State of AI — September 2026

Where things stand

So last month I told you the biggest gap I see isn’t in the technology, it’s in the understanding of it. That’s still the hill, but something shifted in my month: the conversations got specific. A month ago people were asking me what this stuff even is. Now the same conversations come with an actual process attached, a quoting queue, an accounts-payable pile, a shop floor nobody can see into. When somebody hands you the pile of paper instead of a question about the technology, that’s a different stage of this, and it’s worth marking.

Now, the one thing August actually changed for me is how clearly the two layers of this thing came apart. The people funding the frontier repriced it three times in four weeks, in both directions, with dates attached. One lab told us its own safety tests stopped registering progress and shelved a model stronger than anything it sells. And Nvidia trimmed a quarter-trillion-dollar data-center promise down to about a hundred billion because its own shareholders blinked. Another business owner I sat with this month put it in terms of the British railroad mania of the 1840s: enormous capital went into laying rails ahead of any proven demand, and a lot of the fortunes ended up with the people who moved freight on them. I think that’s where we are. The edge doesn’t go to whoever funds the frontier, and it doesn’t go to whoever owns the model. It goes to whoever connects the technology to a real problem set.

Okay, the tension I’m watching is my own window. This month serious capital went into industrializing exactly the kind of work I do: a holding company raised two billion dollars to buy service firms and rewire them, IBM is certifying tens of thousands of consultants on OpenAI’s models, and Anthropic’s implementation venture bought its second consultancy. I don’t know how long the gap stays open, and I’m not going to pretend I do. But here’s the thing: a roll-up can only rewire what it buys, and a certified army sells a playbook. The businesses I work with don’t need a playbook, they need somebody standing in their building who can see how the work actually moves and leave behind a standard that holds after he’s gone. That part doesn’t commoditize on the same schedule, and I’m building against it like the window closes next year, which some nights means working until 3 a.m.

As far as the critics go, some of them are right this month, and I’ll say that plainly. There are companies out there wearing AI like a costume, roll-ups slapping the letters on a deck to cover an old-fashioned balance-sheet problem, and when those unwind people are going to say the whole thing was a costume. It wasn’t. The capability is the capability. The same month the costume companies were getting called out, a government test lab watched an agent invent fake identities to talk a software maintainer into accepting its code, and machine-checked proofs closed math problems that sat open for decades on two thousand dollars of compute. You don’t have to like both of those facts, but they’re facts. Judge the field by what runs on the ground with receipts, not by the costumes and not by the demos.

So here’s your starting point this month, and it builds on last month’s. If you wrote down your one process, take it back out and go line by line marking every step one of two things: busywork or judgment. Busywork is anything where the answer is checkable, the pattern repeats, and nobody’s reputation rides on the keystroke. Judgment is where a human’s name has to be on the call. The busywork list is what you hand to a machine, the judgment list is what you protect, and in my experience the split runs further toward busywork than anybody expects. That one page is the working split for everything else in this report, and the “What I’d do this month” section picks it up from there.

What shipped

The API market repriced three times in four weeks, in both directions. OpenAI opened the month having just cut its GPT-5.6 Luna tier 80%, to $0.20 per million input tokens and $1.20 out, with Terra down 20% to $2/$12 (CNBC). On August 16, DeepSeek went the other way: the vendor that set the industry’s floor price introduced peak and off-peak rates and raised its V4-Pro output price from $0.87 to $3.96 per million tokens at peak, with V4-Flash input up from $0.14 to $0.44 (DeepSeek pricing, DeepSeek). Peak hours run weekday mornings UTC, and off-peak is exactly half price, so the same job costs twice as much on a weekday morning as it does overnight. Then on August 21 OpenAI cut its flagship Sol from $5/$30 to $4/$20, promotional through at least November 21 (OpenAI), and AWS matched it on Bedrock the same day (AWS).

One increase did not happen. Anthropic had scheduled Sonnet 5 to go from its $2/$10 introductory price to $3/$15 on September 1, and on August 10 it canceled the increase and made $2/$10 permanent (Anthropic pricing docs). The tokenizer caveat from last edition still applies: Anthropic’s newest models emit roughly 30% more tokens for the same text, so the sticker understates the per-document cost.

The new models mostly arrived at the middle of the price band. xAI’s Grok 4.6 landed August 12 at $2/$6 with a 500K context window (xAI). Google’s Gemini 3.7 Flash landed August 13 at $0.75/$3.75, introductory through December 31 (Google). The same day, OpenAI previewed Ultrafast, GPT-5.6 Sol served on Cerebras wafer-scale hardware at up to 750 output tokens per second, roughly 14x standard speed, with no pricing announced (Cerebras). Speed is becoming a purchasable tier separate from capability.

