that previously required a human analyst to spend hours gathering sources, cleaning up assumptions, creating tables, and writing the first report draft can now be pushed much further in a single agent run. The output is still not perfect. But the baseline is moving fast.

What We Learned

The biggest lesson from the benchmark is that model quality and system design compound.

A better model inside a weak retrieval environment can still waste time and miss key information. A strong knowledge base paired with a weaker model can produce decent results, but may still fall short on synthesis, judgment, and artifact quality. The best results come from pairing strong models with purpose-built data infrastructure and tools.

That is the direction we think enterprise AI is heading.

The winning setup will not be a generic chatbot pointed at a file dump. It will be a structured knowledge system, exposed through MCP and internal tools, connected to high-quality models, and evaluated against real work products.

That is the premise behind Diligence Stack Agent Bench.

We are not trying to measure whether a model can perform well in an artificial sandbox. We are trying to measure whether it can do useful work in the kind of environment businesses are actually building.

And on that front, GPT-5.6 looks like a meaningful step forward.

The practical readout is simple: for knowledge-work agents, the gap between “interesting demo” and “usable first draft” is closing. The models are getting better, but the infrastructure around them is what turns that improvement into something businesses can actually use.

I am not much of a controller person, and gaming handhelds have never fit into my life. I have owned a Nintendo Switch and used a few Windows/Steam handhelds, but they always felt like a compromised way to play games I could run better somewhere else. The battery was fine, the controls were fine, the performance was fine. None of them made me want to keep using the device after I had done the testing.

The Claw did. After a few days with Spider-Man 2, I stopped thinking about what I was giving up and started thinking about the game. That is the entire point of this category, and this is the first handheld I’ve used that nailed it for me.

The larger point is not that one handheld changed my mind. This category now feels mature enough to grow beyond the people who were already sold on it. The performance is there. The controls are there. Windows has a controller-first interface without giving up the PC game library. Thunderbolt turns the same device into a docked gaming system or a normal PC. These products no longer feel like experiments.

That maturity is arriving while gaming hardware gets more expensive. The tools and process complexity behind leading-edge silicon cost more, memory and NAND supply are tight, and OLED and high-refresh displays remain premium parts. A desktop or gaming laptop built around the hardware people now expect is not heading back toward cheap either.

That makes flexibility more valuable. A handheld like this can be the machine in your hands, the gaming system connected to a TV, and the Windows PC on a desk. It brings the games you already own across Xbox, Epic, Steam, and other stores. As every route into high-end gaming costs more, having more ways to use the same hardware makes the whole category more compelling.

There is still a catch. MSI lists this 32GB/1TB model for $1,799. I can rave about the product and still tell most people not to spend $1,799 on a handheld. Both things are true. I care more about the product here; others can care more about the value. The Claw is not the mainstream handheld. It is proof that the category is ready for more products, more prices, and more people.

My excuse to play Spider-Man 2

Spider-Man 2 is a perfect game for this machine. It is demanding, it looks fantastic, and its movement and combat work naturally on a controller. More importantly for me, it is a game I had wanted to play for years. A portable device that I needed to use to review (and more importantly, try to understand Intel’s G series chip here) is a perfect excuse to justify playing this game.

I ran the game at 1920×1200 on the High preset with ray tracing, V-sync, and frame generation turned off. The display was set to 120Hz and XeSS upscaling was enabled. Most of my gameplay sat around 45 to 50 fps, with some scenes climbing into the 50 to 60 fps range. That is enough for this kind of game. It felt smooth, the controls felt immediate, and I never found myself reaching for the settings menu instead of playing.

I want to be careful with the inevitable console comparison. This does not mean the Arc G3 Extreme is a PlayStation 5 in your hands. The PC and console versions use different settings, the exact XeSS mode matters, and an average frame-rate counter does not tell you about frame pacing or image quality. What I can say is simpler and more useful: a demanding former PlayStation exclusive ran well enough on an 8-inch handheld that I restarted the game and kept playing it.

