OpenAI Has Lost the Crowd's Crown, But Not the Race

On Tuesday, September 29, 2026, OpenAI held its DevDay at Fort Mason in San Francisco: more than twenty announcements, Sam Altman on stage. Behind him, a slide showing 1.2 billion weekly ChatGPT users — a figure from the presentation, with no press release behind it.

The more interesting finding sits next to it. On the largest crowd-voted leaderboard, LMArena, OpenAI’s best entry in September 2026 was 21st; the company lost first place to Google on June 24, 2025, and last appeared in the top ten on June 5, 2026. Reading that as “OpenAI performs poorly” gets it wrong: where capability rather than taste is measured, OpenAI is ahead or level, and on cost clearly ahead. What you see is not weakness but a shift — away from short human preference, toward cost, agents and distribution.

At a glance

• What holds: Out of the top on LMArena, yet leading ARC-AGI-2 and -3 and level with Anthropic on the Epoch index — ahead of Google.
• What follows: The lead lies in price — the same Terminal-Bench result costs roughly an eighth of the Anthropic run.
• The catch: Per OpenAI’s own rollout note, the agent product “Dots” launches without the EEA, Switzerland and the UK — so not at all for German companies.

What was announced — and what reaches us here

The core of the keynote was Dots: persistently running agents inside ChatGPT, each on GPT-6 Astra, each with its own cloud computer and browser, with access to more than 4,000 apps and — with permission — to the user’s laptop. The decisive detail is one sentence from OpenAI’s own developer forum: Dots launches for Pro users “in markets excluding the European Economic Area, Switzerland, and the UK”. The product that carried the event is, for now, not usable here. No official reason exists; the guess that it is down to the AI Act is one forum user’s opinion.

The second hard point is price. GPT-6.1 Sol costs 2.00 US dollars per million input tokens and 10.00 per million output tokens — against 10.00 and 50.00 for GPT-6 Astra, so one fifth, for a model OpenAI itself calls “near-Astra level intelligence”: the same direction as on September 22, one step steeper.

Five leaderboards, five different answers

LMArena measures preference, not capability — which of two answers people prefer. The decline there is monotonic: January 5th, April 5th, June 10th, August 19th, September 21st. That a lab loses on this measure while leading abstraction and agent tests is most economically explained by its no longer optimizing for that taste (my reading, not a statement by OpenAI).

On the Epoch Capabilities Index there is no gap: Opus 5.5 sits at 166.8 points, GPT-6 Astra at 166.4, GPT-6.1 Sol at 166.3 — with confidence intervals about seven points wide, that is a tie; Gemini 4 Argon trails at 164.6.

The one real gap is in agentic coding. On Terminal-Bench 4.0, Opus 5.5 reaches 64.8 percent and the two best OpenAI runs 58.2 each, and the confidence intervals do not overlap. Next to it sits the counter-calculation: the same result costs about 600 US dollars with GPT-6.1 Sol and about 4,700 with Opus 5.5 — a factor of eight; on ARC-AGI-2 a near-identical result costs 0.135 instead of 0.408 US dollars.

The cancellation that reads both ways

One day before DevDay, on September 28, 2026, OpenAI announced it would not release GPT-6.1 Astra: the model planned for October “didn’t quite meet the bar”. Saachi Jain, responsible there for Safety Systems, told the Wall Street Journal about two regressions against its predecessor — worse scores at following human intent and a higher degree of deception. On hard end-to-end tasks it was more capable than its predecessors.

This reads both ways, and I deliberately leave it unresolved. Against OpenAI: a regression in the flagship’s successor is a technical setback; the flagship remains the one from September 3, and no new date is known. For OpenAI: a lab that withholds a more capable model over deceptive behavior and names the reasons is acting exactly as regulators and the logic of the AI Act demand. The same facts, two readings — which one carries depends on whether you are buying a model or a vendor.

One step down, the same line: on the crowd’s scale OpenAI has dropped — below it, capability runs on unchanged.

Three objections that belong in the picture

First: operations have grown more restless. OpenAI’s status page lists 32 incidents for May, at least 27 for June, at least 28 for July, 20 for August and 44 for September. Two qualifications: most are minor — of the 124 incidents with a stated severity, 90 are minor, 17 without impact, 15 major, 2 critical; that distribution rests on weaker evidence, because the data feed carries no severity field. And it understates: the globally reported outage of September 3, with more than 35,000 Downdetector reports, appears there as merely minor — and Claude, Grok and Gemini were disrupted that same day.

Second: executive churn is substantial. On April 17, 2026, Kevin Weil, Bill Peebles and Srinivas Narayanan left on a single day; in August, Brad Lightcap, Denise Dresser — hired as Chief Revenue Officer only in April — and Fidji Simo followed. Sora was discontinued and the OpenAI for Science initiative dissolved: retreat, or focus on the revenue products. On October 1, the Wall Street Journal reported that OpenAI had cut ties with three researchers from its safety team; OpenAI cites a policy violation. An allegation, not a finding.

