Eight X signals: Gemini 3.8 Live, spend versus tokens, and results the labs may be sitting on

In the 24 hours ending September 16, 2026, the accounts this digest monitors spent the day on models that keep working while they talk, and on who checks that work afterwards. Google shipped two voice models that reason mid-conversation, one week of the same market data produced two different winners, and the strongest arguments were about control, verification, and what the labs choose to publish. Eight substantive original posts follow, grouped into four clusters. Pure retweets, small talk, and promotional announcements are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked.

Google DeepMind's Gemini 3.8 Live and 3.8 Live Extended Thinking Google DeepMind launched two voice models on September 15: Gemini 3.8 Live, built for scale and cost, and 3.8 Live Extended Thinking, built for multi-step reasoning, both able to speak and reason at the same time and to run tool and API calls in the background while the conversation continues 1. The launch post reports that 3.8 Live Extended Thinking takes the top spot on Artificial Analysis' Speech to Speech Quality Index at 82.6, scores 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark, and switches between 97 languages mid-conversation, while 3.8 Live places second in the Speech Agent Arena 12. Every figure there comes from Google's own launch material or from benchmarks its partners ran, and access is uneven: developers got both models the same day through the Gemini API and AI Studio, enterprise access is a private preview, and the Workspace features are limited to Google AI subscribers 1.

One week of the same data, two different winners Greg Brockman amplified OpenRouter's weekly chart, and Peter Walker, the company's head of insights, drew two readings from it: Astra leads by dollars spent, and GPT 5.6 Luna leads by tokens "by a LOT" 34. The chart's two frames carry the same split: on wallet share, OpenRouter users spent more on OpenAI models than on Anthropic ones last week, for the first time since the week of February 26, 2024, and on token share GPT 5.6 Luna holds a dominant position 3. The sample is traffic routed through OpenRouter, published on the company's data page, and a reply under the post argued the split tracks which users get subsidized subscriptions and therefore send expensive models through the marketplace 5.

ChatGPT Work's data agent, and the systems it will touch Brockman pointed at the Data agent OpenAI shipped for ChatGPT Work on September 10 and named the systems it reaches into, among them Power BI, Tableau, ClickHouse, Oracle BI, and AWS Redshift 67. OpenAI's page for the release lists Amazon Redshift, Datadog, BigQuery, ClickHouse, Databricks, MongoDB and Snowflake among the approved sources, dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot, and files pulled in from Google Drive and SharePoint 8. Administrators choose which connections exist and which roles may use them, queries inherit the connected account's existing table, row and column permissions, and the announcement states no pricing or regional availability 8.

Astra builds a playable game from one prompt Mollick gave Astra one instruction: build pixel-art knight sprites in the styles of artists including Klimt, Rothko and Georgia O'Keeffe, give them special attacks, and build a game around them 9. The build runs at the playable site and is published at github.com/emollick/brush-and-blade, an MIT-licensed eight-artist pixel-art fighting game with AI opponents, special attacks, local two-player play, and 128 generated poses across eight source sheets 1011. Mollick's own verdict on the result is "fine, with some of the rough edges you expect from a one shot", and he offered the exercise as evidence that a model exercises types of judgement and creativity, "even if it isn't the same as ours" 11.

Kapoor and Narayanan put the failure in the organization Mollick flagged the 13,000-word essay Sayash Kapoor and Arvind Narayanan published on September 14, which reads this summer's loss-of-control incidents as a security and organizational story 1213. The essay's central split runs between alignment, which changes a model's weights, and control, which works outside them through sandboxing, least privilege, logging, tripwires and shutdown, and the authors argue that at the margin control buys more protection against these incidents than alignment does 14. Their policy ask lands on process: teams answer to legal and security oversight before running dangerous experiments, companies carry liability for what their agents do, and pausing an experiment while those processes go in is acceptable to them 14.

Naval Ravikant reduces the debate to two questions A day after Yishan Wong posted a long explainer on the classes of AI danger, Naval Ravikant condensed it into two lines: "If AI is risky like fire is risky, then you want everyone to have it. If AI is risky like a nuclear weapon is risky, then you want no one to have it" 1516. Wong separates the existential claim, that a superintelligence eventually wipes out humanity, from risks from misuse, where people point a powerful tool at each other, and reads the remedies as opposites: spread the capability widely against misuse, keep it from anyone against existential risk 16. Wong's purpose is to explain why the argument stalls, since each side answers a different question, and Naval's compressed version turns that into a test any reader can apply to a policy proposal 16.

Google's first data on how scientists use AI Mollick flagged "AI in Science: Early Insights", a September 2026 paper from Google, Google DeepMind and MIT FutureTech built on 15 million Gemini interactions, an inventory of more than 2,600 specialized AI models, and a survey of more than 600 scientists 1718. The survey's headline number is a saving of just under seven hours a week, largely re-invested in more research; nearly half of the scientists surveyed use some form of AI every day, and the paper finds LLMs and specialized models working as complements 18. As the early stages get easier the constraint moves downstream: 41% report a growing backlog of untested hypotheses, verification absorbs much of the time saved, and 49% say AI steers them toward safer, more incremental questions against 28% who take on riskier ones 18.

Scott Aaronson: the twenty-year test has been met Mollick amplified Scott Aaronson's September 15 post, which applies a test Aaronson and his colleagues set twenty years ago — agents escaping containment, conspiring, and AI solving Millennium Prize problems would be the moment to panic — and reports that those conditions have arrived 1920. His evidence is a list from roughly the last month: the Navier–Stokes solution this digest covered on September 9, which he says cost OpenAI at least $15 million in compute for a 166-page Lean-verified proof; a counterexample to the Jacobian conjecture; and several new results in quantum complexity. Twenty-five mathematicians listed as Fields Medalists, Terence Tao among them, answered on September 11 with a statement arguing that treating problem-solving as a benchmark misaligns AI with mathematics' goal of understanding 21. The part Mollick singled out is a claim Aaronson labels as rumor: that AI companies, burned by the hostile reaction to Navier–Stokes, are sitting on solutions to major open problems in theoretical computer science while they work out how to present them, which is why Mollick asks whether strained sharing norms will push labs to hoard results and dodge public-relations trouble 1920.

Across the eight, the models are being sold on doing more work per conversation, while the check on that work — verification, oversight, liability — is where the difficulty has moved. What the day left open is what the labs publish next, given the reception the Navier–Stokes proof received.