I spent some time trying to figure out what a comparison would be for this performance. After a few minutes of searching, this compares pretty well to an RTX 4070 laptop; it’s within the same ballpark. I don’t mean to say this as a bad thing or saying it’s better or anything, but rather the RTX 4070 laptop chip is a really good chip! There are really good gaming laptops with this GPU. The Claw 8 with Intel G3 Extreme is close to it. That’s really cool.

The result held up elsewhere. In the built-in Cyberpunk 2077 benchmark, the Claw returned about 44 fps at 1200p on the High preset across several runs. I did not spend an afternoon finding the perfect low-medium-custom soup to inflate the number. On an 8-inch screen I would rather start with a sensible preset, turn on the right upscaler, and play.

Frame generation is available if you want more visible smoothness, and this form factor makes a good case for it. A handheld has a hard power ceiling. If the GPU can render a solid base frame rate and use generated frames to make motion look smoother on a 120Hz screen, that is a good trade in a single-player game. Generated frames are not the same as rendered frames, though, and they do not magically improve input latency.

One thing worth being precise about is how frame generation interacts with input latency, because the bigger frame-rate number oversells the responsiveness. A natively rendered 120fps game samples your input 120 times a second, so the controls feel as immediate as that frame rate suggests. A game running at 60fps samples input half as often, so its latency is roughly double. If you take that 60fps game and use frame generation to reach 120fps, the motion looks like 120fps, but your input is still tied to the 60fps base frames. The generated frames are interpolated, so the technique actually adds a little latency on top of that 60fps baseline rather than removing any. The point is that frame gen buys you a smoother-looking image, not native-120fps responsiveness. As long as the base frame rate is solid, which it is here, I have no problem with fake frames.

Arc G3 Extreme is built for this job

The Intel Arc G3 Extreme is based on Panther Lake, but Intel did more than put a laptop chip in a smaller box. Arc G3 is Intel’s first family of purpose-built handheld SoCs, and the parts were rebalanced around what a game needs inside an 8W to 35W power envelope.

The CPU has two performance cores, eight efficiency cores, and four low-power efficiency cores. The GPU is an Arc B390 based on Xe3 with 12 Xe-cores, 96 XMX engines for AI work such as XeSS, 12 ray-tracing units, and 16MB of cache. The compute tile is built on Intel 18A. There is enough CPU here to feed a game, but the design is plainly biased toward graphics.

That matters because a handheld has one small pool of power and heat to divide. Every watt the CPU takes is a watt the GPU cannot use. Intel’s Intelligent Bias Control manages that split, and at 14W or below it can park the two P-cores entirely so the E-cores handle the CPU side while the GPU keeps more of the budget. This is the kind of mechanism that disappears when it works. You do not think about core parking while swinging through Manhattan. The frame rate holds together.

Intel’s own testing says Arc G3 Extreme at 35W averaged 42% faster than the Ryzen AI Z2 Extreme at 35W across 36 games at 1080p High, using 2x upscaling where supported. Intel also says its chip at 17W roughly matched the Z2 Extreme at 35W across that same test set. Those are Intel’s numbers from pre-production systems, not mine, so treat them accordingly. The direction matches what I felt using the Claw, but I did not run a controlled 36-game comparison of my own.

XeSS is the other half of the platform. Super Resolution lets the GPU render from a lower internal resolution and reconstruct the image for the 1200p screen. Multi-Frame Generation can add up to three generated frames between traditionally rendered ones in supported games. On a desktop with a giant GPU, these features can feel optional. On a handheld, they are how you turn a limited power budget into a better-looking and smoother game. See the above point on input lag and frame gen, it’s still great!

Intel finally has the CPU, GPU, process technology, packaging, drivers, and upscaling stack working toward one clear product. That is why Arc G3 feels more important than another integrated-GPU bump. It is a complete handheld platform.

MSI built the right hardware around it

The chip would not matter if the machine were miserable to hold. MSI did a good job here.

The Claw has an 8-inch 1920×1200 IPS touchscreen with a 48 to 120Hz variable refresh range. I initially wrote down that it was OLED because it looked that good. It is not. Color and contrast are still excellent, VRR keeps sub-60 fps games looking smoother, and 1200p is the right resolution for this size. I never wanted more pixels.

