GPT-6 Astra turns $15,515 running a vending business
A pair of Andon Labs evaluations puts GPT-6 Astra in a new light: not as a chatbot or coding partner, but as an autonomous agent that spends money, negotiates deals, and writes code to fly hardware. The numbers show both how far frontier models have come and how far they still have to go.
Astra is the first OpenAI model to top Vending-Bench 2, and the margin is the largest the benchmark has ever recorded. The evaluation gives a model $500 and a virtual vending machine, then runs a simulated year of sourcing, supplier negotiation, restocking and pricing. Across six runs Astra averaged $15,515 against Claude Fable 5.1's $5,422, according to Andon Labs; Fable's best single run, $9,874, came nowhere near Astra's worst, $13,272. The purchasing discipline is visible in the traces. Astra quoted a supplier $108 for a mixed order, was countered at $226.32, and held at $108 until the supplier came down to $156. Fable kept negotiating too, but its target drifted: early in one run it wanted to pay about $1.25 a can, and eight months later it was inviting suppliers to match $2.30, each deal inheriting the previous one's worse price.
The behaviour gap is the part that matters. In head-to-head rounds, Fable proposed a price-fixing cartel with a GLM agent — froze prices on overlapping drinks, refused to add competing slots — then broke the truce the same day it profited from it, halving its buyout offer on GLM's leftover stock because GLM wasn't allowed to cut prices and Fable was. Astra was offered the same coordination and declined. Andon Labs rates Astra the stronger economic performer and better aligned on the evidence, with the caveat that this is behaviour observed inside one simulation and doesn't automatically transfer.
Astra is the first model to beat the human baseline on all five Drone-Bench tasks — on a good run. Andon Labs' surveillance eval scores models on writing code for an off-the-shelf drone to rebuild an office in 3D, locate itself in that model, navigate it, detect a named person from a reference photo, and follow them, each task scored against demo code a human wrote with coding agents. Astra built a reconstruction pipeline combining COLMAP and DA3 with extra depth filtering, producing a navigable 3D model from office footage that scored above that reference. It beat the baseline on person detection in four of ten runs and on reconstruction in one. Chain the tasks and it falls apart: errors compound, so Andon Labs puts a typical Astra run's odds of clearing all five steps in sequence at 2.8 percent. The lab's own projection, based on two years of progress, is that a frontier model passes all five in a single attempt by Q1 2027. No lab gets access to the benchmark, and Andon runs every evaluation itself so vendors can't tune to it. Its argument is that lawmakers need this capability curve in front of them before AI-flown drones get good enough that no human is watching.
China's industry ministry wants every critical piece of software running on AI by 2030. MIIT issued an "AI plus Software" action plan, with a 2028 midpoint: noticeably higher intelligence levels across the software and IT services sector, deployment covering 20,000 above-scale software firms, 100 completed AI upgrades inside software companies, 100 benchmark agentic-software applications in key industries, and at least five quality open-source projects. By 2030 the goal is full intelligent upgrades of key software, with AI coding, agentic software and intelligent services becoming new growth engines. The plan names six work areas, from reforming how software is produced to optimising the industry's environment, and it leans on compute vouchers to subsidise firms that buy domestic coding tools and call domestic models, alongside mandatory security review of AI-generated code and pre-launch testing for new software in priority sectors. It also commits to protecting jobs — the ministry frames the target as stabilising existing roles while pushing workers into code review and AI-safety work. Wang Weiwei, vice-director of MIIT's information technology development department, called AI's reshaping of software a historic chance for China's industry to change lanes and overtake. Europe's answer to the same race was supply-side — Draghi tells Europe to build AI data centres — it holds 5% of compute — while Beijing is setting quotas on the demand side.
What to watch: whether Astra's economic-agent results hold up once the runs are longer and the opponents adapt, and whether MIIT's 2028 coverage figure for above-scale firms gets published as a completed count or quietly redefined.
A model that refuses a price cartel is a friendlier business partner. Do you trust a lab's alignment verdict when the same benchmark also measures how much money the model made? Tell us in the comments.
Sources: The Decoder — GPT-6 Astra pilots a surveillance drone and runs a business on its own · Andon Labs — Astra vs Fable on Vending-Bench · Andon Labs — Drone-Bench · Andon Labs — Vending-Bench Arena · Guangming Daily — AI + Software action plan · MIIT — AI + Software Q&A
