A capability jump, not another round of price cuts, is what Wall Street had been waiting for — and OpenAI delivered it on Sept. 4 with GPT-6 Astra. The model's release sent SoftBank Group up more than 10 percent and Oracle about 5.5 percent higher, as Goldman Sachs' Delta One trading desk declared the launch a genuine intelligence breakthrough that resets the demand curve for AI compute and keeps the industry's spending cycle alive.
"Price declines are one thing; a genuine breakthrough in intelligence will reshape the demand curve again," Rich Privorotsky, head of Goldman Sachs' Delta One trading desk, wrote in a market note. "It forces every other lab to catch up, keeps the spending cycle alive, and revives the belief that there is something better worth building."
Astra scored 99.9 percent on ARC-AGI-3, a test of whether AI can solve unfamiliar problems, against an average human tester's 48 percent, and 100 percent on ExploitBench, which measures whether models can turn software vulnerabilities into working exploits. OpenAI says the model can operate browsers, spreadsheets and desktop software directly, and that it completes software-engineering tasks at roughly 57 percent lower cost per finished task than its predecessor. The API runs $10 per million input tokens and $50 per million output tokens, a premium that President Greg Brockman defended by arguing token pricing "doesn't make any sense" now that value is measured per completed task rather than per token.
The market reaction maps OpenAI's valuation directly onto public equities. SoftBank, the startup's largest outside investor, rose more than 10 percent overnight, while Oracle, whose cloud infrastructure stands to absorb much of the new demand, gained about 5.5 percent. Privorotsky called the share moves "the least interesting part of the story," arguing the more consequential signal is that compute oversupply only becomes a problem if model capability stalls — and Astra points the other way. OpenAI also claims to have surpassed Anthropic, maker of the Claude Opus 5 model, on some AGI-related benchmarks, intensifying a rivalry that had recently tilted toward its competitor on price.
September's catalyst calendar
The bullish read faces a crowded test window. Oracle reports earnings next week, where investors will parse its cloud-infrastructure and AI backlog, and the September technology conference season offers more venues for the Astra effect to spread through the broader AI supply chain. Privorotsky flagged energy price trends and next week's consumer-price-index reading as the main potential disturbances to current sentiment.
Skeptics note the gaps in OpenAI's case. The company did not disclose GDPval, its own benchmark for economically valuable work, in Astra's launch materials — an omission that undercuts Brockman's AGI framing, since OpenAI defines the concept as outperforming humans at most economically valuable work. Independent researchers also flagged that Astra shows a "substantial decrease in chain-of-thought monitorability" versus prior models, complicating oversight of an agent that can act inside live software. And benchmark scores can overstate the leap: Nvidia recently reached 100 percent on ARC-AGI-3 by layering Claude Opus 5 with external memory and tooling, even though the underlying model's baseline sat near 30 percent.
For investors, the question is whether Astra's capability jump justifies the spending it implies. Oracle and SoftBank have rallied on the expectation that frontier intelligence will keep data-center demand growing faster than supply, and OpenAI's own economics — a 57 percent cost-per-task improvement despite premium token pricing — suggest the model can monetize its lead. The September earnings and conference calendar will show whether that conviction holds, with CPI and energy prices the swing factors that could break the spell.