The textured plastic feels better than I expected, and the flared grips make the 785g body easier to hold than the number suggests. The Hall-effect sticks and triggers feel like proper controller parts. MSI rounded the ABXY buttons, changed the D-pad to a metal-dome design, and upgraded the haptics with new linear motors. I am not good enough with a controller to lecture anyone about competitive input latency, but the whole control surface felt normal in the best possible way. I did not feel like I was using a tiny PC with controller parts glued to the sides.

The fans are audible under load, as they are on every handheld in this class, but they never became the dominant sound. The two front-facing speakers help. Turn the volume up a little and the game covers most of the fan noise without headphones.

My informal battery experience landed around three to three and a half hours while playing and using the machine normally away from a charger. This was not a controlled rundown with fixed brightness and a scripted path, so do not read it as one. The 80Wh battery is large for a handheld, and Intel also has an Endurance Gaming mode that targets 30, 40, or 60 fps to reduce power. I did not run the full matrix needed to turn that into a stronger battery claim.

The rest of the hardware is unusually complete. My unit has 32GB of LPDDR5X memory and a 1TB SSD. MSI moved to a standard M.2 2280 slot, so storage upgrades should be cheaper and easier to find. There is Wi-Fi 7, a microSD Express reader, a fingerprint sensor in the power button, and two Thunderbolt 4 ports across the top.

Those Thunderbolt ports deserve their own section.

The one-cable dock changes the product

Intel sent an OWC Thunderbolt Go Dock with the review kit. It does not come in the retail box, but it showed me one of the Claw’s best tricks.

I plugged the dock into my monitor, connected the Claw with one cable, and used it the same way I would use a Nintendo Switch. The cable handled the display, peripherals, and charging. Dock it and play. Unplug it and walk away.

That sounds obvious because Nintendo solved it years ago, but it has not been the default experience for Windows handhelds. Thunderbolt makes the setup predictable, and having two ports means charging or docking does not consume the only high-speed connection on the machine.

It also changes what the Claw can be. This is a full Windows 11 PC with 32GB of memory. Add a keyboard, mouse, and monitor and it can handle ordinary desktop work, development, or local AI experiments in a way a closed console cannot. I did not replace my workstation with it, and I am not going to pretend a handheld is the ideal machine for every job. The point is that the option is real and the setup takes one cable.

Full Windows matters before you ever reach a desk too. The newer Xbox full-screen experience gives the Claw a controller-first home screen while keeping access to games from Xbox, Epic, Steam, and the rest of the PC ecosystem. I am apparently in the minority here, but I like buying games on Epic. A device that can bring that library along without asking me to repurchase games is more useful to me than one built around a single store.

Windows can still show its desktop-shaped edges, and a console-first interface does not remove every prompt or launcher. It is much closer than it used to be. More importantly, I did not spend the review fighting the operating system. I spent it playing Spider-Man.

Why PC handhelds will keep growing

The Claw feels like a mature answer to a real problem: how do you make PC gaming portable without asking people to abandon their library, their accessories, or the option to use the machine as a computer? There is no single trick here. The category works because the chip, controls, battery, software, display, and docking finally work together.

The cost backdrop matters too. TSMC says the tools for leading-edge nodes are becoming more expensive and the process complexity keeps increasing. Micron expects DRAM and NAND supply to remain tight beyond 2026, with enough pressure to weigh on PC shipments. Display companies are moving more of gaming toward high-refresh OLED, which remains a premium part. None of that makes a $1,799 handheld cheap. It does mean the old assumption that a handheld only makes sense as an inexpensive second device is becoming less useful.

For some people, this can be the gaming PC. It can live on the couch, go in a bag, dock to a television, and connect to a monitor and peripherals at a desk. A desktop will still deliver more performance for the money, and a laptop remains better if you need a built-in keyboard and larger screen. The handheld earns its place through the number of situations where it works.

This particular machine sits at the expensive end of the category. The important part is what happens next. Arc G3 gives more manufacturers a capable platform to build around, Windows is finally treating handhelds as a real product class, and every new design creates another chance to balance performance, size, display, and price differently. The category does not need every product to be cheap. It needs enough good options that more people can find the one that fits.

