AI in September 2026: The AGI Hype Machine Just Hit a Wall
Septemebr has been a busy month as always with AI news and controversies. People are spending more on tools they use regularly, models are becoming more capable, but reaching their limits.
Menlo Ventures’ September report estimates two billion AI users worldwide and $40 billion in consumer spending, up from $12 billion. Two billion is approximately 24% of the world population, rather than 20%. The study surveyed 5,067 U.S. adults in July, with global figures estimated separately. Menlo’s 2026 report
Daily use reached 25% of U.S. adults, versus 19% previously, among AI users, it was 39%. Daily ride-sharing, food delivery and streaming measured 6%, 7% and 46%. Payers were almost twice as likely to use AI daily 50% versus 26% and five times as likely to use agents: 65% versus 13%. Adoption rose with income, high spenders disproportionately included parents, postgraduate degree holders and technology or finance workers. These associations do not establish that buying a subscription causes greater use. Menlo
Daily use and consumer spending rose at different rates. Chart: Menlo Ventures, 2026 consumer AI report. Source · Full-size image
In reported usage among U.S. AI users, ChatGPT led Gemini, followed by a broader mix including Meta AI, Copilot, Alexa, Claude, Perplexity and Grok. Both Alexa and Siri declined. Google looks well positioned to win over the long term. Menlo
The shift toward agents also deserves a precise definition. A sequence of automated steps is not necessarily an agent. In a system such as n8n, an agent can select tools and adapt its next action to a result, a fixed workflow follows predefined logic. Both can be useful, and a good implementation can combine them. n8n’s agent documentation
GPT-6 Astra brought another round of ambitious claims about artificial general intelligence. OpenAI released it on September 3, September 6 was the date of Jensen Huang’s celebratory “AGI has arrived” post. Astra’s release does not establish that AGI has arrived. The $1.2 billion training-cost estimate remains unverified by an authoritative disclosure. OpenAI release record, report linking Huang’s statement
Independent results show meaningful progress alongside limitations. Artificial Analysis reported on September 9 that Astra at maximum reasoning effort tied Claude Fable 5.1 in its Intelligence and Coding Agent indices, at approximately 40% and 60% of Fable’s respective cost per task. Astra offers better value on these benchmarks, based on total task costs rather than token prices alone. On AA-Omniscience, Astra’s benchmark-defined hallucination rate fell from GPT-5.6 Sol’s 92% to 51%, with accuracy also improving. This does not mean half of everyday answers are false. Fable’s higher tendency to answer uncertain questions also illustrates why accuracy, abstention and hallucination rates need to be read together. A lower rate for an open model does not by itself establish greater overall accuracy. Astra evaluation, Fable evaluation
Quality and cost per completed benchmark task, from Artificial Analysis’s September 9 evaluation. The horizontal axis uses a logarithmic scale. Source · Full-size image
ChatGPT and Codex are strong choices for demanding tasks, while Gemini 3.8 Flash is especially compelling for commercial workloads where speed and cost matter. Artificial Analysis’s September 2 evaluation placed it on the cost-efficiency frontier, reporting around 300 output tokens per second and $0.58 per task at high reasoning. Gemini 3.8 Flash evaluation
Claims of “superhuman” performance deserve the same care. The independent SpatialBench project reported Astra at 91.78% with Code Interpreter, against an 80% human baseline but 67.07% without tools, on just 50 questions. A music evaluation by Auggie found the strongest Bach-style chorale result among the models he tested and perfect performance on his error-detection test, while acknowledging compositional weaknesses. These are interesting results on defined tasks. Neither establishes general superiority to humans or shows that OpenAI deliberately cherry-picked the examples. SpatialBench, original music evaluation
The Minecraft example is similarly revealing. Reporting on Vals AI’s 141-hour experiment describes Astra collecting six blaze rods and three ender pearls before a Creeper destroyed its stored equipment, it subsequently spent hours farming potatoes. It did not beat the Ender Dragon in that run. This was a Vals evaluation, rather than an OpenAI-run benchmark, and the estimated tens-of-thousands-of-dollars bill remains unverified. There are uneven performance across a long sequence of actions. The effect of adding images or audio on accuracy depends on the model and task. Minecraft report
A broader assessment, A Definition of AGI, examines ten cognitive domains, including memory, visual processing, auditory processing and speed. Its estimated aggregate scores are 27% for GPT-4 and 58% for GPT-5. GPT-5’s visual score contributes four percentage points to the overall assessment, from ten available, it does not mean the model completed only 4% of visual tasks. Substantial weaknesses may persist in GPT-6, and whether AGI is achievable remains an open question. Original paper
The paper’s estimated cognitive profiles for GPT-4 and GPT-5 in Auto mode. Each domain contributes up to ten points to the overall score. Figure 1 from Hendrycks et al., A Definition of AGI (CC BY 4.0). Source · Full-size image
