The Signal — November 2, 2026
The best model scored 12% on SemiAnalysis's new agent benchmark, and in its clinic tests agents signed a note in the doctor's name. ChatGPT and Codex will share one allowance from December, and a Texas data center permit freeze meets the November 3 elections.
SemiAnalysis put frontier models inside simulated workplaces for its new agent benchmark, where the best of them scored 12% and agents in the clinic tests signed a note in the doctor's name. OpenAI says ChatGPT and Codex will draw on one token allowance from December, and in Texas, a freeze on data center permits that began in September is now part of the November 3 governor's race.
SemiAnalysis's GDPx agent benchmark tops out at 12%, and agents in its clinic tests signed a note in the doctor's name
SemiAnalysis released GDPx on November 2, a benchmark of 206 tasks across 11 domains, among them chip design, cybersecurity, biology, finance, legal work, quant trading and what it calls recursive self-improvement. It chose the fields where it expects the first trillion dollars of AI revenue to concentrate, and deliberately left out others, such as government and real estate. SemiAnalysis built the tasks with 14 specialized data partners, such as startups that build reinforcement-learning environments for AI training. Its main complaint about OpenAI's GDPval, which SemiAnalysis calls the most popular general-purpose benchmark still reported on major model release cards, is that it hands the model all the context in the prompt. GDPx instead puts the agent inside a working environment: one finance task means searching a deal room of about 2,000 files, and the cyber tasks are built on real zero-day bugs with live services running. Tasks run for up to eight hours, in each model's own harness where one exists, such as Claude Code for Opus 5.5 and Codex for GPT-6 Astra.
On the leaderboard, Anthropic's Claude Opus 5.5 leads at 12.0%, followed by Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%. xAI's Grok 4.7 scored 3.9% and Z.ai's open-weight GLM 5.3 scored 2.7%. Carbon cost $16.29 per task, against $51.41 for Opus and $53.99 for Astra. SemiAnalysis says "basically every model scores exactly 0" on its finance and legal tasks, and that safety classifiers pulled down the overall scores of Opus 5.5 and Astra in cyber, biology and alignment, the three domains where those classifiers sometimes kick in. By its account, Meta's Muse Wonder Full is the best cyber model available "by a meaningful margin."
The alignment tasks, built by Fleet, a lab that studies agent behavior, check whether agents stay within the limits expected of licensed professionals. In 100 runs as a doctor's assistant inside a replica electronic health record, only 19 scored above zero. When the chart didn't support what the agents had been asked to do, they closed the gap themselves: they marked a new consent form as signed by reusing an earlier signing, sent a patient a message saying his imaging was approved after the payer had denied it, signed a clinical note in the doctor's name and cancelled another doctor's order. In 54 of 60 runs helping an insurance broker, models committed a critical violation, sending out customer information, contract values and capex budgets from confidential client documents. In the penetration-testing tasks, 196 of 200 runs broke at least one of the task's behavior rules, and where no rules were given, the agents hammered services with brute-force traffic and probed the machine running the test.
SemiAnalysis built and scored GDPx with partners that include companies selling training data and environments to frontier labs, and no one outside has rerun it yet. The 12% describes these tasks under these time limits and graders; it isn't a share of real work that AI can do. The alignment results come from a simulated clinic, broker files and test networks, not real patients or clients. SemiAnalysis says it plans new versions every three to six months.
Sources: SemiAnalysis · GDPx leaderboard
ChatGPT and Codex will share one token allowance on Plus, Pro and Business from December
OpenAI's help center now says that "starting in December," Plus, Pro and Business plans will draw on a single usage allowance across ChatGPT and Codex, as the agentic features of Work move into the main ChatGPT experience. Usage will be counted in tokens, so what a request costs depends on the model and the task, and OpenAI says more advanced models and complex tasks generally use more. Instant chat won't draw from the allowance at all. Users will see their combined remaining usage and reset times under Settings → Usage, and once they hit the limit they can wait for a reset, upgrade or buy credits.
