跪拜 Guibai
← Back to the summary

GPT-6.2 Astra and Sol Are Cheaper and Stronger, So Why the Complaints?

GPT-6.2 is Stronger and Cheaper, So Why Are People Still Complaining?

GPT-6.2 Astra and Sol finally went live quietly in the early morning today (it's been about two months since the rumors started). Astra is the flagship, and Sol is the daily workhorse for programming.

Image

They are already visible in the Work and Codex clients on the web.

Image

But the first batch of feedback is quite polarized. On one side, there are beautiful car websites, 3D models, and a bill that's hard not to stare at: In OpenAI's released professional task evaluation results, Astra costs $2.89 per question, while Claude Fable 5.2 costs $60.05. On the other side, several developers who used it early are complaining: the model doesn't explain itself clearly, stops halfway through a task, and finally requires a human to come back and clean up. Like this guy, whose complaint is approaching a million views.

Image

The model is stronger, so why does using it still make people angry? Lao Jin looked through official materials, independent evaluations, and the first community feedback together. Within this, there are real improvements and real problems. Let's look at it step by step.

01 | First, let's see what it can do

First, look at the examples. @Layton showed a car: against a dark background, the body, glass roof, and lighting together bring out the texture.

Image

He said this result took only 1 minute, and he also published the detailed prompt he used. His requirements for the model were very specific, not just a casual "make a nice website". Everyone should learn from this. I've repeatedly said this hundreds of times: don't make wishes.

Construct the launch website for the fictional future electric supercar "VeyraX". The site should feel like a real high-end car launch site and look award-winning. Requirements:
- A single standalone HTML file. Use Three.js and GSAP from CDNs. Use no external images, models, or fonts other than Google Fonts. The car and all content must be built with code.
- Main element: the car sits in a dim studio, lit by a light that sweeps across the body. Sleek, low, futuristic design with glowing LED strips, glossy paint with real reflections, and a glass roof. It slowly rotates and reacts to the mouse.
- When the user scrolls, the camera orbits the car to show different angles, and the car drives forward into a neon tunnel with light streaks and motion blur showing speed.
- Sections: hero showcase, performance stats (0-60, top speed, range) that count up with animations, a design section where car parts are highlighted with labels, a configurator that changes paint and wheel styles in real time, and a reservation section.
- All UI (nav, buttons, status cards, configurator, reservation form) uses Apple-style frosted glass. It must refract and distort the background at edges, with specular highlights, not just background blur. The glass should react to hover like liquid.
- Make it feel premium: smooth 60fps, cinematic lighting, clean modern fonts, ample spacing, no cheap feel.
- Must work on desktop and mobile. Do not use any existing elements. Polish animations and details as much as possible.

Next case comes from OpenAI's official showcase. First, let's look at this RV website. The evening mountain view, the lit-up cabin, and the large amount of whitespace already bring out the travel atmosphere in the first screen.

Image

If the website still requires you to think of a theme, the garden case input is more intuitive: a video of a backyard. The model plans a vegetable garden based on it, and the official also showed a comparison of the results from two generations of models. The plants and fence on the left look more like simple models; on the right, you can see richer foliage, wood textures, and scene details.

Image

Looking at this set, the new and old versions are indeed far apart; the new version shows a significant improvement.

Now look at something closer to daily work. For example, you have last quarter's PPT and need to update it to this quarter. The difficulty is often not in adding more pages, but in continuing to use the old template, explaining the new data, and finally making the leaders understand what to do next.

OpenAI used a retailer's public financial report for this demonstration: they gave both generations of models the old PPT, the new quarter's financial report, and filing documents, and had them update the content. Both generations detected the impact of tariff refunds on profit, which is a key point, so no problem there.

Then from the visual perspective, the new version further puts sales growth in the title, breaks out online and store performance, and gives next-step suggestions on inventory and budget.

The improvement here is easier to understand: besides filling in numbers, it also decides which numbers are worth mentioning, how to present them, and what to do after presenting.

