# Astra marks the Claude 3.7 moment for knowledge work > Source: > Published: 2026-09-11 05:20:33+00:00 GPT-6 Astra marks OpenAI’s first major version change since GPT-5 launched thirteen months ago. OpenAI points to three firsts to explain the new series. It is the first model pre-trained on more than 100,000 GPUs at the Stargate site in Texas, the largest training run in the company’s history. It is the first in which an earlier model played a substantial role in the training process. And it is the first flagship to ship with computer use as a core capability. Released on September 3, Astra is now broadly available to paid ChatGPT users, through the API, and on Azure and Bedrock. Pricing is $10 per million input tokens and $50 per million output tokens, 2.5 times Sol and the same as Fable 5.1, with a 1.05M-token context window. **Our view is that Astra marks the Claude 3.7 moment for knowledge work, meaning white-collar work broadly, and that the potential market is an order of magnitude larger than coding.** Claude 3.7, released in February last year, was the point at which agents first became usable. Claude Code launched alongside it as a research preview, Manus followed a month later, and Operator and Deep Research arrived in the same quarter. For the first time, developers could hand over a whole task rather than prompt line by line. Those early agents were rough. They drifted and needed supervision. The breakout came nine months later with Opus 4.5, when reliability improved enough for coding to move from trials to large-scale enterprise adoption, and revenue reached the tens of billions of dollars. Astra’s computer use is not yet as reliable as Opus 4.5 was for coding, but the trajectory is visible, and the market it addresses is far larger. One step further out, the near-term commercialization of coding and knowledge work initially targets existing labor budgets. Recursive self-improvement (RSI), the process by which AI accelerates its own research, points instead toward AI for science, which could create entirely new sources of economic value. Astra’s gains in mathematics and scientific research are another sign of general-purpose models beginning to contribute to science. On public benchmarks and demos, Astra is clearly ahead of Fable 5.1 in computer use, spatial tasks, and mathematical and scientific research. It remains in the same tier on coding, and its general intelligence index is still below Fable 5.1. For context, Anthropic is expected to make its S-1 public by the end of September; Fable 5.1 is the pre-IPO release, and Astra followed roughly a week later. #### 1. Computer Use Opens a Larger Market Computer use means the model operates the same software people use at work. It reads the screen, moves the mouse and types, working inside Excel, CAD tools, Blender or a browser instead of calling an API from a script. This has long been a weak spot. Even moving a PowerPoint text box could require repeated code edits, whereas a person simply drags it. Astra is becoming much better at acting directly within those applications. Its OSWorld score rose seven points in a single generation to 72.6%, and average time per task fell from 75 minutes to 40. On Agents’ Last Exam, which tests professional work such as financial modeling, engineering, and media production inside real software, Astra scores 59.3% against 55.5% for Opus 5, while using 65% fewer output tokens. The demos illustrate progress at different stages of a workflow. In specialized software, Astra opens KiCad and lays out a printed circuit board in under three minutes, a task that until now required extensive manual work. Across office applications, it builds a Power BI dashboard from raw data, completes a Form 1040, formats a legal document to a specification, and finishes a job-search research task that would take a person five hours in under three minutes. In spatial and creative tools, it reconstructs an editable 3D interior from a single photograph of a house, and models a scene in Blender before importing it into Unreal as a walkable environment; these were the demos that spread furthest in launch week. It can also close the loop with verification: building a website and then running front-end QA on it, or installing and testing software and checking its own results. Inside OpenAI, Astra already takes on experiments in the company’s codebase that used to occupy a researcher for a week, from idea to code to a completed run. The most striking change this week is who is using it. When coding models arrived, the enthusiasm came from programmers posting terminal screenshots. This time the feeds are full of architects, 3D modelers and video editors, people who had assumed AI had little to do with them, because a model that writes excellent code still could not work inside their CAD or Blender. In many professions the barrier to entry is fluency in a particular tool: AutoCAD, Excel, Final Cut, KiCad. As specialized software becomes easier to use, professional judgment and the ability to verify results become more valuable, and these professions will see their workflows reorganized much as programmers’ were over the past year. On market size: the addressable market for coding is the wages of tens of millions of programmers plus global software budgets, roughly a trillion dollars, of which tens of billions have become revenue over the past year or so. The addressable market for knowledge work is the wages of hundreds of millions of white-collar workers plus their tooling budgets. Wages alone exceed $10 trillion. Coding models reach the first group; computer use reaches the second, which is where the order-of-magnitude difference comes from. Because we are still at the 3.7 stage, we count this as potential. The first use cases likely to gain traction are those with structured inputs and verifiable outputs: business intelligence, financial modeling, and spreadsheet workflows. The financial-modeling tasks in Agents’ Last Exam and the Power BI and Excel demos fall into this category, and they represent an early source of revenue beyond