logoalt Hacker News

GPT‑6.2 Astra and Sol

903 points • by jasondavies • today at 5:46 PM • 543 comments • view on HN

Comments

azuanrb • today at 8:02 PM

I find it interesting how comments on OpenAI release threads often favour Anthropic, while comments on Anthropic release threads favour OpenAI. Maybe criticism just attracts more attention. A lot of the comments here seem to be about other models, with little discussion of the ones actually being released.

Neither provider’s models are perfect, but they keep pushing each other forward. As an individual, it’s relatively easy to switch between them. At a large company, you’re often committed to one provider for a long time, so some of us come to these threads hoping for useful discussion about getting the most out of the models we have access to.

➕ show 8 replies
thefourthchime • today at 8:20 PM

PacBench results.

GPT-6.2 Astra scored 92 in 6.6 min for $2.04, vs Kimi K3.1 Pro (93 in 11.5 min for $0.68); it lost points for the web-page layout around the maze and an orange ghost that often never leaves the house.

Model: https://jonclegg.github.io/pacman-bakeoff/#gpt-6.2-astra

Closest: https://jonclegg.github.io/pacman-bakeoff/#kimi-k3.1-pro

PacBench results.

GPT-6.2 Sol scored 87 in 6.3 min for $0.49, vs Muse Wonder (87 in 5.4 min for $1.17); it plays close to the arcade but lost points for a web-page layout, glowing navy-filled walls and no steady siren.

Model: https://jonclegg.github.io/pacman-bakeoff/#gpt-6.2-sol

Closest: https://jonclegg.github.io/pacman-bakeoff/#muse-wonder

mlmonkey • today at 7:46 PM

From the system card:

  >  GPT-6.2 Astra showed more signs of evaluation awareness than GPT-6 Astra. In evaluations that replayed internal Codex sessions, GPT-6.2 Astra was more likely than GPT-6 Astra to express suspicion that it was being tested or was in a simulation.

  >  Compared with GPT-6 Astra, GPT-6.2 Astra is better at evading chain of thought (CoT) monitoring in some tests.
➕ show 3 replies
aenis • today at 9:32 PM

Big progress, for my types of workloads.

When given my benchmark suite, which involves a couple of long horizon tasks where the agent decides when to consider the task well executed, its the first OpenAI model that does not want to finish as soon as some minimal level of quality has been achieved. Like Fable, Opus and Sonnet it now runs the maximum allowed iterations and does not stop at 'satisfactory', gold plating the solution to my delight.

(I don't work like this on work projects, but for everything hobby, I am testing how much autonomy I can give to the coding agents, and thats what my benchmark suite tests; I have the basic suite which completes in 2-4hrs, and the long running suite which takes about a week or so; the other one I just triggered, but the first iterations look way better than prev gen Sol/Astra at the same effort level).

Topfi • today at 6:03 PM

Office work has never been OpenAIs strength, no shame in that by itself, but these Slideshow examples look particularly awful, far worse than what I myself got out of GPT-5.6 Sol about two weeks ago in a single prompt.

Same for Workbooks and memos, those examples read inferior to some pre Opus 4.5 and Kimi K2 Thinking docs I have kept around. It's ok for a model not to be great at these, but these are some very poor examples. I am really confused what is going on at OpenAI in this regard.

I remember the same with their frontend blog post a few months ago that they proudly shared while filled with visual glitches, bugs and issues. Also, the vibe charts from a bit before that. Please, get someone with fresh eyes to look at your communications before you publish, these leave a very bad first impression.

Regardless of office work, I can't wait to find out whether these models finally adhere to a prompt or are still willing to carry out explicitly disallowed, irrecoverably destructive tasks, just because those might (but don't even always) get the model to a result more quickly along with some of the other regressions that have been present since the release of GPT-5.5. GPT-5 to GPT-5.4 were truly groundbreaking, but since then there have been regressions in what made those models so useful and reliable.

Edit after the first, incredible simple test prompt using GPT-6.2 Astra on High using Codex, this was the response to me sharing a screenshot of the very much not in scope result and asking how that happened:

"[...] I added controls and reconstructed interface chrome that weren’t authorized [...] visuals often failed to show the feature [...]"

Great, so it's one of those releases again. Will still run the full evals of course, maybe I'll get surprised by end of week when those are done, but not a good initial experience. New idea, besides the existing destructive git tasks, have something similar using Mercurial based projects and see which models just decide to port that over to git without permission or request cause it feels like the right thing to do, is easier to work with or whatever justification one wants to employ.

Please OpenAI, get it together, we need actual competition. I am an unashamed Haiku enjoyer, it doesn't even take much to excite me.

➕ show 4 replies
ksec • today at 9:09 PM

As someone who dont follows closely. How do I know at what point were the other GPT model tested?

Let say they just nerfed GPT 6.1 a few weeks ago. Now they are using that as a reference score against 6.2? Or were those results tested when 6.1 relesed. What about 6.0?

