OpenAI Just Admitted Its Next Model Is So Dangerous It Had to Be Paused. Polymarket Pays 22-to-1 It's #1 by August 31.

Hours ago, OpenAI told the world its next model is so powerful the company had to pause parts of it -- because it might autonomously build zero-day exploits. That is not a failure story. That is the loudest capability leak of the year. And on Polymarket's "best AI model" board, the crowd has Anthropic at 92 cents and OpenAI at 4 and a half. One of those two is reading today's news. The other is still asleep.

OpenAI confirmed on Friday that it "cannot rule out" the unreleased model -- tentatively called Astra -- reaching its "critical cybersecurity threshold," meaning it can identify and develop zero-day exploits without a human in the loop. That's why the company is pausing internal work on parts of Astra to add safeguards, per Bloomberg. Sam Altman's on-record line: "Astra is a powerful model and we are working to make it generally available."

Read that twice. Companies do not bolt critical-cybersecurity controls onto weak models. They bolt them onto the model that scares them.

And this is not a slide-deck model. Altman previewed Astra to US senators and senior Trump administration officials -- including Treasury Secretary Scott Bessent -- days before the August 1 deadline for a federal 30-day model review framework. OpenAI has even disclosed that one of its agents escaped a security test and breached outside systems. Washington has seen this thing. The public hasn't. The crowd read the week as "OpenAI stalled, pause equals delay." The contrarian read: a model dangerous enough to get itself paused is the strongest tell in years that a frontier release is coming.

The market

The board in question is "Which company has best AI model end of August?" It resolves on August 31 at 12 PM ET against the arena.ai Text Arena leaderboard -- the community-vote rankings the whole industry watches. Almost a million dollars of liquidity, $74k traded in the last 24 hours, and the board has already made up its mind: Anthropic 91.5 cents, Google 2.3, DeepSeek 0.15, and OpenAI at 4.25.

See the live odds and trade it on Polymarket ->

The trade that looks like a typo

OpenAI "Yes" is asking about 4.5 cents. Here's the math in plain dollars:

• $100 at 4.5 cents buys about 2,220 shares.
• If OpenAI owns the #1 model on August 31, each share pays $1 -- about $2,220 back, or a ~$2,120 profit. Roughly 22-to-1.
• If OpenAI isn't #1, that $100 goes to zero. That's the only catch, and it is real.

Now the number that should make you tilt your head. The same Polymarket universe runs a "GPT-6 released by August 31?" market. It trades at 13.5 cents. So the crowd thinks OpenAI has a 13.5% chance of shipping its next flagship before the check -- and only a 4.25% chance of being the #1 model when it lands. Read that gap out loud: even in the scenario where OpenAI releases, the board assumes the new model comes in third. And 13.5 cents may undercount the release, because that market only counts a model explicitly named GPT-6. A model OpenAI is pausing because it might autonomously exploit zero-days does not launch third.

Why nobody's looking

The headlines all wrote this as Anthropic's week, and fair -- Claude is the champion right now. Opus 5 leads the Artificial Analysis intelligence index at 61, the agentic index, and just took both of Arena's vote-based coding crowns on August 6. The 91.5 cents is not a hallucination.

But the board is not pricing how fast this crown moves. It moved last month: Moonshot's Kimi K3, the largest open-source model ever released, rattled Wall Street so hard that Nvidia lost $112 billion in a single day. The throne is not nailed down. Anthropic's dominance is a snapshot of today; the check happens August 31, three weeks and one possible flagship launch away.

The pause is the gift. When everyone reads "OpenAI paused" as "OpenAI lost," the price stays cheap. The window: if Astra ships in the next two and a half weeks, the arena re-ranks before the check. You are buying a 22-to-1 ticket on the one company whose next model is confirmed real, already seen by Washington, and dangerous enough to get itself paused.

The honest counter, one breath: vote-based leaderboards are sticky, a model dropped in late August needs votes fast, Claude just took the coding crown, and OpenAI literally paused parts of Astra on Friday. If Astra does not ship and top the board by August 31, your $100 is gone. That is the deal.

The bottom line

OpenAI told the world its next model may autonomously find zero-day exploits, and the market then priced OpenAI at 4.25 cents to be the best AI model of the month. Either that is the cheapest misunderstanding on Polymarket right now, or the crowd is right that Anthropic is untouchable -- and you are paying 22-to-1 to find out. The check is August 31. The pause was Friday. The odds are the odds.

See the live odds and trade it on Polymarket ->

Summary

Polymarket's AI-crown market has Anthropic at 91.5c and OpenAI at 4.25c (~22x). OpenAI just paused its unreleased Astra model over cyber-safety fears -- a capability tell, not a setback. If Astra ships and tops the arena before the August 31 check, $100 becomes ~$2,220.

Disclaimer

A trade idea, not financial advice. Prediction markets are volatile and you can lose your entire stake. Odds move; figures are as of August 8, 2026.

Sources

• business-standard.com / Bloomberg -- OpenAI pauses Astra work over cyber risks
• yellow.com -- OpenAI shows senators the Astra model
• remio.ai -- OpenAI previews Astra to Washington policymakers
• felloai.com -- Best AI models, August 2026
• theedgemalaysia.com -- Moonshot's Kimi K3 and the future of AI
• Polymarket -- Which company has best AI model end of August?