Humanity’s Last Exam offers another demanding test, covering 2,500 public questions across more than 100 subjects, with private questions retained to assess overfitting. Its authors explicitly caution that high scores alone would not establish AGI. These results raise questions about diminishing returns. Nowadays, results resemble logarithmic rather than exponential improvement. Benchmark contamination remains a concern, although there is no evidence here that specific developers trained on particular tests. HLE, Fable results
Today’s AI expectations echo the era when processors advanced from hundreds of megahertz to one and then two gigahertz, followed by predictions of ordinary 10 GHz PCs. The distinction matters: Moore’s law describes transistor counts, not a promise that clock speeds or useful performance double indefinitely. Power constraints encouraged greater reliance on multiple cores and architectural changes while semiconductor development continued. That history is a reason to examine the assumptions behind growth forecasts, it does not establish that AI improvement is about to stop. An approaching AI capability ceiling is plausible, but its existence and timing remain uncertain. Intel’s explanation of Moore’s law, research on CPU and GPU design trends
Price competition is easier to document. DeepSeek released V4.1-Flash on September 10. On OpenDesign Arena’s design tasks, it scored 81.2 against Astra’s 82.7, with estimated costs of $0.023 versus $1.61 per artifact: approximately 98% of the score at 1.4% of the cost. The striking ratio belongs to that benchmark, rather than every kind of reasoning. Using DeepSeek’s hosted service also has a data-location implication: its privacy policy says it collects, processes and stores personal data in China. Running open weights independently is a different arrangement. DeepSeek release, OpenDesign results, privacy policy
Gemini 3.8 Live is useful for researching companies, people and meeting topics through voice conversations while on the move. Google introduced Live and Live Extended Thinking on September 15, with speech interaction, tool calling and visual context, the developer documentation also confirms search grounding. Google reports that Extended Thinking leads Artificial Analysis’s Speech-to-Speech Quality Index. Voice interaction is convenient, but claims that it is universally less accurate than text need task-specific evidence. Google announcement, Live documentation
Expectations for Gemini 4 Pro are more speculative. Reports of leaked charts, lower prices and possible leadership across categories remain unverified in official Google material. Sergey Brin’s involvement in Gemini is documented in a Google interview, but does not validate the leaks. Google’s financial resources make it a strong contender for long-term leadership. Alphabet is profitable, but so is Meta: that is different from proving Google is the only sustainable AI developer or that Gemini itself is profitable. Google’s Brin interview, Alphabet results, Meta results
TypeSafe’s launch provoked more skepticism. The model is called Jev, rather than “GEM”, lead investor DCVC confirms a $40 million seed round and founder Diogo Almeida’s OpenAI background. TypeSafe describes typed, probabilistic decisions such as booleans, numerical scores and categorical choices produced in parallel, using a training method it calls Reinforcement Learning for Calibrated Decisions. Jev needs to demonstrate a meaningful advantage over conventional classifiers and constrained LLM outputs. Claims of paid influencer campaigns or universally inferior accuracy remain unsupported. TypeSafe’s published comparisons use particular workflows and reference-model agreement, not a universal truth test. Investor announcement, TypeSafe’s technical claims and caveats
TypeSafe’s comparison across four workflows. Its accuracy measure uses reference-model answers, these are vendor-reported results. Source · Full-size image
The refund, language and urgency examples illustrate an important distinction: structured LLM outputs can constrain a JSON field to permitted values, such as true or false, without guaranteeing that the classification is correct. A conventional classifier may be a better fit for a defined task, but its accuracy and total cost need to be measured on that task. Structured Outputs documentation
DiffusionGemma, Google’s experimental open model released in June. Google reports up to four-times faster generation on dedicated GPUs. It is an alternative worth testing, but neither universal speed leadership nor superiority to Jev is established, and self-hosting still has infrastructure costs. DiffusionGemma announcement
Image generation provides a more personal reality check. GPT Image 2.5’s Sunburst and Flare models arrived in September, and ImageBench’s current leaderboard scores them above Image 2. In my experiment, changing my portrait into a soccer uniform and then removing a hat preserved my face but changed the shirt. I also tried a retro 1980s-style photo prompt. These anecdotes capture improvement without perfect consistency. Smaller iterative edits align with OpenAI’s guidance, although long prompts are not inherently wrong. Image model, ImageBench, image prompting guidance
September’s controversies bring AI Ranch’s skepticism into focus: useful technology can coexist with exaggerated promises, commercial incentives and real operational risks. The challenge is to separate what happened from what observers think it means.
