Mythos leaks, SpaceX buys Cursor and OpenAI drops GPT Image 2.0
TL;DR
OpenAI’s GPT Image 2 looks like a step-change, not a minor upgrade — Wes says it jumps from roughly 1271 to 1512 ELO on Arena-style rankings, beating Google’s “Nano Banana 2” across UI generation, text rendering, portraits, anime, 3D, and more.
Wes thinks GPT Image 2 is only half the story — he predicts a second OpenAI model could land within 48 hours, pairing image generation with a rumored front-end coding model that can turn screenshots or designs into working websites.
The strongest practical use case may be design-to-code workflows — early testers of OpenAI’s stealth model reportedly say its front-end development ability is “off the charts,” and Wes shows why GPT Image 2’s UI mockups could become inputs for near-perfect website generation.
SpaceX and Cursor appear to be doing a huge AI coding deal — according to Wes, SpaceX secured an option to acquire Cursor for $60 billion later this year, or else paid $10 billion just for the partnership, giving Cursor access to SpaceX’s million-H100-equivalent Colossus cluster.
The Mythos leak story is messy and alarming in a very internet way — Wes says the supposedly unreleased “super dangerous” model had apparently been circulating inside a private Discord-style forum where users were casually running it long before any official public release.
GPT Image 2 is stunning, but its kryptonite may still be weirdly human prompts — after generating blueprints, menus, handwritten notes, plaques, and an almost-perfect periodic table, the model still gets tripped up by a deceptively simple request: a wine glass filled fully to the brim.
The Breakdown
A chaotic AI news day: Mythos, Cursor, and OpenAI
Wes opens in full rapid-fire mode: OpenAI has dropped GPT Image 2, Mythos apparently leaked into a private online community, and SpaceX is reportedly tied up in a massive Cursor deal. His framing is clear from the start — today’s image model release feels like “part one,” with another OpenAI shoe possibly dropping by Thursday.
The Cursor-SpaceX rumor is enormous if true
According to Wes, this isn’t a clean acquisition yet: SpaceX has an option to buy Cursor for $60 billion later this year, and if it walks away, it still paid $10 billion for the partnership alone. The real strategic hook is that Cursor would get access to SpaceX’s Colossus training cluster — described here as a million-H100-equivalent supercomputer — which could supercharge its work on coding agents and its Composer product.
GPT Image 2 blows past the field on benchmarks and first impressions
Wes calls it the biggest image-model gap he’s seen so far, citing an ELO leap to 1512 versus Nano Banana at 1271. He leans on Arena commentary too: the new model’s outputs, especially for UI and design, have that rare effect where you stare at them trying to find what’s wrong and come up empty.
Why Wes thinks this matters for front-end coding, not just pretty pictures
The big theory is that OpenAI is pairing this image model with a stealth coding model that testers say is absurdly good at recreating websites from visuals. Wes speculates the workflow could be: generate a polished interface with GPT Image 2, then hand it to the next model to turn it into a fully functional site in real time.
The demos get weirdly practical: blueprints, SVG code, panoramas, and upscaling
He tests the model with a “high-tech automated chicken coop” blueprint, and the result lands because the dimensions, labels, and layout all look coherent instead of gibberish. One especially nerdy example stands out: the model generates an image of code for an SVG pelican, someone OCRs the text from that image, runs it, and it actually produces the pelican.
The one thing it still can’t quite do: the cursed full wine glass
For all the gains, Wes gets hilariously hung up on a prompt asking for a glass of wine filled to the brim. The model keeps technically satisfying the wording while dodging the real intent, which turns into a live semantic debate about “glass of wine” versus “wine glass” and whether the output looked more like a goblet or even a Guinness glass.
Text rendering is now good enough to fake real documents and menus
This is where Wes seems genuinely impressed: plaques with long text, an Eastern European restaurant menu, handwritten notes, and even an almost-perfect periodic table all come out legible and structured. His reaction is basically that you could print some of these outputs directly — the menu in particular looks like something a real restaurant could use.
OpenAI won the image round, and Wes thinks another release is imminent
By the end, he says OpenAI is “firmly back on top” in image generation, especially because no competitor seems close right now. But he keeps circling back to the same bigger point: don’t judge this launch in isolation, because GPT Image 2 may be the visual half of a much larger front-end generation stack arriving within days.