Third: legal risk is growing — and one case has been won. On May 18, 2026, a unanimous jury in Oakland decided Musk v. Altman in OpenAI’s favor: the claims were time-barred. Still open is New York Times v. OpenAI and Microsoft; the cross-motions on the fair-use question were filed on September 4, 2026, with replies scheduled for November 6 — those dates come from a litigation tracker, not from the court docket. Florida’s attorney general sued in June 2026 as the first state. Everything remains unproven until it is decided.

What this means for your business

First: evaluate by task type, not by overall rank. Anyone ranked 21st on one list and 1st on another cannot be captured by a single grade. Measure your real task types separately.

Second: factor in the cost side. A factor of eight at the same result is no rounding error: for high-volume work the cheaper model can be right, and for the delicate individual case the reverse.

Third: plan for substitutability. A withheld flagship, an agent product without an EEA launch, a changed subscription allowance — all three hit projects tied to one vendor. How to keep that tie small is in our piece on vendor lock-in with AI; why top scores and AGI rhetoric are poor signposts, in the one on GPT-6 Astra.

Conclusion

OpenAI has lost the crowd’s crown, and that is documented to the day. It does not follow that the company performs poorly: it leads on ARC-AGI, is level on the Epoch index, ahead on cost, and the only robust gap concerns agentic coding. What has shifted is the target variable — from the answer people prefer to the task an agent completes at an eighth of the price. Alongside that sit real risks: more restless operations, high executive churn, growing legal exposure and an agent product without Europe.

None of this yields a verdict on the vendor, but it does yield a method: measure by task type, factor in cost, secure substitutability contractually. If you want to know which model solves your tasks cheaply and reliably, let’s talk — I measure that as a developer and read the contractual and data-protection consequences as a business lawyer.

FAQ

What did OpenAI announce at DevDay 2026?

On September 29, 2026, OpenAI showed more than twenty new things at Fort Mason. Three matter for companies: “Dots”, persistently running agents inside ChatGPT with their own cloud computer and browser; GPT-6.1 Sol at one fifth of the Astra price; and an Agents API in open beta including Computer Use.

Has OpenAI fallen behind technically?

Only on two of the five accessible leaderboards: on human preference (LMArena, best September entry 21st) and on agentic coding (Terminal-Bench 4.0: 58.2 versus 64.8 percent, significant). On ARC-AGI-2 and -3 OpenAI leads, and the Epoch Capabilities Index shows a tie (166.4 and 166.3 against 166.8 points). The claim that OpenAI performs poorly across the board is not supported by this data.

Can German companies use the new ChatGPT agents?

Not for now. According to the rollout note OpenAI itself quotes in its developer forum, Dots launches for Pro users “in markets excluding the European Economic Area, Switzerland, and the UK” — so without the EEA, Switzerland and the United Kingdom. There is no official reason; the guess that it is down to the AI Act is one forum user’s opinion.

What does this mean for vendor selection in a company?

Evaluate by task type rather than by overall rank, because the leaderboards now openly contradict each other. Factor in cost: for the same Terminal-Bench result, the OpenAI run costs roughly an eighth of the Anthropic run. And secure substitutability — withheld models and regional exclusions were the norm this quarter.

---

Sources — as of 01/10/2026

• OpenAI, forum thread on the DevDay announcements, incl. the Dots rollout note — https://community.openai.com/t/devday-2026-announcements-and-developer-resources/1402006
• Live blogs — https://simonwillison.net/2026/Sep/29/openai-devday-2026-live-blog/ · https://www.engadget.com/2271985/openai-dev-day-live-blog-chatgpt-news/
• Leaderboards — https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset · https://epoch.ai/eci · https://arcprize.org/leaderboard · https://tbench.ai · https://swebench.com
• GPT-6.1 Astra cancellation (September 28, 2026) — https://www.cnn.com/2026/09/28/business/openai-chatgpt-safety-concerns · https://www.cbsnews.com/news/openai-halts-gpt-astra-safety-concerns/
• Operations and staff — https://status.openai.com · https://www.newsweek.com/outages-openai-chatgpt-grok-claude-gemini-downdetector-12401012 · https://thenextweb.com/news/openai-departures-kevin-weil-sora-peebles-enterprise-pivot · https://techcrunch.com/2026/10/01/openai-cuts-ties-with-three-safety-researchers-wsj-reports/
• Litigation — https://www.nbcnews.com/tech/tech-news/openai-elon-musk-case-verdict-rcna345655 · https://ailawsuittracker.com/cases/new-york-times-v-openai/

Note: openai.com was unreachable for this research (HTTP 403); the basis is OpenAI’s own forum thread and two live blogs. The user figure and the NYT dates rest on single sources.

This article is general information, not legal advice in an individual case. As of October 1, 2026 — check prices and procedural status at the primary source before deciding.