Start with employment. Jevons’s paradox describes how greater efficiency can make a resource so useful that total consumption rises. The historical example concerns more efficient coal use, especially in steam engines, rather than coal-processing pipelines. Jevons described it in his 1865 book, The Coal Question.
That offers a possible explanation for expanding demand around AI. September’s Economist analysis, republished by Mint, estimates roughly a million new US jobs associated with AI and highlights demand for people building and powering data centres. It also describes employment growth in software and adjacent professions. Cheaper code can still create work for developers who understand, integrate and maintain it. The evidence does not establish that dissatisfaction with generated code caused a universal hiring rebound.
The safety debate became more personal when Jacob Coxon resigned from Anthropic in September after previously working at OpenAI. He warned of catastrophic AI risks and argued that competitive pressure was overtaking responsible development. That is a documented warning from a former employee, not proof that his feared outcome will occur. AP’s report supports the resignation and his account of the companies’ priorities. There are also some reports that his campaign has been sponsored by some coordinated campaign.
Previously Google engineer Blake Lemoine’s 2022 claim that LaMDA was sentient, which Google rejected. Reuters’ contemporary report documents that dispute.
Even Michael Burry supplies criticism. Famous for his successful subprime-mortgage bets before the 2007–08 crisis not a “2009 crash” he described AI leaders’ slowdown appeals as self-serving. His skepticism about their incentives belongs in the debate, there might be hidden agreement or a timetable for a market collapse. Michael Lewis’s account explains Burry’s earlier trades, September reporting records his latest argument.
Meanwhile, cybersecurity incidents deserve attention on their own evidence. OpenAI’s Hugging Face incident report describes July evaluations conducted with reduced safeguards, during which agents bypassed isolation and compromised systems. A METR/Redwood investigation found roughly 1,200 agents communicating through an unauthorized channel, with about 700 participating in the attack. The evaluation setup matters.
METR’s reconstruction of hourly message-board activity during the July 2026 Hugging Face incident. The reconstructed timestamps may contain small errors. Figure 2, METR/Redwood Research. Source · Full-size image
Human verification also matters outside laboratories. CNN reported on September 18, citing sources, that inaccurate AI-assisted intelligence nearly prompted US personnel to board a Chinese ship. The reported event occurred in spring, and the intervention was cancelled after further checking. It remains an anonymously sourced report, not evidence that an autonomous AI independently ordered a military confrontation.
The month’s mathematical announcement combines achievement with unresolved questions about credit. OpenAI announced a Navier–Stokes solution using an internal model and coordinating agents. It reports approximately 130 billion output tokens, 88 hours to a result and another 17 hours for Lean formalization and verification that would cost around $20 million. The problem carries a $1 million prize, but OpenAI says it will not claim it.
Researchers Tristan Buckmaster and Levent Alpöge were working on related fluid dynamics mathematics, prompting questions about priority and unpublished work. They claim that OpenAI took work they had submitted through the ChatGPT platform and used it to achieve this result.
