A writer for Entrepreneur says the same AI model can score 62.7% or 99.9% on one test, depending on the setup around it. That matters before you buy any AI tool.
The same AI model can look average or nearly perfect depending on the software wrapped around it. That is the main point of an opinion piece in Entrepreneur by an AI engineer. On one test, OpenAI's new GPT-6 Astra scored 62.7% with a standard setup and 99.9% with another. The model did not change. Only the setup did.
OpenAI released GPT-6 Astra on Sept. 3. Nvidia's Jensen Huang posted on X that "AGI has arrived," and said "400K GPUs" are "coming online next." AGI means artificial general intelligence, a system that can match or beat humans at nearly any thinking task.
The reported scores are strong: 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench, up from 78.5% for the earlier GPT-5.6 Sol. Still, OpenAI itself did not call Astra AGI. It stressed speed and versatility. Wharton professor Ethan Mollick calls today's systems "jagged AGI," meaning superhuman in some areas and unreliable in others.
The piece points to something called a harness. It is the software around a model that builds the prompts, gives it tools, manages its memory and helps it recover from errors. One description it quotes: "A model is the engine. The harness is the car. Most benchmarks test the car."
The gaps can be large. Research from Cursor found one model scoring 46% with one harness and 80% with another. On ARC-AGI-3, Astra's jump from 62.7% to 99.9% came from a harness that let it reuse earlier reasoning instead of starting cold at each step. That is a 37 point swing from setup alone, according to the sources the author cites.
If setup drives results, a leaderboard may say less about the model than it seems to. The author argues that some of what gets called model improvement is really harness improvement.
Real use shows the same split. Astra and Anthropic's Claude Fable 5.1 look close on several coding and reasoning tests. Engineers report different output. Fable tends to write more production-ready code, while Astra shines at computer use and multi-step automation. Snorkel AI found Fable tends to lose valid work midway through a task, rather than running out of budget.
The author also reads SpaceX's $60 billion purchase of Cursor this way. Cursor is a harness company, not a model maker. OpenAI has cut Cursor's access to its models over trust concerns tied to Musk's companies.
The author's advice for anyone weighing an AI tool for a business is to set the AGI talk aside and check three things:
Before paying for an AI tool, test it on your real tasks and ask the vendor how it fails, not just how high it scores.
Nvidia's board is investigating CEO Jensen Huang after reports tied him to an alleged Chinese spy, a story that matters to anyone who relies on AI.
Read more: Nvidia's board will investigate CEO Jensen Huang over alleged spy linkMicrosoft reportedly cut some AI usage limits by 90 percent, and Meta's use of Claude Code reportedly halved. Here is what that could mean for your tools.
Read more: Microsoft cuts some AI usage limits 90%; Meta coding tool use halvesApple's new M6 Mac mini costs $900 and handles everyday creative work well, but Engadget says NVIDIA graphics cards are far stronger for gaming.
Read more: Apple's new M6 chip trails NVIDIA graphics cards but costs far less
