Who's Winning the AI Race Right Now

As of this afternoon, Tuesday the twenty-seventh of October, the prediction markets say Anthropic is winning. Fifty-five and a half percent odds of having the best AI model at the end of this year. Google DeepMind is second at thirty-three and a half.

Okay, but odds on what exactly? Who decides "best"?

Polymarket. Fifteen companies listed, exactly one resolves yes, about three point two million dollars traded over the life of the market. It's people betting real money on which lab ends the year on top.

So it's a vibe, but an expensive vibe.

Expensive vibes move fast, too. A week ago Anthropic was at fifty-two and a half and Google was at forty and a half. Google dropped seven points in seven days.

Why? Did something go wrong for them?

Nothing public that I can point to. What I can tell you is that on October twenty-second, Google was actually ahead on one widely watched scoreboard, ninety-six to Anthropic's ninety-four point four. By yesterday Anthropic had retaken it, ninety-seven to eighty-seven.

That's a lot of lead changes for a two-week stretch.

Which is the real story. Back on October fourteenth, The Verge ran a piece called "The AI race is back on," and the reason was two launches in three days.

Go on.

Google shipped Gemini 4 Argon, their most capable model ever deployed, priced to undercut rivals by as much as eighty percent during an introductory period. And Meta shipped Muse Wonder, their first genuinely frontier model.

Meta? The company everyone had written off in AI?

That one. They spent billions last year rebuilding, bought Scale AI, brought in cofounder Alexandr Wang. Meta says Wonder competes with GPT-6 Astra and Claude Opus 5.5 on numerous benchmarks.

Says Meta. On benchmarks Meta chose.

Right, that's their claim. Independently, Muse Wonder sits second on Arena text but six points behind Claude Opus on the Artificial Analysis index at the same price. So: arrived at the frontier, not leading it.

And where's OpenAI in all this? They were the whole story a year ago.

Fourth on that weekly power ranking, with GPT-6.1 Sol. And on the betting market they're at five point nine percent.

Under six? That feels wrong. ChatGPT is still the thing my mum uses.

Usage and frontier-model bragging rights are two different scoreboards, and they've come apart this year. On one weekly board OpenAI's best model ranks twenty-first on Arena, while their revenue is fifty billion dollars.

Okay, so people pay them regardless.

And they are shipping. Yesterday, October twenty-sixth, they launched GPT-6.1 Terra, a computer-use model. Fifty-two outlets covered it. API prices are half of the previous version's.

Half. Is everyone cutting prices at once?

Pretty much, and that's the quiet war underneath the rankings. Argon undercut on price, Terra halved its price, and the open Chinese models keep landing. DeepSeek's new V4.1 Flash came in on the board under an MIT license at thirty cents in, a dollar twenty out.

Which has to be terrifying if you're the one with the trillion-dollar price tag.

Anthropic literally lists it as a risk. Their IPO prospectus came out Monday, and CNBC's read is that competition from lower-cost open-weight models is one of the bear points.

Give me the numbers on that filing. How big are we talking?

Revenue of sixteen point three billion dollars in the first half of this year, up thirteen hundred percent on a year earlier. Valued at nine hundred and sixty-five billion back in May. Reportedly seeking one and a half to two trillion at IPO.

And the losses? There are always losses.

Eighty point seven billion dollars in six months. Though most of that is the change in fair value of convertible notes and derivatives, not cash burned on servers.

"Most of it is accounting" is a sentence I've learned to be nervous about.

Fair. The real cash worry is hundreds of billions in compute commitments, plus over ninety pages of risk factors.

So if I had to answer the title question at a dinner party tonight, what do I say?

That it depends which scoreboard. The money says Anthropic. Arena's text leaderboard says Gemini 4 Argon. One index actually had Meta top as of today, at ninety-three point seven.

Three answers. Brilliant.

And one independent benchmark in early October put Claude Opus at eighty point two against GPT-6 Astra at seventy-nine point nine, which the testers themselves called not a measurable difference.

So what settles it?

Thirty-first of December, when that market resolves. Until then, watch Argon's introductory pricing, which ends on December first, and watch whether Anthropic actually prices this IPO.

Two months of lead changes, then. I'll check back in a week and it'll all be different.

Probably by Thursday.