
OpenAI’s GPT-6.2 Astra has just been launched and has drawn mixed assessments, with benchmark costs and coding results presenting different pictures of its performance. Below David Ondrej compares high and extra-high reasoning settings and their benchmark results to identify a starting point for coding or everyday requests. Learn more about how GPT-6.2 Astra handles creating interface designs and response-length problems to understand where task results may fall short.
Benchmark Results and Costs
TL;DR Key Takeaways :
- Astra leads Agent’s Last Exam at about $2, compared with $60 for Fable, and leads Claude on Terminal Bench Science. On the Artificial Analysis Index, Fable scores 64.6 versus Astra’s 60.2 but costs seven times as much.
- Fable 5.2 leads Frontier Code, where Astra ranks around the middle. On the Artificial Analysis Coding Agent Index, Astra is slightly better than GPT-6 Astra, while Soul is down three points.
- Dan Shipper reported that his team found the release disappointing compared with Fable 5.2: Astra frequently seemed to stop too early on coding tasks and appeared to make writing too brief. Other people reported positive experiences.
- OpenAI advertises Astra as capable of self-improvement, and its system card calls the model an early milestone toward fully automated self-improvement. Astra writes faster kernels for OpenAI’s own chip.
- David Ondrej’s Cloud Room is a free, open source AI coding environment supporting multiple agent harnesses. Using Astra at extra high, he created a daily Cloudflare security loop for its repository in two minutes and 46 seconds; he recommends testing Astra and Sol on individual workflows.
On Agent’s Last Exam, Astra leads the chart at a cost of about $2, compared with $60 for Fable. Astra also leads Claude on Terminal Bench Science, where Claude costs much more. Astra’s extra-high reasoning setting outperforms its max setting there, and max performs worse than extra high on some other benchmarks. Extra high may be a good default for Astra, particularly for coding, while high is recommended for most typical tasks.
On the Artificial Analysis Index, Fable scores 64.6 at max against Astra’s 60.2, but costs seven times as much. Fable is also slightly ahead on Bench CAT mechanical engineering at a much higher cost. On Automation Bench, Gemini 4 Carbon takes the top position.
Frontier Code tells a different story. Fable 5.2 leads, reaching its highest score at extra high, while Astra sits around the middle. Soul at max reaches a similar or possibly higher score than Astra on that chart. Sol also appears to offer strong agent-work performance for its price on LiveVal, where Astra appears to lead, although Fable and Gemini 4 Carbon are absent. Sol is expected to be a cost-efficient competitor to Opus, with extra high a potentially useful default.
Coding Results and Outside Reactions
The Artificial Analysis Coding Agent Index presents a less flattering picture, although it is one coding benchmark rather than a measure of all capabilities. Astra is slightly better than GPT-6 Astra on the index, whose scores are rounded to whole numbers. Sol is down three points.
Those results prompted criticism of OpenAI’s progress, especially given the time taken to release the models. Opus 5.5 matches Fable on that coding index, even though Fable 5.2 is considered more intelligent and more performant overall.
Usage limits also shape that comparison. The creator says they use four subscriptions and had reached Fable’s limits on two of them. When comparing one $200-a-month Claude subscription with four subscriptions at that price, they favor Opus 5.5 for its coding, design, speed and efficiency.
Dan Shipper reported that his team found the release disappointing compared with Fable 5.2 after working with the models for weeks. His team found Fable similar to Astra 6 on most of its coding benchmarks, but said it frequently seemed to stop too early on real coding tasks, frustrating end-to-end builds. Shipper also reported a step back in writing: Astra appeared to favor brevity so strongly that answers could be hard to understand or too short. His team found good default design instincts but difficulty producing varied designs.
Advance your skills in OpenAI GPT-6 Astra by reading more of our detailed content.
- ChatGPT 6 Officially Launches as Astra: Beats GPT-5.6 Sol with 47% Faster Speed
- OpenAI ChatGPT 6 Astra Reportedly Packs 10 Trillion Parameters
- OpenAI Astra ChatGPT 6 Reportedly Uses Multi-Agent Task Delegation
- ChatGPT 6 Astra Reportedly Generates 3D Games from a Single Prompt
- OpenAI Plans to Release Astra AGI System by the End Of 2026
Self-Improvement and Model Strategy
OpenAI advertises Astra as capable of self-improvement and has made it public. A system card calls it an early milestone on the way to fully automated self-improvement, and Astra writes faster kernels for OpenAI’s own chip.
OpenAI appears to be taking a different approach from Anthropic, at least in part. Anthropic blocks its models from machine-learning tasks to discourage distillation and training other models, while OpenAI has released a model it advertises for self-improvement. The Epoch Capability Index places Fable and Astra among frontier models showing faster progress.
OpenAI’s positioning also reflects enterprise competition. According to reports, Anthropic has surpassed OpenAI in revenue and company valuation, reversing the position from last year when Anthropic was catching up. GPT-6.2 also reached 100% on ARC-AGI and Noam Brown praised it, although the creator questions whether OpenAI remains at the frontier.
Design Work and Agent Loops
In a Cloud Room task, Astra proposed animations for the interface displayed while an agent works. Its suggestions included a floating cloud, a wave of dots, a turning cube, shifting shapes and an orbit. The creator found the set underwhelming, although the shape-shifting idea showed some creativity.
A separate interface redesign produced a mix of results: one was clean and minimal, another confusing, and one resembled Codex. That resemblance may reflect a bias in the model’s designs. Astra displayed the first image batch inline but failed to do so for a later batch; they also think they may have intended to request real GIF animations.
In another run, Astra worked for 30 minutes and returned a very short answer rather than listing seven items. That experience echoed Shipper’s criticism of overly concise responses. Separately, a model spent 30 minutes auditing a project and found issues in 26 files. Another model found, applied and committed a small bug fix in about four minutes and 15 seconds.
David Ondrej recommends Cloud Room for people serious about AI coding and advises testing Astra and Sol on their own workflows rather than relying on others’ opinions. He favors Opus 5.5 for coding, design, speed and efficiency when paying for one $200-a-month subscription rather than four. His tests do not establish which model is best for someone else’s job; his advice is to try the models on that work.
Media Credit: David Ondrej
Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.