The best thing I can say about the MSI Claw 8 EX AI+ is that it got out of the way. It gave me a reason to restart Spider-Man 2 and ran it well enough that I stopped watching the frame counter. When I finished testing, I kept playing.

Intel nailed Arc G3 Extreme. MSI nailed the hardware around it. This is the best Windows handheld I have used and the first one that did not feel like a downgrade. I love the thing.

I do not think every PC gamer is about to replace a desktop or laptop with a handheld. I do think many more people will choose one as gaming hardware gets more expensive and the handheld becomes capable in more places. A machine that works in your hands, on a television, and at a desk is easier to justify than one with a single job.

The Claw 8 EX AI+ is too expensive to take this category mainstream by itself. It is mature enough to show why something like it will. There are other versions with Intel’s Lunar Lake, and there are likely more affordable SKUs on the way. It’s only a matter of time!

A Transaction Layer That Never Existed

HTTP 402 has been in the web’s specification since 1991, a status code reserved for payment-required responses that never found a use case at scale. The reason is straightforward: until recently, the web’s consumers were humans, and humans had browsers, credit cards, and login sessions. The payment infrastructure built around those assumptions works well enough. Agents have none of those things. They cannot log in, cannot click through a payment form, and have no persistent identity that maps to a billing relationship. The result is that agents currently operate on the web the way someone might walk through a store taking things without any mechanism to pay, not out of bad intent but because no checkout exists for them.

That is the structural problem AWS is addressing. The WAF AI traffic monetization capability returns a machine-readable HTTP 402 response using the x402 open protocol, a standard developed by Coinbase for machine-to-machine payments. An agent that hits a monetized resource gets a price manifest in JSON: the cost in USDC, accepted blockchain networks, destination wallet, and payment timeout. Any x402-compatible agent runtime can complete that flow autonomously. No human needs to be in the loop.

This is not AWS inventing something from scratch. x402 is an open protocol, and Coinbase built the facilitator infrastructure. What AWS contributes is operationalizing it at cloud scale, at the network edge, across CloudFront’s global distribution network, with no changes required to origin infrastructure. That is a meaningful difference between an interesting protocol and something enterprises can actually deploy.

The Traffic That Pays Nothing

The urgency behind this is real. AI bot traffic now exceeds 50% of web traffic for many content providers, with AI-specific crawlers growing more than 300% year-over-year. Unlike traditional search crawlers like Googlebot or Bingbot, which returned measurable referral traffic, AI bots consume content to generate summaries and responses with little to no traffic sent back. The web’s existing monetization model has no answer for that. Blocking agents entirely is a short-term response that trades one problem for another. Licensing deals negotiated individually are not scalable. A programmable transaction layer is the only structural solution.

AWS WAF Bot Control already classifies over 650 distinct AI bot and agent types, including GPTBot, Claude-Web, and Perplexity-Bot. The monetization capability lets content owners price by content path, bot category, or verification tier. That last variable is worth pausing on. AWS distinguishes between verified agents, whose identity is confirmed through cryptographic signatures or documented IP ranges, and unverified agents, recognized through behavioral fingerprinting but not cryptographically confirmed. Pricing those two tiers differently creates a commercial incentive for AI companies to formally identify their agents. That is a shift in the accountability dynamic that goes beyond content monetization.

Why Amazon Is the One Building This

Amazon’s durable advantage has never come from selling things. It comes from owning the infrastructure that others use to sell things, and taking a position on every transaction that flows through it. AWS, Amazon Marketplace, Fulfillment by Amazon, Amazon Pay, Buy with Prime. The pattern across two decades is consistent. Find a commerce flow, build the indispensable intermediary layer, and embed deeply enough that switching becomes costly.

The WAF monetization announcement follows that logic. The feature is available at no additional charge to existing AWS WAF customers. AWS takes no fee on content revenue, with disbursement going directly to the content owner’s wallet. That looks generous until you consider that the value Amazon captures is not transactional margin. It is infrastructure position and the data visibility that comes from sitting at the center of every agent commerce flow.