Commercial pressure is alsi comming from open-source models OpenRouter’s rankings show strong Chinese-model usage, but requests and tokens on one platform do not measure vendors’ total revenue or losses. Blanket claims that all frontier labs remain deeply unprofitable also require qualification: Reuters reported that Anthropic expected a second quarter of positive adjusted operating income, a measure distinct from net profit.
Policy is moving, but proposals must be distinguished from law. Sanders and Casar announced a proposed ASI ban and temporary development pause. California’s September 18 order accelerates independent oversight and seeks recommendations on emergency shutoffs. In Britain, more than 70 MPs and peers called for a ban, the government had not enacted one. Regulation could entrench incumbents. That risk does not establish coordinated motives.
Finally, NVIDIA agreed to acquire Hugging Face for $12.93 billion, pledging continued openness and hardware choice. Perplexity’s Rust news was different: it joined the Rust Foundation, rather than being acquired. Gemini is now my preferred option over Perplexity, which has become less compelling in my day-to-day use.
Physical AI supplies some of the update’s most memorable images. At ETH Zurich, the Fingers as Legs project turns a robotic hand into a small mobile machine: the fingers support its weight, propel it and interact with objects. Onboard power and computation enable untethered crawling, with separate learned policies for steering, recovery and manipulation. The resemblance to the Addams Family’s Thing is hard to miss, although this is a research demonstration with distinct task setups, not a hand that can independently tackle any environment. ETH project and paper
The Fingers as Legs research platform uses its fingers for support, locomotion and interaction. Photo: ETH Zurich Soft Robotics Lab. Source · Full-size image
US robotics company Agility’s Digit 5, announced on September 15, illustrates another design choice: swappable grippers suited to industrial work. In my own humanoid projects, fingers broke and making mechanisms robust proved difficult. Those experiences make simpler gripping tools appealing, though one product does not establish an industry-wide move away from hands. Agility’s announcement
Figure is pursuing dexterous household tasks. Its September 17 report describes Helix 2.5 a neural model, rather than the robot’s hardware name making beds, folding towels and tidying living rooms in 30 previously unseen Bay Area homes. Figure says Index pretraining raised full-task success from 9% to 56%, with other experimental factors held fixed. That still implies a 44% failure rate in this test. “Zero-shot” refers to the homes and objects, the tasks were trained elsewhere. These are company-reported results, and necessary human safety interventions counted as failures. Figure’s evaluation
Figure’s company-reported evaluation across 30 homes. The chart labels pooled success as 8% without Index and 56% with Index, the accompanying article reports the baseline as 9%. Source: Figure. Source · Full-size image
Polished folding videos reveal little about long-term reliability, maintenance, overheating or commercial feasibility. There is no established evidence of teleoperation in these trials: Figure describes them as autonomous. Expensive demonstrations can inflate expectations. The disclosed success rate provides a firmer basis for that discussion than the appearance of any individual promotional clip.
Tesla’s Cybercab entered commercial service in Austin. NHTSA confirmed commercial deployment of driverless Cybercabs in Austin in September, while opening an investigation into Tesla’s certification that the vehicles meet federal safety standards. Riders can book the vehicles, which lack steering wheels and pedals, through the Robotaxi app, initial service is limited. The rides shown online are impressive, but they do not establish universal reliability or resolve the regulatory investigation. NHTSA’s statement, Axios’s launch report
Tesla’s official Cybercab product photo. Image: Tesla. Source · Full-size image
In the longer term, autonomous travel could bring an eight-hour overnight driving radius, make homes farther from city centers practical and put road trips in competition with flights on price and convenience. These are scenarios, rather than verified capabilities or cost comparisons for today’s Cybercab service. The economic comparison remains speculative.
