Simon Willison on OpenAI DevDay: The Price War Is Back, and China Is Writing the Playbook

Simon Willison was walking the grounds of OpenAI's DevDay 2026, photographing the pelicans that happened to be flying overhead, when he offered an assessment that cut through the keynote's polished demos. This year, he said, OpenAI is "going for safer, obvious bets," in stark contrast to last year's event, which he described in hindsight as "a bit of a disaster." The launches from DevDay 2025 — agent drag-and-drop and Sora features — no longer exist. They evaporated.

Willison, creator of the Pelican Bench benchmark and one of the more skeptical hands-on practitioners in the AI developer community, was speaking on the Every (AI & I) podcast in a candid live debrief with host Ben Hylak. The conversation unfolded less as a formal interview than as a peer-to-peer postmortem of what OpenAI did — and did not — announce. The headline product was Dots, a personal agent. The headline infrastructure move was an ultra-fast inference tier. But the most important story, Willison argued, was happening in the pricing pages, not on stage.

The price war returns, with a Chinese accent

A week before DevDay, several frontier models quietly got cheaper. GPT-6 dropped 6%. Luna was cut in half. Anthropic's Opus 5.5 fell roughly 20%. This was notable because, for a stretch, neither OpenAI nor Anthropic had been cutting prices at all. The reversal was sudden and broad.

Willison has a theory about why, and it points away from Silicon Valley. "Some of the Chinese AI labs have been putting out new optimization papers," he said. "DeepSeek definitely put out paper describing some of their optimizations and then GPUs get faster like a month later." In other words, the cost curve is no longer being bent primarily by US labs shipping new architectures. It is being bent by open research published in China, which US labs then absorb and operationalize within weeks.

Hylak, who had not previously heard the theory, called it "very interesting." But the data from developer platforms backs the broader trend. Chinese models from DeepSeek, Z.ai, and Alibaba — which Willison noted are "incredibly cost-effective compared to most models from American labs" — have surged from roughly 6-13% of token traffic on OpenRouter and Vercel in February to 55-67% by September. The pricing pressure is not theoretical; it is showing up in actual usage patterns.

Model	Price change	Timing
Luna	50% (half price)	Week before DevDay
Opus 5.5	~20% down	Week before DevDay
GPT-6	6% down	Week before DevDay
Ultra-fast tier	~6x cost for 6-8x speed	Announced at DevDay

The ultra-fast tier is where the pricing war collides with product strategy. OpenAI announced a new inference mode that runs six to eight times faster at roughly six times the cost, bundled into a $500-per-month plan. Willison presumes the hardware is from Cerebras — a company he has directly benchmarked against — and noted he was surprised OpenAI did not name the chipmaker on stage.

Willison's own reaction to the $500 tier was careful. "If I try it, I get hooked on it," he said, given he already pays $200 a month for a tier he described as a toy. Hylak was more blunt. "It's fun to play with, but unless you have 100 million dollars or whatever, wait a couple months for them to make it cheaper." The implication is that ultra-fast is a luxury product today but will not remain one. If the DeepSeek optimization theory holds, the price of that speed is already on a downward trajectory.

Dots: a personal agent wearing a developer's costume

The most anticipated announcement, according to Willison, was some form of personal agent. Meta Muse had been sitting at the top of the app store for roughly two weeks, and OpenAI opened its "personal AI assistant" category back in January. The product arrived as Dots — but Willison found the framing revealing.

"I thought it was interesting that they still pitched it as a developer tool," he said. "Like their big example was look at it doing a complicated database migration to your team."

Hylak, who saw the keynote rehearsals, confirmed the positioning was hotly debated inside OpenAI. The decision to frame Dots as a developer tool rather than a life companion came down to distribution. Dots lives inside ChatGPT, where the user base skews heavily toward people who already do work and build things. Trying to compete with Meta Muse for general consumers would mean fighting for app-store attention from scratch. Inheriting ChatGPT's existing developer-heavy audience is the path of least resistance.

The economics support the decision. "90 plus percent of every company is developers," Willison noted, and developers drive the vast majority of token spend. You do not need to win over the general public to build a large business. You need to win over the people who write code.

Hylak's hands-on experience with Dots is a mixed review. He has been using it for three or four days, and his verdict contains both enthusiasm and frank criticism. "It's not fully featured enough. It's a little bit buggy, but it has actually become the main way that I use ChatGPT." He called it "super early" and less feature-complete than competing tools. Still, the team is moving fast, and in early tests the Dots API — built on Luna, multimodal, and by Hylak's account outperforming a competitor on some benchmarks — looked promising. Whether it was "knocked together in the last week," as Willison wondered aloud, is a question Hylak says he plans to put directly to Sam Altman.

