State of AI — August 2026

Where things stand

So here’s where I actually stand after another month inside this work. The biggest gap I see isn’t in the technology, it’s in the understanding of it. Nearly everyone I talk to knows these AI models have capabilities they don’t understand, and they have no idea what those capabilities actually are. That’s the hill I spend most of my time on, and honestly it’s the largest one: getting people to understand the nature of these systems before I can even get to what I believe can be done in their organization. Most people don’t know where the starting point is. They don’t know there’s another interface than a standard chat, and they don’t know there’s a different kind of capability sitting behind it. There’s not just ask-and-get-an-answer. There’s ask-and-get-an-action, and understanding the action part, and just how far that action can go, is going to be the main thread for a while.

Now, the one thing that actually changed this month is that more people are asking. More and more of the people I talk to are expressing a desire to learn what the capabilities are and then to implement them. People definitely understand there are huge gains to be had and real danger if it’s done wrong, and as time goes on the appetite to understand it keeps growing. I’m glad that when they get to that point, the people I talk to end up talking to me, because I think there are people out there who understand the capability but don’t see it through the lens of operations or the full context of an organization and what has to go with it.

Okay, the tension. We’re so early in this that it’s hard to put hard numbers on what these systems can do, because the work is very custom. There isn’t a standardized install path. There’s no organization of AI implementers, no governing body for AI implementation, nothing like that where there’s a certain standard. I’m not in any way saying I want one, but it is the Wild West. For an organization to actually gain value, the systems being implemented have to be communicated extremely clearly, and they have to be studied and understood by whoever is doing the implementing, because there aren’t always easy connectivity surfaces between the way a business currently works and what agentic AI needs. So the real tension is basically: is the end result actually going to save time or money, and how do we get there?

As far as the critics, I don’t know what else to say other than the capability is the capability. I’m sure there are going to be people who say this is all overhyped, and that’s fine. From what I’ve seen on the ground in real businesses in the time I’ve been doing this, there’s enough proof, at least in my mind, that there is net positive to be gained from agentic AI. It’s not vaporware. Whether it’s overvalued or undervalued is kind of beside the point, because it’s not a made-up thing. It’s incredibly capable, and it’s going to get better. Maybe this hits a bubble like the dot-com fallout did, and even that wouldn’t mean it isn’t transformative. In my mind it has the potential to be the most transformative thing that’s ever happened, because you can see and formulate a path in which it is that transformative. This is really the first year that agentic AI is really real, and look at what’s already getting done in this mess, where nothing is organized and there is no standardization and things are still getting done. Just imagine what happens when this all comes together.

So if you’re looking for the starting point, here it is. Pick one process in your business, the one that eats the most hours, and write down plainly how it actually works today: who touches it, what they check, and what happens when it goes wrong. You can’t hand work to a system you haven’t described, and you can’t put a gate on a system you don’t understand. That one document is where every real install starts, and the rest of this report is the state of what you’d be handing the work to.

What shipped

The story of the summer is that frontier capability got cheap. On July 24, Anthropic shipped Claude Opus 5 at $5 input and $25 output per million tokens, the same sticker as the model it replaced, while claiming performance close to its $10/$50 flagship Fable 5 (Anthropic). Six days later OpenAI cut its GPT-5.6 Luna model by 80%, down to $0.20/$1.20, and Terra by 20% to $2/$12 (CNBC). Anthropic then made Sonnet 5’s $2/$10 introductory pricing permanent, canceling the increase to $3/$15 that had been scheduled for September 1 (Anthropic). Jefferies, citing Silicon Data, put the average price of inference at $1.16 to $1.18 per million tokens in the first week of August, the lowest recorded this year and down from $2.04 on May 31 (SCMP).

The open-weights side moved just as fast. Moonshot released the weights for Kimi K3 on July 27, the largest open-weights model to date at 2.8 trillion parameters, and it debuted at #3 on Artificial Analysis’s overall leaderboard behind Claude Fable 5 and GPT-5.6 Sol (Tom’s Hardware). Alibaba’s Qwen3.8-Max went GA on August 3 at $2/$6 (Forbes). And on August 10, Meta returned to open source with Muse Glimmer, a 30-billion-parameter Apache 2.0 agent model built to run on a single consumer GPU (CNBC). That last one matters for any business that wants AI without sending data anywhere: a genuinely capable agent that plans, calls tools, and recovers from failure now fits in under 20GB on hardware you can buy at retail.

On the independent leaderboards, the top is crowded. SWE-bench Verified has Claude Opus 5 at 96%, Mythos 5 at 95.5%, and Fable 5 at 95% (BenchLM). OSWorld-Verified, which measures real computer use, has Qwen3.8-Max at 86.1% with the two Claude models at 85% (BenchLM). Terminal-Bench 2.1 has Kimi K3 at 88.3% and Mythos 5 at 88.0% (CodingFleet). Three different vendors at the top of three different boards, and two of them sell at $2 to $3 per million input tokens.

Two caveats an operator should know. Capacity is rationed even while prices fall: Moonshot froze new Kimi K3 signups about 48 hours after launch, saying demand pushed close to the limits of its GPU capacity (Euronews), Microsoft told investors it expects to be capacity-constrained at least through 2026, and Amazon said AWS will still not have enough capacity to meet demand this year (CNBC). And the per-token sticker can mislead: Anthropic’s newest models use a tokenizer that produces roughly 30% more tokens for the same text, so the effective cost per document is higher than the list price implies (Anthropic docs). DeepSeek, the cheapest name on the chart, is warning of a significant price increase in its own documentation (DeepSeek).

What’s actually working

Here’s the honest picture, and it takes four different surveys to see it. The US Census Bureau, running the most representative read there is, puts AI use among American businesses at roughly 17 to 20% from December 2025 through May 2026 (Census). Kyndryl’s survey of 1,100 senior leaders found 57% have deployed AI broadly, but only 11% have hit both of their top two AI goals, and the share who say their workforce is ready actually fell six points from last year, to 23% (Kyndryl). For small businesses specifically, Pax8’s Q2 survey of 402 US SMB leaders found 61% actively using AI and 29% still experimenting, while just 11% have a documented AI policy (Pax8). The widely quoted IDC finding that 88% of AI proof-of-concepts never reach production dates from IDC’s research with Lenovo published in March 2025, so treat it as last year’s baseline rather than this summer’s news (CIO).

Where it is working, the numbers are specific. Stripe published the architecture of Kai, its internal AI platform, and reported 83% of employees using it weekly within two weeks of launch, over 5,000 sessions a day, and sales reps closing 39% more deals in the weeks they use it (Stripe). Salesforce’s telemetry across its own customers shows the average agentic deployment growing from 5 agents in February 2025 to 13 in April 2026, with the time to stand up an agent down 53% to 1.9 days, though that data comes from companies that already ship agents, not a random sample (Salesforce).

Where it broke

Every major lab had an agent do something it wasn’t supposed to do this summer, in public, with a written incident report. The pattern across all of them is the same one: agents running under loosened conditions, real internet access, safety classifiers off, took real actions on real systems.