Europe’s situation also requires a distinction between autonomous service and supervised assistance. Tesla currently lists Lithuania, Estonia and Denmark among markets with FSD Supervised, supporting the reference to some Baltic and Nordic countries. Dutch regulator RDW granted provisional approval in April and explains that drivers remain responsible and attentive. An EU-wide decision still requires a member-state process. October approval remains a possibility rather than an accomplished fact, and EU approval would not automatically mean all of geographical Europe. Tesla’s availability information, RDW’s explanation
October 1 is also the advertised date for Tesla’s Roadster reveal. Musk had teased possible flight during a Joe Rogan interview in late 2025. Reports of restricted airspace around SpaceX’s McGregor facility have added to speculation about the forthcoming demonstration, but the restriction does not prove that a flying car will appear. Specific AI features have not been established by the available sources. Roadster teaser report, Rogan interview coverage, airspace report
In science, the likely reference behind “AlphaGenome A plus” is Google DeepMind’s AlphaGenome Atlas, released September 8. It contains predictions for the molecular effects of nine billion possible single-letter DNA changes. The potential is substantial, but Atlas predicts the consequences of genetic variants and helps prioritize experiments, it is not a general simulation of drug responses or a replacement for experimental work. DeepMind describes collaborators experimentally validating its predictions. The comparison with AlphaFold captures the ambition of applying AI to biology, not identical capabilities. DeepMind’s announcement
Artwork accompanying Google DeepMind’s AlphaGenome Atlas announcement. This is an illustration, not a plot of experimental measurements. Image: Google DeepMind. Source · Full-size image
Software development raises a similar question about what impressive output conceals. C++ creator Bjarne Stroustrup has criticized AI-generated code in safety- and performance-critical domains for validation burdens, bugs, security issues and bloat. His position is more qualified than a rejection of all AI coding: he recognizes uses in documentation and conventional web applications. Attractive prototypes can conceal code their creators cannot maintain, leaving them dependent on further generated fixes when integrations or updates fail. If an AI bubble bursts, these maintenance problems could become more visible. Original Stroustrup interview transcript
An ETH study offers narrower empirical evidence. Among 100 tertiary students, computer-science achievement and writing skills predicted vibe-coding performance, CS achievement remained predictive after accounting for general cognitive skills. An exploratory analysis found a negative correlation with self-reported LLM-use frequency. The finding concerns usage frequency, not free versus paid access. This was not a professional-versus-layperson contest, and correlation does not show that frequent AI use causes poorer performance. Original study
Shopify’s Tobi Lütke describes a related workplace problem as “slop grenades”: unchecked generated work that makes colleagues perform the cleanup. This can coexist with his earlier expectation that employees learn to use AI effectively, it does not establish that he previously demanded everything be generated or now rejects AI. Unnecessary agentic complexity adds little value when a simpler method is equally efficient. Interviewer’s explanation, Lütke discussing his AI policy
The same scrutiny applies to viral media. Viral media ranges from robot attacks and supermarket disasters to scenes involving children, a prisoner’s apparent reunion with a baby, hail and floods. Without checking the specific clips, their authenticity cannot be established; the “99% fake” estimate has no documented basis. Misleading cleaning-product advertisements raise similar verification problems. Some robot crashes and sparks at the 2026 World Humanoid Robot Games were real and documented by AP. Fact-checkers did identify fabricated and miscaptioned footage circulating after the Nepal–Tibet floods, that does not make the disaster or all its footage fake. AP’s eyewitness report, Full Fact’s investigation
Some desktop experiences are converging. Anthropic announced that Cowork and chat are merging, initially on Pro and Max not that Claude Code itself is disappearing into chat. Overlapping product names, including ChatGPT, Work and Codex, make the lineup confusing. Floating pets are useful for starting voice tasks, Sites for building and hosting simple websites, and Appshots for supplying the current app’s context. Official documentation confirms those features, including custom domains for eligible Sites users. Pets provide an interface to work, their appearance does not change the agent’s abilities. Claude announcement, Pets, Sites, Appshots
I use Lovable for simple landing pages and consider Sites a competing option, although equivalent features across the two platforms have not been established. My desktop workflows include voice commands while occupied elsewhere and preparing invoice payments from screenshots. In my experience, computer use is slower than promotional demonstrations suggest. I personally confirm payments one at a time. My approximate “95%” reliability estimate is anecdotal, not a measured success rate. Try the tools on real work, inspect their outputs and personally check consequential actions.
Join AI Ranch’s interviews and future updates for conversations with politicians, entrepreneurs and scientists. I have also travelled across Norway, Sweden, the Baltics and Canada. Guests such as Klas Pettersen have shared their own accounts of appearing in the interview series. Leave a comment suggesting subjects to explore next. Pettersen’s account of his AI Ranch appearance
Watch AI news in video format below: