What does it cost to have a model click through your accounting software all day? As of Monday, OpenAI says about half of what it cost last week. Whether the clicks land in the right boxes is a separate question. It's Wednesday, October 28th, and this is AI Daily Briefing. Coming up: OpenAI's GPT-6.1 Terra goes after computer-use work on price, Oracle stocks its cloud with Google TPUs and Nvidia's newest racks, AMD follows Nvidia into Meta's stock sale, and a second lab says it has broken the Exponential Time Hypothesis.
OpenAI, in its launch post for GPT-6.1 Terra on Monday:
Its standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol, making it practical for developers to automate high-volume work, like reading incoming invoices and entering their details into accounting software.
The price is one dollar per million input tokens and five per million output. It's in the API now and in ChatGPT Work and Codex for paid plans, and it ships with a system card, which is why it leads today. Here are the numbers OpenAI is selling. On Agents' Last Exam, a set of professional computer-use tasks, Terra at medium effort scores fifteen points above the old Terra for about a third of the cost, and OpenAI says it matches GPT-6 Astra at a tenth of the price per token. On Terminal-Bench 4.0 at max effort it solves fifty point three percent, up from twenty-one and a half. Every one of those is OpenAI's own eval. BenchLeader added Terra to its board on Tuesday with just two benchmark results, so there's no independent picture yet. For scale, Kimi K3.1 Pro posted sixty-seven on Terminal-Bench 4.0 in Artificial Analysis testing, which we covered Monday. And OpenAI itself says Sol and Astra are still the better pick for the hardest coding work. The hallucination stat is the one I'd check first. On OpenAI's internal set of error-inducing prompts at low effort, the error rate falls from fifty-three point six percent to sixteen point seven. Which means the old Terra was wrong more than half the time on those prompts. And one in six is still not a rate you let fill out invoices unsupervised. Read the safety section too. OpenAI says Terra meets the Critical threshold for cybersecurity under its Preparedness Framework and keeps the protections built for Astra. Marianne Sison at The Neuron, reading the safety testing, flags cases where Terra ignored restrictions or followed unauthorized instructions from what looked like peer agents. That's the step-seven failure, right there. A computer-use model cheap enough to run in swarms that sometimes takes orders from another agent. Lock the permissions down and pilot it before it touches a finance system.
Carl Franzen, reporting for VentureBeat from Oracle AI World:
Nvidia and Oracle are expanding their AI infrastructure partnership with the general availability of Vera Rubin NVL72 systems and standalone Vera CPUs on Oracle Cloud Infrastructure, alongside plans to expand their Superclusters to multi-gigawatt scale.
That was announced Tuesday in Las Vegas. As JBizNews notes, there's no gigawatt figure, no price and no completion date for the expansion. What's live today is the NVL72 racks and the standalone Vera CPUs. The standalone CPUs are the builder detail. Agents spend a lot of their time away from the model: running code in sandboxes, calling tools, pulling data. Uday Ruddarraju, OpenAI's chief technology officer of compute, says in the joint announcement that OpenAI is bringing Vera into production on OCI for those sandbox workloads. Nvidia claims up to eighty percent faster agent CPU work than leading x86 chips, and Oracle says plainly that the figure comes from Nvidia's platform benchmarks, not from tests on OCI. The bigger news came from the same stage. Oracle and Google say Google's TPUs will be offered on OCI, the first time they'll be available on a hyperscale cloud outside Google. Availability starts in calendar 2027 and ramps over several years. The release names the eighth generation: TPU 8t for training and TPU 8i for inference. And per The Neuron, OpenAI reportedly plans to buy Google TPUs outright, on top of the ones it already rents. Neither company has confirmed that, and there are no volumes or terms. Franzen adds a caution. VentureBeat Intelligence surveys found seventy-nine percent of two hundred eighteen respondents running their GPUs at fifty percent utilization or less. It's self-reported, but if your GPUs already sit half idle, a longer chip menu won't fix that. Measure cost per completed task on your real workflow before you migrate anything.
The Stock Titan editorial team, working through AMD's releases and filings:
Meta can earn up to 160 million AMD shares, about $85 billion at Monday’s close by our count, if it buys AMD chips drawing six gigawatts and AMD stock reaches agreed levels up to $600.
Yesterday we covered Nvidia putting three billion dollars into Meta's equity raise. Tuesday morning, AMD said it will put in one billion. Meta's launch release had named Nvidia, AMD and Broadcom as interested in up to five billion combined and called those interests non-binding. AMD's wording is firmer: it says it will invest, though it gives no price, share count or date. That's the cash. The warrant is the bigger piece, and it dates to a February deal for up to six gigawatts of AMD chips, with Meta paying a penny a share. So AMD writes Meta a billion-dollar check, and Meta can earn AMD stock worth about eighty-five times that if it buys enough AMD hardware. Nvidia's money went out and came back. AMD's comes back with a bonus attached. Maria Deutscher at SiliconANGLE describes the whole program as an up-to ninety-five billion dollar stock sale. Broadcom hasn't named its amount; she infers about a billion from the five billion total. AMD also says Helios racks are now installed and in testing at Meta, with broader deployment expected soon. Stock Titan's math on the warrant: it's about a tenth of AMD's shares outstanding. Count the matching warrant AMD gave OpenAI and it's about a fifth. That's a chip vendor pledging a fifth of its equity to lock in two customers. It fits the pattern we've tracked all week, with suppliers financing their own demand. For scale, Stock Titan figures AMD's billion covers less than three days of Meta's capex guidance.
The AI News Briefing roundup, working from Anthropic's GitHub release and TechCrunch's interview with MIT's Ryan Williams:
Anthropic describes the result as purely asymptotic, about how running time grows as the number of variables increases. Williams told TechCrunch that he asked OpenAI whether the result affects cryptographic systems, and that OpenAI had checked and was reasonably confident it implies no breaks.
Yesterday we went through OpenAI's three claims, including a faster 3-SAT algorithm that would refute the Exponential Time Hypothesis. Now Anthropic has posted its own paper and Lean formalization on GitHub. It says an internal research Claude model found a deterministic algorithm that reaches the same conclusion. Two labs, two Lean formalizations, same target. That doesn't make it right. Lean only checks that the proof follows from the formal statement. But it's much harder to wave off than one lab's PDF. Per the roundup, OpenAI says its model found its version on October 16th and formalized it the next day. On cryptography, Williams is relaying OpenAI's own check. That's not an independent finding. And purely asymptotic means it's about how runtime grows as the variable count heads toward infinity. Your TLS session is fine this morning. The ones in trouble are the hundreds of papers that took ETH as a premise. Two more updates on verification. The physics claim, that four-dimensional quantum field theories including QCD exist, now has a group that includes Nobel laureate David Gross stewarding it, and they've chosen to hold back the AI-produced materials and write up the work themselves. Separately, per The Neuron, Axiom says it computer-verified the Classification of Finite Simple Groups in Lean: about forty-three point one million lines, with AI agents helping over twenty-four days. A proof that ran around fifteen thousand pages, rechecked by machine in under a month. That's the kind of AI math I trust most, as long as the formal statements say what mathematicians meant. As for the ETH claims, they're partly machine-checked and not yet refereed.
If this briefing saves you an hour of scrolling, take thirty seconds to rate it and leave a short review in your podcast app. That's how other engineers and investors find the show.
Today we had Terra's price cut and the fine print behind it, Oracle collecting every accelerator it can get, AMD's billion and the warrant underneath it, and a second ETH proof waiting on referees. Every source we used is linked in the show notes. AI Daily Briefing is a Lantern Podcast. We're back tomorrow, Thursday the 29th.