New Intelligence Report
Is front-end development really about to be wiped out in one move by OpenAI?
Today, developer @Lentils80 discovered by chance that OpenAI has expanded internal testing of Astra, under the codename mozaik-alpha-fdm.
The two latest outputs released were all generated zero-shot in Max effort mode.
One is a 3D pixel-art castle, the other a front-end web page — both produced in a single pass.
At the same time, a wave of developers granted early gray-release access posted the first batch of Astra tests simultaneously.
What truly astonished developers was its ability to control a massive number of details in one shot: a unified visual language, spatial relationships among complex elements, page hierarchy, and the code structure needed for interactivity.
After seeing it, someone flatly declared: "Frontend is solved!"
The consensus prediction is that around next Thursday (September 3), Astra will officially debut and go head-to-head with Fable 5.1.
To sum up, Astra will boast the following powerful capabilities:
Exceptionally strong front-end performance, with zero-shot, one-shot generation;
End-to-end trained multi-agent orchestration, easily handling ultra-long-horizon tasks;
A "persistent" reasoning mode, with instant self-correction.
One Sentence Generates a 3D Web Page — Astra's Front-End "Ascends to Divinity"
As everyone knows, front-end work has long been a weak spot for large models.
No matter how slick the back-end logic, the moment it comes to generating UI, the flaws show — bloated CSS, eye-searing color schemes...
This time, however, the level of detail control Astra demonstrated under zero-shot conditions genuinely hints at a qualitative leap.
This interactive 3D isometric kingdom map required the AI to simultaneously handle geometric perspective, layer stacking, dynamic interaction, and visual consistency.
With maximum thinking effort enabled, Astra could nail it in a single pass.
In Max mode, the Astra model tends to think much more than Sol, but as the output below shows, its attention to detail is extraordinarily high.
The spaceship it generated is remarkably realistic in every detail and structural element.
Another developer, MeSun, used Astra to build a 3D bicycle test page (HTML) with physics effects — even the bike's body proportions could be adjusted at will.
Then they asked it to generate a 3D first-person roguelike game focused on satisfying combat.
The result: once again, it "passed in one take," nailing it on the first try.
After watching this wave of stunning Astra demos, one prominent leaker even proclaimed that, for the first time in history, "ideas" are beginning to crush "execution."
"Whatever you can imagine, AI can build for you."
The Permanent Agent: Four Core Technologies Laid Bare
Why did Astra's front-end ability transform seemingly overnight?
Combining the leaked characteristics with industry predictions, the core technical breakthroughs of the mozaik-alpha-fdm version likely center on the following four dimensions.
First, end-to-end multi-agent orchestration.
Astra may no longer be a mere "monolithic code generator," but instead natively integrates a multi-agent architecture at its foundation.
For example, letting Codex handle programming tasks, using another sub-model to process natural-language queries, and having the main model integrate the results.
When generating a 3D web page, the model internally runs "a UI designer, a front-end engineer, and a code reviewer" concurrently, thereby guaranteeing the trinity of layout, styling, and logic.
This is in line with the Multi-agent feature rolled out in beta in GPT-5.6, but Astra may have internalized it as a natively "end-to-end trained" capability.
Second, ultra-long-horizon task capability.
Previously, when writing projects involving multiple components and complex CSS/JS interactions, models were highly prone to "forgetting" and "hallucination."
The completeness Astra displays proves that the attention mechanism within its context window has been revolutionarily optimized, enough to support building complex engineering projects from scratch.
Judging by Astra's performance solving ten math problems, it possesses the ability to maintain reasoning consistency over extremely long time spans.
This is hugely significant for continuous code-project development, research tasks at the level of academic papers, and autonomous execution of complex workflows.
Third, a persistent reasoning mode.
Astra's thinking volume under Max effort significantly exceeds Sol's, pointing to a more "persistent" reasoning mechanism.
Rather than outputting an answer in one go, the model repeatedly verifies, corrects, and iterates internally until the output reaches a satisfactory level.
This means it can "remember" the current DOM tree structure and maintain absolute consistency in design language and code conventions across subsequent generations.
Fourth, instant self-correction.
This is the most frightening part: Astra may complete a "virtual render" or a logical pass of the AST (abstract syntax tree) in an internal sandbox while generating code.
It automatically detects front-end misalignments, completes the refactoring, and only then presents the flawless zero-shot result to the user.
Reproducing a PS5 controller — a highly detailed rendering task — zero-shot is itself indirect evidence of self-correction capability.
First Week of September: A Head-On Clash With Fable 5.1
But the real show is still to come.
Just this week, Anthropic was also reported to be accelerating Fable 5.1.
Even more coincidentally, its release window lands right around the same time as Astra's.
Judging from feedback shared by gray-release testers, Fable 5.1's upgrade priorities collide almost perfectly with Astra's:
More stable front-end code, with complex component logic formed in one pass;
Stronger long-chain-of-thought reasoning, less prone to "forgetting the premise" midway;
Further enhanced tool calling and step-by-step planning, edging ever closer to a mature AI agent.
In other words, the two companies have charged onto the same battlefield: stronger coding ability, longer-running tasks, and minimal human hand-holding along the way.
Now, the two sides have even begun sparring remotely.
Generating the same SVG vector image of a "woman's self-portrait": Astra on the left, Fable 5.1 on the right.
Whose details are steadier, whose structure is cleaner, and whose single-pass completeness is higher can now be compared directly, out in the open.
But what ultimately decides the winner may not be model capability alone.
According to leaks, beyond raw performance, Astra may also push prices down further on cost-effectiveness.
If OpenAI can truly deliver — at lower API costs — a model with built-in Max effort, stronger self-correction, and no need for complex engineering orchestration, then Fable 5.1 will be facing more than just benchmark pressure.
Next Thursday: the battle of the titans. Who will claim the new throne? Let's wait and see.
References:
https://x.com/pankajkumar_dev/status/2093623079192891882
https://x.com/lentils80/status/2093617080327127456
Editor: Taozi