coding, before broader automation becomes reliable. Enterprise usage falls into roughly four types. Operational tasks: filling in forms, updating CRM records, managing calendars, researching online and writing the summary into an email, building a site and running QA. Document production: generating decks, documents, spreadsheets, and analysis from existing templates. Astra scores 41.4% on AutomationBench, against 18.1% for Sol and 31.4% for Fable 5.1, and Hebbia measured 17% better adherence to the brief and 19% higher citation accuracy when building decks. Research: On Perplexity’s research-workflow benchmark, Astra scores 13.5% higher than Fable 5.1 at 6.1% lower cost. Specialized professional tasks: Harvey reports a clear improvement over Sol on complex legal tasks, with the model distinguishing documents from the established record, flagging unsupported assumptions and turning gaps into concrete drafting positions; Jane Street reports a clear gain on its trading-intuition evaluation; and on the science side, Terminal-Bench Science is 64.6% against 52.6% for Fable 5.1. Our RSI report described a progression from language models to chain-of-thought (CoT) reasoning, to agents, to continual learning, and finally to self-improvement, with each stage creating new commercial opportunities. Coding corresponds to the agent stage, and Astra extends agent capabilities to a much wider range of applications. The next stage, continual learning, is the model accumulating a company’s tacit knowledge in deployment, and it is being addressed. Context windows have reached 1M tokens; Codex now keeps notes across context windows and can retrieve earlier ones, an early form of continual learning. OpenAI’s Frontier team and its consulting partners serve a similar purpose by working on-site with customers to supply the necessary business context. Coding’s own history is instructive. Claude 3.7 made agents usable; nine months later Opus 4.5 delivered reliability, 4.6 made long tasks stable, and Fable got long-running tasks and merge-ready code right. Each step felt spaced out at the time, but in hindsight the path is clear. We describe it as agentic scaling driven jointly by the model and the harness, the tools, memory and verification systems that surround the model. With each generation the model learns more long-horizon work in reinforcement learning (RL), and the harness gains another layer of tooling, memory and checks. Together, these advances took agents from being usable with Claude 3.7 to producing reliable results with Opus 4.5, and set the stage for this year’s rapid adoption. The computer-use harness is only getting started. Codex’s cross-window memory and the verification and rollback built specifically for computer use are first versions. There is good reason to expect knowledge work to follow a similar adoption path. **Computer use is the foundation for extending the model into more domains. Once that foundation is in place, the market it opens is on the order of ten trillion dollars.** #### 2. AI Research Begins to Accelerate Itself Astra is the first flagship in whose development an earlier model played a substantial role, and that fact matters more to our RSI thesis than any single benchmark. OpenAI’s RSI progress note on September 6 put numbers on how fast this loop is turning. The goal set last fall was an “automated research intern” by September, a system that can complete tasks that would take a skilled researcher a few days, under human direction. By internal measurement it has been reached, and the next target is an automated AI researcher by March 2028. The median researcher in OpenAI’s research organization now consumes more than $600 a day in inference tokens, and researchers in the top decile of usage consume $7,000 worth of inference tokens per person per day. Before June, total agent hours were below human hours; by mid-August, each human workday corresponded to 3.1 agent workdays. Experiments per researcher reached a record in August. Success rates on four-to-eight-hour tasks are rising, although more than half of the successful tasks in that range involved at least one human intervention. In OpenAI’s own words, compute is one constraint on progress, and its importance may grow as other bottlenecks ease. RSI has implications beyond the commercialization of coding and knowledge work. Near-term commercialization of coding and knowledge work mainly displaces existing labor spend: the programmer’s work, the analyst’s work, the designer’s work. However large the pool, substitution comes first and expansion second. One aim of RSI is to extend the acceleration of AI research to scientific research more broadly, and the long-term value of AI for science lies mostly in new output: new drugs, new materials, new energy sources, higher crop yields, results that were previously out of reach. Astra shows early signs. It solved 2 of 68 open Erdős problems; it improved one bound on large prime gaps by a loglogn factor after more than eighty years without progress; Terminal-Bench Science, run in real research environments, rose from Sol’s 22.4% to 64.6%; and sequencing quality checks and variant visualization now run unattended inside specialized software. OpenAI is targeting an automated AI researcher by March 2028. Some of the tools and methods developed to automate AI research could also prove useful in other scientific disciplines, but progress there will still depend on experimental data, validation facilities, and the cycle time of physical experiments. **The loop in which models build models is already turning, and compute is one of its main constraints. Coding and knowledge work mostly capture existing labor spend; AI for science, driven by RSI, creates new output.