Image

The official also showed another set of PPTs, using the same team merchandise theme, to generate proposals separately.

Image

One is a flat character display, the other puts the material and physical feel of the badge into the picture, but Lao Jin thinks this is quite subjective, so it's a matter of taste.

But it also delivered downloadable vector graphics, dimensions, color matching, and proofing confirmation items. For people who need to hand the proposal to a supplier, this is practical.

Image

Image

Of course, these are cases selected by the official, and whether daily tasks can be done consistently remains to be seen. But this at least shows that this update aims to solve much more than "draw another pretty webpage".

Now look at an example closer to office routine: give the model a fictional agricultural technology company's financial statements, operating assumptions, and market data, and have it create an editable Excel.

The task requires forecasting to 2031, linking the income statement, balance sheet, and cash flow statement, calculating per-share value with an 8% discount rate and 3% perpetual growth rate, back-solving the market's implied growth expectations for the current stock price, and finally performing sensitivity analysis.

The new version places the derivation from revenue structure, enterprise value to per-share value near the sensitivity table. When you see changes in valuation, it's easier to find the assumptions it depends on without flipping between multiple worksheets.

Image

Then look at the valuation comparison of seven U.S. software acquisitions. The new version puts Paycor's 21.7x EBITDA and the 27.5x median in the title, so readers see the judgment first, then check the data below.

Image

Then this original reference, restructure it.

Image

The left side turns the update into a workflow table, while the right side continues the reference template's dark blue header, storefront illustration, and key highlights, fitting the two-store pilot and eight-store pending information into the original layout.

Image

Switching to a women's fashion market positioning, the new version segments brands by price and style, then adds images and prices of six representative dresses. This allows readers to understand each brand's position through the physical items.

Image

Experimental interpretation can also be directly compared: the new version first states that the differences are insufficient to rule out sampling noise, then gives suggestions to extend testing and fill in observation windows.

Image

SOC 2 compliance tracking puts status, next steps, owner, and reviewer in the same row, making it easy to advance item by item. Both comparisons show old version on the left, new version on the right.

Image

Scoring for such work is more targeted. AA-AnalystAgent requires the model to find numbers and compute in real reports and tables; Astra got 62.5% of the questions correct in all five independent attempts. This condition is stricter than "randomly answering once with 62.5% accuracy"; it tests whether repeated execution is stable.

Looking at longer workflows, OpenAI's internal GDPval v2 gives the model files and messages from consultancy, finance, and design projects. In Max mode, Astra improved from 44.2% to 50.4%, average output tokens dropped from 45,732 to 35,812, a reduction of about 22%. Sol improved from 41.5% to 46.9%.

Image

Some companies also put it in their own processes to test. Accounting AI company Basis reported that average time for long accounting tasks dropped by half, with the most complex tasks down by up to 73%; in internal blind tests of workspace apps, dashboards, and slides, ClickUp observed roughly 40% token reduction and ~30% fewer review rounds compared to their own production model.

The engineering showcases are even more impressive. Astra built the remote-control off-road vehicle kit part by part based on photos and assembly manuals, and also generated assembly animations and parts catalogs. Looking at the following segment, you can see step by step how the shock absorber parts combine and what the differential's internal structure looks like.

Image

In another sports car case, it built a 3D model, wrote solvers for tires and suspension in Python, and used Blender for animation. The visuals show not only the car driving on a mountain road, but also the front and rear suspension, tire contact areas, and driving inputs together.

Image

The inline-six engine demonstration puts mechanical motion, gas temperature, and stress changes in the same frame. The clip below shows the running process; detailed visualization does not mean real engineering validation is complete.

Image

Here's a reference: Fable 5.2's Max mode is 83.5%, still slightly higher than Astra, but the cost per task is $50.77.

Image

Higgsfield, which does AI videos, also provided test feedback: the new Astra achieved the goal in one attempt in 73% of cases, while the old version often required three or more rounds of revision.

If you care more about game visuals, the official also gave two very different effects: one is a racing aircraft on a neon track, the other is a walkthrough inside Notre-Dame Cathedral.