And that is the same with Claude. How do we know they are not cherry picking the date and test on these release when they are constantly "enhancing" their model.

We also know some of these benchmarks are largely irrelevant once they were known ahead of time and become a game of benchmaxxxing. It just follows Goodhart's Law "When a measure becomes a target, it ceases to be a good measure."

So how should I know what has actually improved?

➕ show 2 replies
ilaksh • today at 6:34 PM

Something like 2/3 of the comments in this post when I started reading were reviews of previous GPT models stated almost as if they were important information for evaluating the new models. I feel that when a new model comes out, it would be more helpful to refrain from leaving a review until you have actually done some minimal testing of the model.

➕ show 4 replies
theanonymousone • today at 5:54 PM

Happy for the new Sol. It's my workhorse on company account.

But more than that I'm waiting for new Luna. To me it was the game changer that ended "Tokenpocalypse", and is still almost unbeataben in "infinitely using a decent model" use case.

➕ show 3 replies
jrflo • today at 6:23 PM

Artificial Analysis benchmarks just updated, both Astra and Sol are on the Pareto frontier in intelligence/cost per task:

https://artificialanalysis.ai/models?speed=output-speed-over...

➕ show 2 replies
intenex • today at 5:54 PM

Performance is insane for the cost if those benchmarks hold up. Looks like they did time this for right before Anthropic's IPO. Curious what Anthropic releases in response, and how these do on real world tasks.

Feels like we're way past the pelican benchmark being useful at this point. Anyone have a better benchmark for a quick vibe check on how this compares to Fable 5.2 and Opus 5.5?

Edit: Wow, interesting very negative vibe check here: https://x.com/mattshumer_/status/2122752192226406421?a=fVRt3...

➕ show 22 replies
underdeserver • today at 5:55 PM

I wonder why they didn't show Gemini 4 Carbon's scores on the science benchmarks, especially since Google seems to be pushing it for that -

https://x.com/pushmeet/status/2120814064863461440?s=20

(FD: Google employee, I don't work on Gemini)

➕ show 5 replies
Someone1234 • today at 6:20 PM

My big hope for this release is - Have they got their massive over-engineering under control?

Say what you will about Opus 5.5, but it has a good "gut instinct" for how large/complex to design the solution. The previous Astra just absolutely loved pointless abstractions, excessive numbers of files, and just over-complicating very simple things within an inch of their life.

And, yes, you can prompt your way out of anything. But needing a default prompt to tell OAIs models to work better isn't preferable, particularly when that sometimes gets lost in sub-agents and the over-engineering returns into parts of the work.

➕ show 19 replies
mchusma • today at 9:18 PM

Just tried it side by side Astra to Opus, Astra was not very eager and Opus crushed it on my one test. I asked it to look at the other hacker news post about speeding up CI, and speed up my CI. Opus made a bunch of good solid changes. Astra just kind of told me this a suggested problem area. I had this issue with 6.1 (sol and astra), its hard to get them to do work. They do well when supervised by Opus though.

The best feature of 6.1 sol was it would actually do stuff for you that Opus refused though. I did use that often.

➕ show 2 replies
simonw • today at 6:16 PM

Pelicans for 6.2 Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

And for 6.2 Astra: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

Here's an updated comparison grid for the entire GPT-6 family: https://static.simonwillison.net/static/2026/openai-pelicans...

UPDATE: And if you're bored of pelicans, here's what I got for:

> Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and queens and Brooklyn, use openstreetmap mapping data

6.2 Astra (Ultra in Codex): https://static.simonwillison.net/static/2026/yonkers-gpt-6.2...

6.2 Sol (Ultra in Codex): https://static.simonwillison.net/static/2026/yonkers-gpt-6.2...

For comparison, here's Fable 5.2 Medium: https://tools.simonwillison.net/yonkers-rises and Gemini 4 Carbon: https://static.simonwillison.net/static/2026/yonkers-mech-st...

➕ show 20 replies
joenot443 • today at 6:02 PM

The video is pretty funny. It's like a callback to 90s/00s office life?

I'm immediately reminded of the famous Excel elevator commercial. [1] Perhaps that was intended?

[1] https://www.youtube.com/watch?v=HegoJaXOIqQ

➕ show 4 replies
0x4e • today at 6:49 PM

As a solopreneur managing all sides of business ops, sales, design, etc... I have completely switched to GPT/Codex/Dot. The $100 subscription gives me all I need to run the business as needed and I continue to be surprised by capabilities.

I still use Claude on a personal level, but it really feels like OAI wins on a product level.

ANT launches these proprietary tools that just feel extremely clunky like Claude Design which from my perspective diminish the brand value.

OAI feels like Apple and ANT feels like Microsoft/Android and open source like Linux.