What This Does Not Solve

It would be a mistake to read this as AWS completing the transaction layer for the agentic web. What they have built handles one well-defined use case: a content owner pricing and collecting payment from an agent requesting access to a protected resource. That is an important piece of infrastructure. It is not a complete answer.

Agent identity at scale remains unsolved. The verification tier distinction AWS draws today is a start, but a functional agent economy requires identity frameworks that travel across platforms, not just within AWS’s classification system. Liability is unresolved. When an agent transacts autonomously, the legal and financial accountability structures that govern human commerce do not map cleanly. Interoperability is an open question. x402 is an open protocol but adoption outside of AWS and Coinbase is early. A transaction layer that only works within a subset of the agent ecosystem is not a transaction layer for the web.

None of that diminishes what AWS shipped. It is the most serious production-grade infrastructure move on this problem to date. But the agentic web will need contributions from standards bodies, regulators, and the broader industry before the transaction layer is genuinely complete. AWS has built the first significant piece. The rest is still ahead.

At WWDC, the expectation is for a long list of new features rolling across every OS Apple supports in its portfolio. This year, Apple reframed that. Rather than leading with what is new, the keynote centered on perfecting and homogenizing the experience across products, addressing the kind of friction that quietly erodes user satisfaction over time, and laying the groundwork for what was the main focus of the event: Apple Intelligence and the new Siri AI.

Perfecting the Experience

Craig Federighi framed the releases around three explicit priorities: making platforms more responsive and easier to use in daily life, advancing trust and safety, and delivering a meaningful leap forward for Apple Intelligence with a new Siri architecture. That ordering was deliberate. Performance came first, and it showed throughout the keynote.

Across iOS, macOS, iPadOS, and watchOS, Apple’s updates centered on responsiveness and polish. App launches are up to 30% faster, with the system preloading key data apps need when opened, applying to third-party apps as well. Photos appear in libraries up to 70% faster after being shot. AirDrop transfers are up to 80% faster. File browsing and transfers from iPad to external drives are up to five times faster. These are the moments users encounter daily, and their cumulative effect on satisfaction is significant even when users cannot name the specific change that made things feel better.

Apple also rebuilt the foundation of search across iOS, iPadOS, and macOS, re-architecting the Search Index that powers Spotlight, Photos, and Mail to be more stable, more comprehensive, and faster at indexing new content. In Mail, a new ranking system surfaces more relevant results in top hits, so the email you need is more likely to appear first regardless of how long ago it was sent. Search is one of those features that users only consciously notice when it fails. Making it work reliably is the kind of improvement that adds up quietly into trust.

Apple extended these gains to older hardware as well, bringing an optimized CPU Scheduler all the way back to iPhone 11, meaning iOS 27 supports the same iPhone models as iOS 26 and extends Apple’s already notable device longevity record. At a time when sustainability commitments are under scrutiny across the industry, keeping older hardware performant is both a user satisfaction story and an environmental one.

The design layer received attention too. Liquid Glass, introduced last year as Apple’s most ambitious cross-platform design update, and one that did not please everyone, was refined based on user and developer feedback, with improved diffusion behind complex content for better readability. A new slider in settings lets users adjust Liquid Glass from ultra clear to fully tinted, applying the preference across apps including third-party ones that have already adopted it. None of this generates headlines, but it signals a company willing to revisit and improve rather than defend a prior decision.

Apple Intelligence and Siri AI: Grounded in the Real World

On Apple Intelligence, the company showed compelling everyday use cases that may not be revolutionary but are built to hold up outside a demo environment.

The Google collaboration on Siri AI generated the most discussion coming out of the keynote. Creative Strategies’ Max Weinbach explained it clearly after the keynote: Apple’s new advanced on-device model for A19 Pro runs 20 billion parameters on 12 gigabytes of memory, loading only the active parameters needed at any given moment. That model architecture is unprecedented. Apple does not lack AI research capability; it lacks large foundational models, which is what Google contributes to the collaboration. The reason Apple’s AI work has been underestimated is straightforward: companies like OpenAI, Anthropic, and Google publish benchmarks aggressively because they are selling AI as a service to enterprises and need to compete publicly. Apple is not selling that, so it has not played that game. The collaboration with Google fills the large model gap without Apple ceding control of its own architecture.

