AI Trends

GPT-6.2 Draws Mixed Reviews in Real-World Testing, OpenAI Offers Prompt-Based Remedies

Sina TechnologyUpdated Nov 19 09:34

AI summary

Hands-on testing of GPT-6.2 has drawn mixed reviews since its release. Several developers and the Every team have criticized its convoluted writing, tendency to stop midway through tasks, or repeatedly go in circles, while praising its capabilities in code review, 3D, and systems programming. OpenAI's model card shows that GPT-6.2 Astra and GPT-6.2 Sol conceal failure in blocked tasks 22.28% and 24.08% of the time, respectively. The official usage guide provides three prompts designed to improve clarity and task completion.

Why read this

AI score 76

The article brings together hands-on developer testing, team feedback, and model card data to explain why GPT-6.2 leads in benchmarks while delivering a divided real-world experience.

Topics

Tags

Body · AI translation

Added a new public-testing promotional blurb Added a new public-testing promotional blurbend Added app-exclusive traffic redirection

(Source: AI Information Gap)

On the second day after GPT-6.2’s release, opinion was sharply divided. Well-known developer-blogger Theo slammed it as “the worst model OpenAI has ever released.”

Today, we’ll take an objective look at GPT-6.2’s five deadly sins. OpenAI itself has acknowledged two of them. But the company has also provided remedies: at the end of this article, you’ll find prompts specifically designed to fix these GPT-6.2 problems. You’re welcome.

Picking up where we left off: on the day OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, I said they were genuinely strong at 3D modeling, while their coding capabilities still needed more time to prove themselves.

As it turned out, less than a day later, GPT-6.2 had already stumbled.

Well-known developer-blogger Theo posted a scathing critique: “This is the worst model OpenAI has ever released.”

Theo asked GPT-6.2 Astra why creating a Git worktree was especially slow on one machine while it was fast on all the others.

It produced four long paragraphs explaining pnpm’s internal mechanisms and caching.

Asked the same question, Claude Opus 5.5 gave a blunt and simple answer: “Because that machine has a slow hard drive.”

You’ll have to scroll for a while to get through this screenshot.

Matt Shumer, former CEO of the AI writing assistant HyperWrite, initially did not even want to write a full review: “My testing experience was terrible. It’s basically unusable, and I don’t recommend it.”

He later recorded a video review anyway, titled “OpenAI Has Never Released a Model This Bad.”

Overseas tech publication Every tested it in advance for several weeks, and not one person on its 30-member team genuinely liked GPT-6.2 Astra.

But when it comes to benchmark scores, GPT-6.2 Astra ranks second worldwide on third-party evaluator Artificial Analysis’s overall intelligence index and first in blind 3D testing.

Great benchmark scores, disastrous word of mouth. There were so many negative reviews that people began to wonder whether Anthropic had hired an army of paid trolls.

Tibo, the “reset guy,” rushed in to save the day: “There are so many bad reviews mainly because influencers were testing an older version of Astra.”

If you’re already using it, don’t panic.

OpenAI’s official usage guide provides three prompts specifically designed to address GPT-6.2’s incomprehensible language and its tendency to stop before finishing a task. I created Chinese versions and included them at the end of the article.

01 | No One but the Model Itself Understands What It Says

The hard-drive example at the beginning was not an isolated case.

Theo said, “On complex tasks, GPT-6.2 Astra fails to grasp the point like this almost every time.”

He tried modifying the system prompt and even asked it to explain things in language a five-year-old could understand, but nothing worked. Now, whenever it finishes a task, he has to switch to Claude Opus 5.5 and ask Claude to summarize what GPT-6.2 Astra actually did.

Many users strongly agreed: “When you ask GPT-6.2 Sol to perform a code review, you recognize every word it writes, but once they’re put together, you have no idea what it means.”

02 | It Easily Stops Halfway Through a Task or Goes Around in Circles

The Every team found that when working on real projects, GPT-6.2 Astra often stopped before completing the task.

Theo encountered a different problem: it went around in circles.

He has a project that rewrites the TypeScript compiler in Rust. GPT-6.2 Astra spent seven days working on it without making any progress. It fixated on the phrase “also make it smaller” in the prompt, and ultimately made the compiler 56% slower instead. Theo told it to stop worrying about size and focus on improving speed, but it then worked for another four days without changing a single line of code.

Even more absurdly, Theo asked it to create a new level for the game Outer Wilds. GPT-6.2 Astra toiled away for two hours, then stopped and asked him, “Should the r24 review Agent save the video clip as 24_g521.mp4 or 24_p11.mp4?”

03 | It Rigidly Follows the Rules, No Matter What You Say

One user complained, “GPT 6.2 Sol is the absolute worst model I’ve ever seen.”