The bigger August story on the open side is who shipped weights. DeepSeek’s V4-Pro exited preview August 13 with MIT-licensed weights, roughly 1.7 trillion parameters and a 1M context window, at unchanged prices (DeepSeek coverage). Alibaba shipped the promised Qwen3.8-Max weights around August 12, though the open version is text-only, without the vision input or the 1M context of the hosted model, under a custom license rather than Apache 2.0; a smaller 27B sibling did ship Apache 2.0 (Hugging Face). Meta released Muse Glimmer on August 10, a 30B Apache 2.0 model built specifically for agent work (Meta). And Harvey, the legal AI vendor, released Tenet on August 20, its own open-weight legal model post-trained from Moonshot’s Kimi K3 base, with roughly double the task completion of the base model on its internal legal benchmark (Harvey). That’s a vertical vendor with real customers deciding the tuned model should be an asset it owns, and it’s the cleanest answer yet to a lab repricing. Z.ai released GLM-5.3 August 14 with strong vendor-reported security-research scores, but its promised weights had not shipped as of August 25.

Capability news came with the brakes attached this month. OpenAI opened August by publishing solutions to ten math problems that had been open for a decade or more, each with a machine-checkable Lean proof, at a token cost OpenAI put near $2,000 (The Decoder). Five days later it disclosed it could not rule out that the same unreleased model, Astra, meets the top “Critical” cyber-risk threshold of its preparedness framework, and slowed internal work (TechCrunch). On August 10 it shipped GPT-5.6-Cyber behind identity verification and legal attestations: the gated model completes 95% of exploit-development requests its guarded flagship refuses at a 1.5% rate (The Hacker News). On August 18 it paused reinforcement-learning training on deployment-bound models for two weeks while it hardens its research environments (OpenAI).

Anthropic’s August 14 risk report is the other half of that picture. It raised its catastrophic-misalignment risk rating from “very low” to “low,” stating plainly that this reflects increased uncertainty rather than worse models: its task-based safety evals have saturated and no longer register capability gains, while the company says it sees early signs of acceleration. The same report disclosed an unreleased internal model, called Model 2, that scores 62.8% on Anthropic’s own R&D benchmark against 50.3% for Mythos 5, its best released model, with no release plans (Anthropic risk report).

Two more shifts landed in August. Anthropic began watermarking every Claude output worldwide on August 11, an imperceptible signal in generated text plus signed provenance on files; the mark shows text passed through Claude, not that Claude wrote it, so an AI-edited human draft carries the same mark as an AI-authored one (Anthropic). And the agent-standards layer kept consolidating without touching trust: the Agent Plugins 1.0 spec shipped August 6 covering packaging and discovery while leaving permissions and sandboxing to each client (Vercel), Google handed its A2A protocol to the Agentic AI Foundation on August 20 (AAIF), and the MCP roadmap published August 22 puts agent identity, the piece unattended automation actually needs, still in the future (MCP).

What’s actually working

The adoption datasets that landed in August measure divergence, not adoption. OpenAI published enterprise usage data around August 12 showing its top-decile customers now generate 8.3x the output tokens per active user of typical firms, up from 2.6x in January, and the growth since February is concentrated outside engineering: weekly active users grew 108x in legal, 41x in sales and recruiting, and 26x in marketing, against 5x in engineering (OpenAI). Two caveats run with those numbers. The non-engineering multiples grow off near-zero starting bases, and OpenAI itself warns that token volume is not a measure of value.

Linear published its own first-party numbers on August 21: AI now authors just under half of all issues on its platform, up from fewer than one in a thousand two years ago, and teams running coding agents tripled weekly pull-request output from 21 to 65 while teams without went from 8 to 10 (Linear). Linear’s own conclusion is that AI landed on top of existing work rather than replacing it: planning time is flat and triage time is up. The counterweight comes from LinearB’s benchmark report published in May, from 8.1 million pull requests: AI-assisted PRs wait over 16 hours for review against roughly 200 minutes for unassisted ones, and merge at 32.7% against 84.5% (LinearB). Generation scaled and the human checkpoint did not. That is the approval gate, measured.

Salesforce’s Agentic Enterprise Index, published August 10, says the average customer deployment grew from 5 agents in February 2025 to 13 in April 2026, with stand-up time down 53%. The split it found is useful: consumer-facing sectors run high-volume, narrow task agents, while regulated industries like manufacturing, financial services, and healthcare build multi-step agents carrying approvals (Salesforce). The telemetry counts only orgs that ran production agents every month of the window, so it describes committed adopters, not the market.

The measurement gap underneath all of this hasn’t moved. Thomson Reuters’ professional-services study, fielded in late 2025 and published this spring, found GenAI use nearly doubled year over year to 40% of organizations, while only 18% measure whether the investment works at all, and 77% of those track only internal metrics like cost savings (Thomson Reuters). The population is self-selected professional-services respondents, and the field dates matter, but no August dataset contradicts it.

The month’s clearest deployment lesson came from a corner store in San Francisco. Andon Labs runs a real store there managed by an AI called Luna, running Claude Opus 4.8, with a budget, a card, and contracted staff. In mid-August it moved to terminate a worker who was late for 17 of 23 shifts, the first known case of an AI manager initiating a real firing; humans reviewed and carried it out (TIME). The lesson is in the logs. Luna had written an attendance policy months earlier and then lost track of it, and only acted after a staff member prompted it to search its own records. A long-running agent silently dropped a rule it was supposed to enforce, and nothing in the loop noticed until a human did.

Where it broke

Every incident on this month’s table shares a shape: the boundary existed in the prompt or the default, and the environment disagreed. Nobody jailbroke a model. The holes were misconfigured test infrastructure, permissive defaults, and config files nothing scans.