Apple chose to show what this means in practice rather than where it sits on a benchmark. The demos were grounded in real tasks: finding content by describing it naturally, drafting and editing in context, taking action across apps without requiring users to navigate between them manually.

Apple continued with the same architectural philosophy that has defined its approach to AI from the start: ground Apple Intelligence in the OS and in Apple’s own apps, then expose APIs so developers can bring those capabilities into their products. Apple Intelligence uses a new system orchestrator to coordinate across system-wide capabilities, including personal context understanding, broad world knowledge that can reach the web via Private Cloud Compute, and App Actions that draw on an App Toolbox to find the right tools from installed apps to complete a request. The result is AI that surfaces where users already are, which limits some of what Apple can announce at any given moment but also limits the gap between what Apple promises and what users actually experience.

The Siri AI app is the most structurally significant announcement from an architecture standpoint. It serves as the persistent home for user interactions with Siri, where conversations live, are searchable, and are accessible across devices. Apple is giving users a coherent AI identity that travels with them rather than a series of disconnected exchanges that disappear after each session. That continuity across devices matters practically, but it also reflects a considered view of how an AI relationship should work: one personality, one history, accessible from every device a user owns.

What stood out across the keynote was how grounded it was, and how grounded the demos were in particular. Apple showed real situations without overpromising what the technology does in those situations. Federighi framed Apple’s position directly: truly helpful AI must be centered around the user and their needs, integrated deep into the products people use every day, grounded in personal context and the apps they rely on, and designed with privacy at every step.

After leaving consumers waiting for a meaningfully improved Siri, Apple had to deliver. Releasing Siri in the developer beta on the day of WWDC keynote, with public availability in English in the US later in the year, points to a September launch alongside the new iPhone models on iOS 27. The stakes were high and the pressure was real. While we are only seeing the first glimpse through the beta, the signs are promising.

The company reported record third-quarter revenue of $15.8 billion, up 12% year over year, driven by demand for AI infrastructure and networking products. The announcements at Cisco Live made a clear statement about what Cisco has become and where it is positioning itself: the infrastructure layer for the agentic era, the same role it played for the internet.

The Playbook Has Run Before

Cisco lived this transition once,. When it built the networking backbone the internet ran on, it hit a constraint that had nothing to do with technology. There were not enough engineers in the world who knew how to deploy, configure, and manage what Cisco was selling. So Cisco built the talent pipeline itself. Networking Academy launched in 1997 with 64 institutions across seven U.S. states. It has since reached 28 million students across 195 countries. The CCNA, first awarded in 1998, became the industry benchmark for network engineering. Cisco did not wait for the talent market to catch up. It created the talent market.

The same logic is now being applied to AI. Cisco has committed to providing AI and digital skills training to one million student participants in the United States over the next four years, running from the beginning of fiscal 2026 through the end of fiscal 2029. The certification portfolio has been rebuilt accordingly. The CCNA v2.0 blueprint is organized around four pillars: network infrastructure, troubleshooting and problem-solving, a security-first mindset, and understanding the role of AI in network management and operations. AI literacy is now treated the way networking and cybersecurity once were: as a baseline competency, not a specialization.

This is a strategic investment. Cisco needs the people who will deploy its AI infrastructure to know how to use it. Enterprises need to trust that those people are qualified. Cisco is building both sides of that equation simultaneously, as it did with the internet generation.

The Infrastructure Argument

AI agents generate 450% more network traffic per task than a human performing the same work, and AI traffic overall is expected to triple in the next three years. The network is the constraint of this transition, not a background utility. Cisco Cloud Control is a new management platform bringing together networking, security, observability, and collaboration under a single interface, built for an environment where AI agents operate continuously across enterprise systems and need to be managed, secured, and audited in real time. Cross-domain telemetry sits at its core: data flowing across any connected system is compiled and viewable in one place, so humans and agents can act on coordinated, contextual information to manage uptime, control agent behavior, and prevent token overspend.