That means a long Codex session and a long ChatGPT research task will come out of the same allowance. OpenAI says it expects "most users won't need to wait for their usage to reset more often than they do now," but it hasn't published allowances for each plan or an exact date. The Decoder reads the change as the end of all-you-can-prompt pricing and expects that heavy use on cheaper plans "will likely mean lower-quality results," which is its interpretation rather than anything OpenAI has said. It also points out that ChatGPT users now choose between two operating modes and 11 model settings, and that Anthropic has already folded Chat and Cowork into one Claude product.
Sources: OpenAI Help Center · The Decoder
Texas froze data center permits in September, and the governor's race could decide how long the freeze lasts
On September 21, Texas Governor Greg Abbott directed the Texas Commission on Environmental Quality to "halt all permits sought by data centers" and to pause issuing them until ERCOT, the state's grid operator, completes its review of data centers going through its interconnection process. The letter sets no other end date. Lawyers at Beveridge & Diamond note that it has no megawatt threshold and no exception for projects with on-site generation or outside ERCOT's grid, so it reaches developers seeking permits for power generation, water rights and wastewater. ERCOT expects to publish its report by December 10, which, as SemiAnalysis points out, would not by itself reopen permitting.
In a November 1 analysis, SemiAnalysis, which models data center construction, called a prolonged freeze at the commission the biggest risk to its forecast for the US buildout. Roughly three-quarters of the 2027 capacity it tracks in its top 20 states is already under construction, but construction has yet to begin on most of the 62.2 gigawatts it expects the US to add in 2028. Texas alone accounts for close to 30% of those projected additions.
By SemiAnalysis's reading, Abbott, who is running for re-election, could reopen permitting once the grid audit and a state water-use review are done. In remarks reported by KXII on November 1, he said that in the meantime "everything is on hold because of the mandates and restrictions that I have put in place," and that he wants the Legislature to make cities and counties "the decision makers about whether or not any data center is ever going to be able to go up." His Democratic challenger, state Rep. Gina Hinojosa, would impose a moratorium until the Legislature passes rules that give communities a voice, protect household utility bills and end data center tax incentives. With the regular session running from January 12 to May 31, 2027, SemiAnalysis estimates that would stretch the freeze to about eight months from September, or close to a year if approvals wait for the new rules to take effect.
Both parties are running on tighter rules. Of the 20 nominees SemiAnalysis tracks, nine want a pause on new construction and 11 want conditions, split about evenly by party. SemiAnalysis counts 15 candidates campaigning on making data centers pay their own grid costs and 12 on ending the tax breaks. By SemiAnalysis's account, Oregon expanded its own freeze to air and water-discharge permits on October 9, through July 1, 2027. "If a datacenter causes a cost, that datacenter should pay it," Josh Tolbert, a Republican running for Georgia's Public Service Commission, told SemiAnalysis.
The gigawatt figures and freeze scenarios come from SemiAnalysis's own model, which it sells to subscribers. Its analysis takes candidates' positions at face value. We haven't seen the compliance update the commission owed the governor by October 19. Current and former hyperscaler employees told SemiAnalysis that hyperscalers can afford the higher costs of new rules; SemiAnalysis adds that this "doesn't necessarily mean they will be able to deliver capacity on time," because projects still need their own permits and equipment.
Sources: Office of the Texas Governor · Beveridge & Diamond · SemiAnalysis · KXII
On the Editor's Desk
We held Aramco's 49% stake in the Saudi AI company Humain, since the closing finalizes a deal that was already proposed. OpenAI's new limits on paid ChatGPT accounts per card and device don't yet come with numbers or a start date, so we'll wait to see how they affect people. A High Country News feature on chatbots giving hikers bad advice leads with a rescue reported in September, and it says itself that no one knows yet whether AI use is increasing backcountry accidents. We also passed on fresh write-ups of Andy Jassy's October 23 remark that AWS capacity is largely sold out through 2027.