The speed fantasy and the practical reality

The ultra-fast tier is built on a simple promise: cut latency to the point where interacting with a screen feels like speaking to a person. Willison has direct experience with Cerebras, the presumed hardware vendor, and he described the compelling demo: voice-driven iteration on a live canvas. You say, "show me a chart like this, now add this to the house, now add more wheels to the bicycle," and the screen updates in near real time.

His own reaction to those demos was measured. "They were fun, but it didn't feel like something I'd be using on a daily, day-to-day basis." He acknowledges other developers swear by fast mode for coding, and Hylak says he always uses fast mode but has not adopted ultra-fast. The gap between an impressive demo and a daily habit is where many AI products go to die.

Willison also pointed to a cautionary precedent. Google shipped a Gemini diffusion model roughly a year and a half before this DevDay that was "crazy fast" — and then did nothing with it. He called the decision puzzling. Speed alone is not a product. It is a feature that needs a use case.

The distribution problem that will not go away

If there is one thread that runs through Willison's analysis, it is distribution. OpenAI launched a marketplace at DevDay — a place where developers can publish and users can find AI tools. Willison's reaction was not excitement. "They keep trying that," he said. When Hylak asked what happened to the previous attempts, Willison's answer was essentially that they evaporated.

This is OpenAI's unresolved problem. The company builds models so capable they can perform database migrations, generate images, and write code. But getting those capabilities into the hands of users at scale has repeatedly proven harder than building the models themselves. Altman himself nodded to the problem on stage, Willison noted, acknowledging that OpenAI has said it will try to solve distribution. Acknowledging the problem is not the same as solving it.

The marketplace is at least the third attempt. Last year's agent drag-and-drop was another. The question Willison leaves hanging is whether this time is different, or whether the marketplace becomes the next discontinued offering, remembered only as a DevDay slide.

What the safer bets tell us

Willison's overarching read on DevDay 2026 is that OpenAI traded ambition for reliability. Last year's event was, in his view, "a bit of a disaster" in hindsight — launches that did not survive contact with the market. This year, the company made choices that are harder to get wrong: a faster tier that everyone wants at some price, and a personal agent that "feels like it's going to work."

The pelican interlude, which gives Willison's benchmark its name, offered a small but telling data point. He rendered a pelican with the new Sol 6.1 model and critiqued the output: a chunky neck and a "garbage bicycle frame." Checking his comparison grid, he found that the earlier 5.6 Sol had produced the correct bicycle shape at all thinking levels until low, where it degraded — suggesting the earlier success may have been a lucky roll rather than a genuine capability difference. The benchmark is, in Willison's own words, "the stupidest of all of the benchmarks," and yet it keeps surfacing errors that more sophisticated evaluations miss.

That is the through-line of the entire event. The safe bets are safe because they are measurable: speed, price, API availability. The ambitious bets are risky because they depend on distribution, habit formation, and the messy question of whether a tool becomes something people reach for daily. OpenAI has proven it can make models faster and cheaper. It has not yet proven it can make a marketplace that matters.

For the industry, the pricing reversal is the signal to watch. If the prior week's cuts across GPT-6, Luna, and Opus 5.5 are the start of a trend rather than a one-off — and if the optimization pipeline from Chinese labs continues to feed US cost curves — the $500 ultra-fast tier's economics will erode within months, just as Hylak predicts. Developers who wait will get the speed at a fraction of the price. Investors should read the price war less as a margin threat to any single company and more as evidence that the competitive moat in AI is narrowing from a technology lead to a distribution lead. And distribution, as Willison keeps reminding, is precisely the problem OpenAI has not solved.

The coding landscape beyond the keynote is similarly unsettled. Willison names Codex as his primary tool, and mentions picking up another tool again after a 5.5 release he called "so sick." He relayed a blunt complaint about a competing tool he was forced to use while building his live-blogging system: "Claude sucks so badly." The two also sparred lightly over whether OpenAI actually delivered its stated 20 launches at DevDay — Hylak counted around 20 depending on how items are grouped, while Willison suspected it fell short and called the 20-launch framing simply a bad idea.

For the reader tracking this space, the practical takeaways are clear. The era of stable AI pricing is over, and it ended not because any US lab decided to be generous but because open optimization research from China made higher-speed, lower-cost inference possible. The products that matter now are not the flashiest demos but the ones that solve distribution. And the personal agent category, for all Dots' rough edges, finally feels like it is going to work — even if the question of whether anyone can reliably get these tools into users' hands remains, in Willison's words, "the biggest question."