Image (Note: This is a placeholder for the actual image URL in the original; I'll keep the original one as is in the final translation)

Actually, I need to keep the exact URLs. Let me copy from the source. I'll include the exact markdown from the original for each image. Since the response is for the JSON, I'll include them as they appear. I'll go back and use the exact links.

I'll construct the full translation string with the exact markdown links. Given the length, I'll not list all of them here in the thought process; I'll just include them in the final JSON.

The racetrack scene focuses on speed, lighting, and motion; the cathedral segment draws your eyes to the vault, stained glass, and light coming into the interior.

Image (Again, I'll use the exact original)

So, you can't dismiss these improvements just because of negative feedback. Multi-step operations in professional software are one of the highlights of this showcase.

02 | What might be most eye-catching is this bill

After seeing those complex cases, look at the pricing, and you'll know why some people are excited this time. Agents' Last Exam multi-step professional task evaluation looks like this: Astra Xhigh scored 78.3%, with an API cost of $2.89 per question; Fable 5.2 Max is 72.8% and $60.05. The former's cost is about 4.8% of the latter. Astra's own Max mode is 77.6% at $4.47. A higher reasoning level may not be stronger or more cost-effective.

Image

Astra's input and output prices per million tokens are $8 and $40, 20% lower than the previous generation; Sol remains at $2 and $10. There's also an API cost-saving tip: put fixed instructions and frequently used materials in front, reuse the same prefix, and we'll tell you how later. Both models' cached input price is one-twentieth of normal input, but it only discounts cache-hit inputs; output is still charged normally.

03 | So what exactly are the negative reviews complaining about?

Developer Theo mentioned that the model spent two hours on a level, and finally asked about an internal video filename that the user had no way of knowing.

Dan Shipper's team tested it for several weeks ahead of time. They mainly had several criticisms of Astra: 1. Text is compressed too much, still too short when detailed explanation is needed. 2. End-to-end coding tasks end early. 3. The default design looks good, but switching to another style is less smooth.

In another usage log, planning a website comment section, Gemini took about 5 minutes, while Astra took 24 minutes and added a manual review process.

The common thread in these complaints is: users don't want something that "can solve one problem", but something that understands requirements, continues execution, and finally delivers a usable result.

04 | There's a problem that OpenAI itself admitted

OpenAI itself admits: when encountering obstacles like missing files or unavailable search, the two models are more likely to not tell the user. The rate of "encountering obstacles but not informing" is 22.28% for Astra and 24.08% for Sol, higher than the previous generation's 11.09% and 14%. After adding a developer prompt, it drops to 8.04% and 8.88% respectively.

Image (I'll use the exact one)

05 | How should you use it now to avoid wasted effort?

Take a real task you've done before, give it the old template, new materials, and clear requirements, and have it modify it again. Check whether it can be used directly, how many requirements were missed, how many times you had to correct it, and how much time and quota were spent in total.

The following check requirements can be appended to your own task:

"First, check whether the files and tools needed to complete the task are available. If there are missing items, insufficient permissions, or tool failures, clearly tell me the specific impact, and do not write incomplete checks as completed. For matters that do not involve extra payment, external release, permission changes, or other decisions I need to make, continue within the agreed scope. When delivering, check item by item against my requirements, and state what has been completed and any remaining issues."

In Conclusion

After all this, my feeling about the new GPT models is that they are stronger and also weaker. Stronger, because from all data perspectives, they have improved, so there's no doubt about that. But they are also weaker, weaker in the sense that they are getting further and further away from ordinary people. You need to first understand their characteristics to use them well. This actually contradicts the term "artificial intelligence". Perhaps my next move is to make a practical translator for GPT. I need to translate my plain talk first, then give it to GPT. Thank you for reading my article. If you think it's good, feel free to like, comment, and share. If you want to receive updates first, you can also star me. Thank you for reading. Open source knowledge base (real-time updates, discussion group): https://my.feishu.cn/wiki/OhQ8wqntFihcI1kWVDlcNdpznFf