Hybrid Mesh Firewall extends unified protection across networks, applications, and both Cisco and third-party firewalls, designed to limit the blast radius in an environment where AI agents have extended the security perimeter and can introduce attack vectors at machine speed. Patel framed the underlying challenge as an “AI trust deficit,” the gap between what enterprises want to do with AI and what they currently feel confident deploying. The full portfolio is structured around closing it.

The Leadership Structure That Made This Possible

Patel’s consolidation of product strategy under one role is visible in what was announced last week. Networking, security, observability, and collaboration are no longer separate conversations with separate roadmaps. They are one platform. Liz Centoni as EVP and Chief Customer Experience Officer leads the organization responsible for ensuring enterprises can realize what Cisco is selling. Her keynote brought on stage two customers operating in environments where the margin for error is essentially zero. For GEODIS, Cisco IQ replaced guesswork with full visibility across 12,000 devices and a clear end-of-life roadmap. For GlobalFoundries, where maintenance windows are not an option, Cisco IQ cut vulnerability assessment from days to hours. Both cases made the argument that an integrated platform delivers where point solutions cannot.

The through-line from 1987 to 2026 is consistent. Cisco has understood, at each major infrastructure inflection, that the technology alone is insufficient. You need the infrastructure. You need the security model. You need the platform to manage it at scale. And you need the people qualified to operate all of it. Cisco built that full stack for the internet era, including the workforce development program when the market could not supply the talent on its own. The conditions are the same now. The difference is that Cisco arrives at this moment with a leadership structure finally organized to execute it as a single coherent strategy rather than a collection of business units making separate bets.

I was watching the movie The Hitman’s Bodyguard a couple of weeks ago, and one of the characters says “boring is always best.” While that might be a line written for an action film, it’s also a perfect rule of thumb for consumer technology. When a laptop is meant for everyday users, the best place it can be is a spot where you can’t think of a single major issue. Almost every laptop has some weird, annoying quirk that ruins the experience. I can’t find any of those game-breaking flaws on this machine. It just works, it’s comfortable, and it’s easily my default recommendation.

Refinements over the First Generation

I spent a lot of time with the first-generation Snapdragon Yoga Slim 7x, and it was a bit of an odd machine. The trackpad felt mushy, the speakers were mediocre, and the 14.5-inch chassis was slightly wider than it needed to be.

This second-generation model fixes those issues. Lenovo trimmed the screen size down to 14.0 inches. That half-inch reduction makes the footprint a lot smaller. It now weighs 1.17 kg and measures 13.9 mm thick, which is a much more comfortable size to throw in a backpack.

Lenovo didn’t mention it in their marketing, but they completely redesigned the trackpad. Clicking the old and new models side-by-side reveals the change immediately. The new one has a tactile, crisp click that feels vastly superior to the mushy mechanism from last year. The keyboard layout is roughly identical, but it feels like they made a subtle tweak to the key travel that makes typing more comfortable. They also upgraded the audio system to four stereo speakers (two 2W woofers and two 2W tweeters). It sounds clean and punches well above its weight class for a thin-and-light laptop.

I actually passed the first-generation Snapdragon laptop down to my dad when he needed an upgrade. He chose it over several other laptops because it hit all his basic needs. He was a big fan of it, but when I showed him this new version, he liked it even more (but I am keeping this one for myself). The refinements are subtle, but they add up to a much more polished product.

Elite Silicon without the Extreme Compromise

The silicon landscape on Windows on ARM has split into distinct tiers. Under the hood, this laptop runs the standard Snapdragon X2 Elite chip. Qualcomm also offers a flagship X2 Elite Extreme, but I think the standard Elite is the smarter choice for a thin-and-light laptop.

To understand why, we have to look at how these chips are packaged. The Snapdragon X2 Elite Extreme uses on-package memory, which sits right next to the silicon to deliver higher memory bandwidth (228 GB/s, against 152 GB/s on the standard Elite) and better native efficiency. In a vacuum, that should mean better battery life.

