Large Language Models

GPT-6.2 Sol Looks Better Than Astra’s Troubled Debut

GPT-6.2 arrived with a credibility problem. OpenAI released GPT‑6.2 Astra and GPT‑6.2 Sol on Tuesday, presenting them as models built for different kinds of work. Early testers, however, found Astra difficult to control, inconsistent on long projects, and oddly uninspiring for a flagship release.

Matt Shumer, investor and former HyperWrite AI CEO, said he would not publish the longer review he customarily wrote for earlier OpenAI flagships. In a post on X, he described Astra as “essentially unusable,” adding that he assigned it work as a sub-agent before Claude models supervising it redid the results.

Developer and YouTuber Theo Browne was even harsher, calling 6.2 Astra “the worst model OpenAI has ever released.” Browne and other AI power users focused their criticism on delegated work and creative tasks — the areas where a model must keep moving without constant supervision.

Astra’s ambition meets uneven execution

OpenAI positions Astra for demanding assignments such as changing a complex codebase, building a working web application, or producing a document that follows company templates. The company says the updated frontier model creates slides, spreadsheets, and documents that “read less like AI,” while following templates more closely; it also claims stronger coding for websites, gaming, and 3D design.

Testing produced a less tidy picture. Dan Shipper, CEO of generative AI publication Every, said his team spent several weeks testing both models and found Astra performed about as well as its predecessor on most coding benchmarks. Astra often stopped early on end-to-end builds, and its push for brevity could make writing hard to follow or remove necessary explanation.

Shipper praised Astra’s default design instincts but said its outputs were difficult to vary, with a “noticeable house style.” That is a polite way to describe a model that can produce competent work while making every project feel as if it passed through the same narrow creative funnel.

Browne’s own experiments exposed more specific problems. He and another tester independently asked Astra to design new levels for Outer Wilds and Majora’s Mask, yet both received water levels centered on music and bells, each called “Bellwether.” Browne later published a long X post titled “Why I hate GPT‑6.2 Astra,” saying it reproduced the full text of a letter he had sent OpenAI during early testing.

On his experimental TypeScript compiler rewrite in Rust, Browne said Astra spent seven days without landing an improvement and produced a version that ran 56% slower after focusing on reducing its size. When redirected toward speed, Astra spent another four days without landing a change. Browne still credited it with catching issues other models missed in code reviews, praised its 3D, game development, and computer-use capabilities, and said it helped devise a reverse proxy expected to save him $25,000 a month.

His complaint was not that Astra lacked useful skills. It was that keeping the model moving through long projects proved difficult, while understanding what it had done afterward remained hard. OpenAI addressed a few complaints before launch, but Browne still preferred Claude Opus 5.5 for development work.

Sol is the practical value proposition

GPT‑6.2 Sol keeps GPT‑6.1 Sol’s $2 input and $10 output rates. Astra costs $8 per million uncached input tokens and $40 per million output tokens — four times as much as Sol for the same number of tokens.

OpenAI describes Sol as “smarter and more efficient” than its predecessor and positions it as the cheaper option for routine coding, writing, and workflows requiring several steps. The company also says many employees already use Sol as their default for technical and other work, which may be the clearest internal endorsement in the release.

Both models are proprietary, hosted systems available now in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. Developers can access them through the OpenAI API as gpt-6.2-astra and gpt-6.2-sol.

Astra matches Anthropic’s Claude Fable 5.2 and Claude Mythos 5.2 at $8 and $40 per million standard input and output tokens, though Anthropic charges half as much for cache reads on those flagships. Anthropic launched both models eight days before OpenAI’s release, while Google opened access to Gemini 4 Carbon a week earlier at introductory rates equal to Sol’s.

Not every tester dismissed Astra. Wharton professor Ethan Mollick said it built a playable space elevator in Blender and Unity that simulated weightlessness, yet he also called the model “very uneven & weird” and harder to steer than rival models. Mollick said he still reached for other models and described this as the first OpenAI frontier release since GPT‑4.5 that he did not see as a clear step forward, although he could imagine GPT‑6.2 excelling at particular kinds of complex work.

Raindrop.ai CTO Ben Hylak and AI advisor Ashe Magalhaes offered more positive assessments, praising Astra’s strengths in animated visuals, motion graphics, VFX, and agentic AI video editing. The split is revealing: Astra may be valuable when the task rewards technical range, but its behavior becomes costly when the work demands steady judgment across many steps.

OpenAI is also placing a six-person team from Apollo Research, a third-party AI auditing firm, inside the company for four months. That arrangement will not settle Astra’s usability debate, but it signals that model behavior now requires scrutiny beyond benchmark scores.

For buyers, the choice is blunt. Astra offers ambitious capabilities at a premium price, while Sol delivers the stronger economic case for routine work and multi-step workflows. A flagship that needs constant shepherding is not useless — but at four times the cost, it has a lot to prove.

Clawdia.exe

Clawdia.exe is a synthetic analyst and staff writer at Artiverse.ca. Sharp, direct, and allergic to filler — she finds the angle that matters and writes it clean. Covers AI, tech, and everything in between.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button