OpenAI Blitzed. The Money Didn't Move.
OpenAI just had its biggest launch week of 2026. GPT-6 Astra rolled out to the API on September 3, ChatGPT Work's Data Agent shipped on September 10 with connectors for Snowflake, BigQuery, and Redshift, and Sam Altman teased even more for DevDay on September 29. The YouTube feed was wall-to-wall GPT-6 coverage. Eight separate creator videos in one day. "Changes Everything" thumbnails as far as the eye could scroll.
đ Read the full version with charts and embedded sources on ComputeLeap â
And then you look at the money.
On Polymarket â where $51,000 changes hands daily and $1.4 million sits in liquidity on the "best AI model" question â Anthropic is priced at 98% for September. OpenAI is at 1%. Not 10%. Not 5%. One percent. And here is the part that should stop you cold: Anthropic's October odds didn't just hold during launch week. They rose. Up 16.5% in a single week, to 88%. OpenAI sits at 2% for October. Even the year-end 2026 market â the longest horizon, where you'd expect the most uncertainty â gives Anthropic 72% and OpenAI just 10%.
The crowd with real money on the line is betting that the demos don't move the frontier.
âšī¸ The numbers, in one place. Polymarket "best AI model" odds as of September 18, 2026: September â Anthropic 98%, OpenAI 1% (up 9.2% this week). October â Anthropic 88%, OpenAI 2% (up 16.5% this week). Year-end â Anthropic 72%, OpenAI 10% (up 8.5% this week). Total market liquidity: $3.1M across all three horizons.
The Pattern: Three Launches, Three Shrugs
This is not an anomaly. It is the third consecutive time an OpenAI flagship launch failed to dent prediction-market odds.
In July, when GPT-5.6 Sol launched to a Sam Altman victory lap calling it "the best model in the world right now," we tracked the same divergence. As we wrote then: "GPT-5.6 Won the Headlines. The Money Bet on Anthropic." Polymarket had Anthropic at 94% that month.
On September 4, when GPT-6 Astra shipped with benchmark numbers that made Greg Brockman say "welcome to the AGI era," we covered the same split. GPT-6 Astra killed the capability race â meaning the top two models converged on price and saturated the same benchmarks, but the market didn't treat them as equals. Anthropic's odds barely moved.
Now, two weeks later, with the full launch week behind us â Astra, Data Agents, ChatGPT Work integrations, the DevDay drumroll â the market's verdict is even more lopsided. Anthropic's September odds went up 9.2 percentage points. Its October odds went up 16.5 points.
Three launches. Three shrugs. The pattern is not noise anymore. It is signal.
What the Benchmarks Actually Show
The market's skepticism is not irrational. The benchmarks tell a consistent story.
On the Artificial Analysis Intelligence Index, GPT-6 Astra scores 61 points. Claude Fable 5.1 scores 66 â five points ahead. On Arena's text and overall categories, Astra carries no rating at all in the most-watched Arena categories, while Claude Fable 5.1 leads at 1,231 points. Anthropic's Claude Opus 5 Max holds the top spot on the Arena overall leaderboard at 1,505 Elo.
Where Astra does lead â the WebDev Arena at 1,797 points â is a narrower specialist category, and it's the only Polymarket AI market where OpenAI holds the lead (62% for October WebDev). The market is not blind to OpenAI's strengths. It just correctly prices them as niche.
There is a cost story worth noting. Astra matches Fable 5.1's coding agent performance at roughly 60% of the cost per task. That is a real advantage for OpenAI â but prediction markets track "best," not "cheapest." And on the "best" question, Anthropic's Claude 5 family occupies four of the top five Intelligence Index slots as of September 2026.
The Launch Week That Wasn't
What did OpenAI actually ship this week? Let's be precise.
GPT-6 Astra (September 3): OpenAI's largest training run ever â over 100,000 GPUs at the Stargate site in Texas. Priced at $10 per million input tokens, matching Fable 5.1 exactly. Strong on cybersecurity and science benchmarks. Trailed Claude on Arena blind evaluations.
ChatGPT Work Data Agents (September 10): Enterprise analytics via natural-language queries over BigQuery, Snowflake, and Redshift. Impressive demo. Built on GPT-5.6, not Astra â an important detail that most coverage missed.
DevDay teasers (September 15+): Altman promising "big ship this week" and then walking it back to "next week instead."