However, Qualcomm has to justify the “Extreme” branding. In practice, they push the Extreme chip’s clock speeds higher to secure impressive peak performance numbers in synthetic benchmarks. Pushing those clocks spends the efficiency gains, which means the battery life ends up leveling out between the Elite and the Elite Extreme.

The Extreme chip is destined for larger, heavier chassis designs that can handle the thermal output of sustained high clock speeds. For a 1.17 kg thin-and-light like the Yoga, the standard X2 Elite is the right fit. It keeps the chassis thin, stays cool, and runs almost completely silent (the fans are barely audible in daily use).

To be clear, none of this is confirmed by Qualcomm, but given what I know about the silicon and have experienced myself, I tend to believe this is true.

The Screen Configuration Mismatch

My review unit came configured with the Snapdragon X2 Elite, 32 GB of RAM, and a 1 TB SSD, but it was paired with the base WUXGA (1920 x 1200) 60Hz OLED display. Lenovo sent me this unit directly, and while it’s a fine screen, I don’t think this display configuration makes sense.

This 60Hz 1200p OLED belongs on the lower-end Snapdragon X2 Plus laptops. If you’re configuring a laptop with Qualcomm’s premium X2 Elite chip, you deserve a premium display. Lenovo offers a gorgeous 2.8K (2880 x 1800) 120Hz OLED screen as an upgrade option on their website.

The markup on Lenovo’s website to jump from the 60Hz WUXGA screen to the 120Hz 2.8K display is $60. That’s a rounding error on a laptop in this class, and with a gap that small, Lenovo should just kill the base 60Hz display option on the Elite configurations entirely. Pairing a high-end chip with a low-refresh-rate screen is a compromise that doesn’t need to exist. If you’re custom-building this on Lenovo’s site, spend the $60 and get the 2.8K display. It’s a no-brainer.

Real-World Performance: The Agent Benchmark

I’m not going to dump standard benchmark numbers here. We already covered those in my Snapdragon X2 Elite Extreme reviews, and synthetic benchmarks don’t tell you what a laptop is actually like to use.

Instead, my ultimate test for real-world performance has been running local AI coding agents. For the past few weeks, I’ve been working on a side project using Claude Code and a local co-work agent. The project is an implementation of SwiftUI that compiles natively on Windows and Windows on ARM. Unlike other ports, I wanted this to have its own custom GPU rendering engine (similar to how GPUI works in Rust) rather than mapping to WinUI 3.

This is a heavy workload. It requires constant compiling and building of native Swift code on the machine. Any CPU bottlenecks immediately slow down the loop between agent runs.

I’ve run this project on my main desktop (which rocks an RTX 5090 and Intel Core i9-13900K) and several heavy desktop-replacement laptops. When I ran it on this Snapdragon X2 Elite Yoga, I was stunned. It’s one of the fastest laptops I’ve ever used for this workflow. The CPU compiles native Swift code incredibly fast. The agents finished their tasks so quickly that it felt like a night-and-day difference in how much work they could get done. A lot of this is purely from the higher performance per core, not every workload is using all the cores to compile.

I also used a co-work agent to build TypeScript and Electron applications, which are notorious for compiling slowly on Windows. On this machine, those builds are snappy. When you pair that level of performance with great battery life, fast Windows Hello face recognition, and a thin chassis, it makes the Yoga an incredible development machine.

The Easy Default

Qualcomm’s chips make this laptop what it is. The thinness, the silence of the fans, the performance under agent workloads, and the outstanding battery life are all direct results of the Snapdragon X2 platform.

But Lenovo deserves credit for nailing the chassis design. The product you touch and feel is great regardless of the configuration you choose. Whether you buy the base Snapdragon X2 Plus SKU or custom-build a top-tier Elite model, the physical experience is excellent.

We’re missing easy defaults in the Windows laptop space. It’s hard to find a machine that doesn’t have some glaring flaw that will annoy you or break your workflow. Lenovo built a laptop that has none of those. It’s an easy recommendation for almost anyone. If I’m throwing a Snapdragon laptop into my bag, it’s going to be this one.