** #### 3. Where Astra Leads, and Where It Does Not Start with the results that are easiest to grasp. Of 68 open Erdős problems in mathematics, no model had solved any; Astra solved two, and it advanced a bound on large prime gaps that had stood for more than eighty years. On FrontierMath’s hardest tier, Tier 4, it scores 97.6% against 87.8% for Fable 5.1. Terminal-Bench Science, which runs full workflows in real research environments, gives Astra 64.6% and Sol 22.4%. BenchCAD, from design drawing to CAD code, is 95.9%. ARC-AGI-3, built specifically to test abstract reasoning, was below 1% at the end of last year; Astra reaches 99.9% with the official harness, and François Chollet says the benchmark saturated twice as fast as he expected. OpenAI’s official table is organized in four blocks, with each model’s best score at any effort level. On coding, Astra has not pulled ahead. Artificial Analysis’s Coding Agent Index gives it 67, roughly in line with Opus 5 and Fable 5 and below Fable 5.1’s 70. Cognition’s own FrontierCode run puts it within 0.4 points of Fable 5, at 64% lower cost. Developers who have used it tell a consistent story: it runs 2,000 steps overnight without breaking and gets UI geometry right, but it tends to over-engineer, returning a large pull request when asked for a targeted fix, and its writing is slightly weaker than the previous generation. On the general intelligence index it scores 61, level with Sol and five points below Fable 5.1. On AA-Omniscience, a benchmark built specifically to measure hallucination, the rate fell from 92% for Sol to 51%, a meaningful improvement for enterprise use, though it reflects one evaluation under one configuration and should not be read as the error rate of everyday answers. On cost, Artificial Analysis measured about 10% fewer output tokens than Sol at maximum effort, but at billed rates each task still costs about 75% more. The headroom in computer use deserves separate attention. ScreenSpot-Pro at 92.7% suggests that locating elements on screen is largely solved. OSWorld 2.0 at 72.6%, on the offline set with partial credit, shows that complete workflows still have room to improve, with the gap concentrated in the reliability of multi-step tasks. A related signal comes from the UK AI Security Institute’s mathematics evaluation: without emitting a chain of thought, the difficulty at which the model succeeds half the time, measured in human solving time, rose from 3.6 minutes for Sol to 30.9 minutes, a clear step up in implicit reasoning. The next one to two model generations have a real chance of delivering a substantial improvement in multi-step task performance, depending mainly on progress in RL training and in the harness. That would be computer use’s 4.5 moment. **What remains is multi-step reliability, which RL and the harness will have to supply together. Once that gap closes, computer use will have its 4.5 moment.** #### 4. Why Compute Demand Could Keep Growing According to industry conversations, GPT-6 used roughly ten times the training compute of the GPT-5 generation. As a rule of thumb from scaling laws, an order of magnitude more compute corresponds to about one model generation. OpenAI’s public statement is pre-training on more than 100,000 GPUs. An easily missed point is that this figure refers to the pre-training hero run, the single largest complete pre-training job, which is not where most of a generation’s compute goes. Based on our industry discussions, experiments, RL, synthetic data generation and supporting infrastructure can together consume several times as much compute as the main pre-training run, potentially as much as ten times. OpenAI’s September 6 note points the same way: experiments per researcher hit a record in August, and agent hours across the research organization are 3.1 times human hours. After Astra was assessed on August 7 as potentially reaching the Critical threshold for cyber capability, Astra-class GPU allocation fell a further 59% that week, while GPU allocations to other model classes rose by roughly 17%, offsetting about 85% of the decline in Astra-class allocations. This shows that compute can be quickly redirected to other research workloads. GPT-6’s parameter count has not reached 10 trillion, and this generation’s additional compute went mainly into post-training and RL. We therefore see substantial room for further scaling in pre-training. On timing, the hardware for the next round of pre-training arrives in volume between the fourth quarter and the first half of 2027. Vera Rubin ships its first units in the third quarter, ramps in the fourth and continues through the first half of 2027, with Rubin Ultra Kyber NVL576 in the second half. Google’s TPU 8t superpod links 9,600 chips with 121 exaflops and 2 petabytes of shared HBM, and the Virgo network connects 134,000 TPUs into a single fabric; general availability is this year, and Anthropic is expected to receive 3.5 gigawatts of TPU capacity in 2027. Add the 400,000 GPUs Jensen Huang mentioned, and this is the base for pre-training scale to rise by another order of magnitude in 2027. Pre-training and inference cycles favor different hardware. Inference puts greater emphasis on memory requirements and cost per token, while pre-training places greater demands on scale-up fabrics and inter-rack connectivity. The training footprint of very large models cannot be estimated from weight capacity alone. Gradients, optimizer state and activations together are several times the weights, and as parallelism increases, rack-to-rack communication carries a growing share of the load. NVL576 is now confirmed to use optics for scale-up at the second tier, and the rumored 6D torus on the TPU side would raise wrap-around optical ports per chip from 1.5 to 6. This supports a central argument in our RSI report: training demand recurs, and pre-training becomes a continuing production requirement rather than a one-time cost. In a pre-training cycle, the biggest beneficiaries are large-scale interconnect and optics. **An order of magnitude more compute corresponds to roughly one model generation, and pre-training still has room to scale. Per industry sources, the hero run is the smaller part; experiments, RL and synthetic data are the larger part. Once TPU 8t and Vera Rubin ship in volume, the next pre-training acceleration can begin, and large-scale interconnect and optics benefit most in that cycle.**