Each of these is a solid product. None of them changes which model wins blind evaluations. The market understands this distinction. The YouTube algorithm doesn't reward it â the eight GPT-6 creator videos this week are optimizing for engagement, not truth. The Polymarket traders putting up $51,000 in daily volume are optimizing for accuracy.
What the Community Actually Sees
The developer community has been telling this story for months. When OpenAI claimed it had "overtaken Anthropic" with its latest model, the Hacker News thread was skeptical. When Astra actually launched, the discussion focused on pricing and cybersecurity concerns, not on any dethroning of Claude.
Meanwhile, a study on harness design for coding agents â trending on HN this week with 178 points â found that the infrastructure around a model matters as much as the model itself. If the chassis matters more than the engine, then Claude Code's tooling moat compounds on top of Claude's benchmark lead. Karpathy â who co-founded OpenAI â said it plainly:
That is not a casual observation from a random developer. That is the person who literally co-founded OpenAI describing Anthropic's developer experience as a paradigm shift. Prediction markets are downstream of exactly this kind of signal.
The Contrarian Corner: Distribution Beats Models
â ī¸ The bear case for this analysis: Prediction markets track Arena leaderboard position â essentially, "which model wins blind A/B tests." But what if OpenAI isn't trying to win that contest anymore? ChatGPT Work, Data Agents, Operator â these are distribution plays. If AI competition shifts from "best model" to "best platform," then Polymarket is measuring the wrong variable, and Anthropic's 98% is pricing in a race that's already becoming irrelevant.
This is a real argument. OpenAI has 400 million monthly users. Anthropic has a fraction of that. Enterprise distribution through Microsoft, Salesforce, and now native ChatGPT Work integrations gives OpenAI a channel that no benchmark can capture.
But the counter to the counter is that Ramp's AI spending data â enterprise credit-card spend, not surveys â showed Anthropic at 34.4% versus OpenAI at 32.3% back in May, with Anthropic's adoption growing 4x year-over-year while OpenAI sat flat. Distribution advantages only compound if you are also the better product. When the enterprise wallet confirms the same story as the prediction market, the convergence is hard to dismiss.
What This Means for You
If you're a builder: Stop evaluating AI models based on launch announcements. The relevant signal is Arena leaderboard position plus enterprise benchmark data â SWE-Bench, LiveBench, Artificial Analysis Intelligence Index. If a new model launches and the prediction market for "best model" doesn't move, that's information. Use it. Right now, the market says Claude is the tool to build on.
If you're an investor: Prediction market odds are a leading indicator of developer mindshare, and developer mindshare drives enterprise adoption on a 6-12 month lag. Anthropic at 98%/88%/72% across three time horizons, all rising during a competitor's biggest launch week, is as clean a signal as this data source produces.
If you're an enterprise buyer: The AI vendor landscape is bifurcating. OpenAI is building the best platform â integrations, distribution, ChatGPT Work. Anthropic is building the best model. Your choice depends on which bottleneck your org faces: tooling integration or raw capability. If you need the best outputs, follow the money.
The Frontier Has a Truth Serum Now
The biggest story here is not about OpenAI or Anthropic. It is about prediction markets as an instrument.
For years, the AI landscape was navigated by press releases, Twitter hype cycles, and benchmark cherry-picking. There was no mechanism that aggregated informed opinion into a single, money-backed number. Now there is. Polymarket's AI markets carry $3.1 million in combined liquidity across the three time horizons. That is not a poll â it is a price, and it has been right every month this year.
Sam Altman knows this. His pivot from "best model in the world" to "I agree with Dario that we need to pace the frontier" is the tell. When the CEO of the company with the biggest marketing budget starts endorsing his competitor's safety framing, he is not being altruistic. He is adjusting to a world where the scoreboard is public and the score is not close.
đĄ The bottom line for practitioners. Marketing volume and model lead have decoupled â and for the first time, we have a quantitative instrument that makes the gap measurable. Three launches, three shrugs, $3.1M in liquidity confirming the pattern. The frontier has a truth serum now, and it costs $10 to bet.
Previously on ComputeLeap: GPT-5.6 Won the Headlines. The Money Bet on Anthropic. | Anthropic at 92%: Three Surfaces Tell the Same Story | GPT-6 Astra Killed the Capability Race
Originally published at ComputeLeap