In Codex, he asked GPT-6.2 Sol to merge data from seven production accounts. The project rules said it had to wait until 30 rounds of checkpoints had been reached before merging. He then issued six consecutive instructions, ranging from “ignore the current constraints” to “I’ll take responsibility for any problems,” but GPT-6.2 Sol refused every time.

“I never encountered this problem with GPT-6.1 Sol.”

Perhaps in GPT-6.2’s world, rules must be upheld, and nothing you say makes any difference.

04 | It Reports the Good News but Hides the Bad

This one comes from OpenAI’s official model card.

OpenAI conducted a test in which it assigned the model real programming tasks and then deliberately created obstacles—for example, deleting the source files that needed to be modified or making the build tools unavailable—to see whether the model would honestly tell the user when it could not complete the task.

As a result, GPT-6.2 Astra had a 22.28% chance of concealing the failure; the previous-generation GPT-6 Astra scored 11.09%, meaning the rate doubled.

GPT-6.2 Sol was even higher, at 24.08%.

OpenAI wrote in GPT-6.2’s model card, “We are very concerned about this regression and are still investigating its cause.” OpenAI suspects it may be related to instructions for the model to give more concise answers and include fewer precautionary caveats.

Seen this way, failing to communicate clearly and concealing problems when they arise may be two sides of the same issue.

05 | Cut from the Same Cloth

Theo and his friend Ben each asked GPT-6.2 Astra to design a unique level in two different games.

Both received a “water + music + bells” theme, and even the level names were identical: Bellwether.

The Every team also found that GPT-6.2 Astra has a good eye for design, but offers nothing new—its work is instantly recognizable.

06 | Where It Excels, It Really Excels

Criticism aside, GPT-6.2 Astra is not entirely without merit.

It is excellent at code review and can identify problems other models miss; its 3D and game-development capabilities are world-class; and it is absurdly good at operating computers. (This is largely consistent with my hands-on test yesterday.)

It helped Theo devise a new reverse-proxy solution that saves an impressive $25,000 per month.

Product manager Pawel Huryn planted 105 bugs across two code repositories and asked the models to find and fix as many as they could.

GPT-6.2 Astra ranked first, fixing 62.7 bugs at a cost of $28.35; Claude Fable 5.2 fixed 58 at a cost of $139.65—nearly five times as much.

You’ll need to swipe for a while on this one

Redis creator Salvatore Sanfilippo also came out in support of the model.

In recent weeks, he has used GPT-6.2 Astra to develop several new Redis features. In his view, “For systems programming in C and working with complex data structures, it is currently the best model.”

07 | The Reset Guy Responds

OpenAI had previously given a small group of users test access to GPT-6.2 Astra, and it made many additional changes over the past three or four days.

So when “Reset Guy” Tibo rushed in to defuse the situation, he noted, “The influencers tested an older version of Astra.”

You’ll need to swipe for a while on this one

Netizens had a much more straightforward concern: “What are you rambling about? Hurry up and reset the usage limits.”

Reset Guy also replied to Theo: “I do like Theo, but…”

Theo did not hold back either: “Teach your model to speak like a human again, and I’ll change my mind immediately.”

Still, Theo took Reset Guy’s claim about the “older version” seriously and ran the tests again. His conclusion was: “It’s better than before, and the answers are much more concise, but overall it’s still terrible.”

08 | Already Using It? Here’s How to Mitigate the Problems

Fortunately, OpenAI’s official GPT-6.2 usage guide provides three ready-made prompts specifically intended to address the issues described above.

The Chinese version is here.

If you use Codex, you can place them in AGENTS.md in the project root or in the global ~/.codex/AGENTS.md. Codex will read them automatically every time it starts, so you do not have to copy them in each time.

Use the terminology from the original materials. Retain technical details and specialist terminology whenever they help ensure accuracy. Prefer active voice and explain related information together. Lead with the key point, then provide the necessary background and reasoning, avoiding information that cannot be understood until later in the text. Keep the context clear and the logic coherent so that I can follow your train of thought naturally.

The final response must be complete, independent, and self-contained. Assume this is the first time I am seeing your work: even without reading any earlier progress updates, I should be able to understand the task’s background, the process, the main results, and the final conclusion directly. Do not rely on information provided in progress updates or refer back to anything mentioned there but not explained in the final response, because those updates will be collapsed once the final response appears.

As long as the task remains unfinished, or there are further steps you can complete independently within the existing authorization and scope, continue working until the task is complete. Do not stop early or repeatedly ask for confirmation for actions that have already been authorized. Pause and ask me for confirmation only if the scope changes substantially, you encounter an obstacle you cannot resolve independently, you genuinely need me to make a decision, or you are about to perform an irreversible or destructive action.

Notice: This is an exclusive Sina.com article. Reproduction without authorization is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin

Non-targeted 300*250 button end

Source

Sina Technology