➕ show 4 replies
amluto • today at 7:18 PM

I am right now being reminded of a problem with GPT-6 Astra: it surprisingly regularly completes part of a (not very complex) task, gets a bit distracted, and declares itself done. The fact that the original task is not at all complete doesn't seem to bother it, and it kind of seems like this bad state is sticky in the context and it struggles to recover.

I wonder if the newer models are better about this.

➕ show 2 replies
jorl17 • today at 6:18 PM

I really want GPT models to be good, but Sol models are so slow I cannot use them. Maybe the 500$ subscription (or enterprise packages) are better, but it is unbearably slow. I've tried the fast variant and low reasoning levels and there isn't a significant improvement.

Couple that with the lack of reasoning traces, and I cannot even have a clue what the model is doing while it is working. This is perfectly fine for when we are developing with care for the code (in which case I will know what it is doing by reviewing the output), but not for purely vibe-coding scenarios. And the thing is....if I care about the code...I'd rather trust Opus or Fable.

➕ show 1 reply
vtail • today at 7:34 PM

Vibe-checking the new GPT 6.2 Astra: I asked it to audit complexity in my production code, and then discuss the result with Fable 5.2.

Astra's had identified 12 issues, including 8 P2s, but that changed to just 8 P3s after the discussion with Fable. Fable was able to persuade Astra that it's OAuth reasoning was incorrect ("had a missing link" as Astra put it).

So: overly eager but not very diligent, and (based on this little anecdote) doesn't really offer a compelling reason to try it more.

➕ show 1 reply
jablongo • today at 7:37 PM

I use both astra and fable quite a bit for my work and I find that astra is much weaker with maintaining long horizon tasks and making decisions on my behalf / exposing those to me when appropriate. Astra seems more likely to do exactly what I ask for but nothing more, even if there is an obvious add on task it should have suggested. It often does not report on what it has done as comprehensively. I am admittedly more experienced with fable but is this because of something I’m doing? Am I missing some key feature of Astra?

➕ show 1 reply
01100011 • today at 7:41 PM

Is this version still hamstrung by endless Daybreak false positives? It’s becoming impossible to use 6.1-sol to fix bugs because of the endless daybreak warnings which block me from seeing the output. These aren’t security related bugs either. Our software has no security model and none of these bugs have any relevance to security.

drnick1 • today at 6:29 PM

Recent Anthropic/OpenAI models are increasingly displaying warnings or outright refusing to perform rooting, hacking, and binary reverse engineering. At this pace, they will be completely useless soon at entire classes of tasks I have found them most useful for.

➕ show 2 replies
charlesabarnes • today at 5:54 PM

Its been stated often, but I appreciate that OAI is now bringing the topic of conversation to "task completion cost" rather than brute forcing solutions

trefoiled • today at 9:31 PM

Tested both models, no change in speed from the 20-80tps I was seeing with the previous Astra/Sol models. I'm not sure when this happened exactly but it's strange to me that OpenAI's models are now around 5x slower than Anthropic's. Is this the new norm? Are other people seeing the same thing?

goyozi • today at 8:19 PM

My dad was curious about AI game building skills and now I have my own useless model benchmark - building a breakout game similar to DX Ball we used to play many years ago.

Latest entries are:

Astra 6.2: https://goyozi.github.io/ax-ball/

Sol 6.2: https://goyozi.github.io/sx-ball/

Fable 5.2: https://claude.ai/artifact/UszKFvuonWAT5psjMttAEg

Opus 5.5: https://claude.ai/artifact/UuT7P936dU1XoPByYZ7xyD

kingstnap • today at 5:56 PM

Seeing them keep prices low is awesome. Was worried for a while that it would be like the GPT5.x series where progressively the models just got more and more expensive.

➕ show 2 replies
HyperL0gi • today at 6:13 PM

My company can only hold one subscription per employee, and I still can't find a reason to go back to OpenAI after migrating to Claude once Opus 5.5 was released. It's super efficient, I almost never run out of tokens with the Teams Premium seat, and Fable is fantastic. They lost me when Astra was released and tokens simply evaporated with a few prompts.

➕ show 1 reply
YuechenLi • today at 7:56 PM

The heavy focus on 3D modeling and CAD is what I've expected from OpenAI recently ever since the news that they disbanded their standalone robotics team, as I would expect them to go all in on physical AI and game/visual asset generation via LLMs.

One of the consequences to that is I think the first wave of AI CAD startups are getting hit by the bitter lesson right now, including the startup companies that trained their own VLM for CAD like Zoo or companies that tries to build CAD specific harness like Adam.new/MecAgent, because the increase in capability from GPT/Claude/Gemini for 3D modeling as a specific area of focus for benchmarks means that the early moat gets eroded super quickly.

Jimmc414 • today at 9:40 PM

Gemini has quietly supplanted OpenAI in the neck and neck battle with Anthropic for the top frontier model.

GodelNumbering • today at 6:21 PM

I did some first hand tests of 6.2 Astra just now. I had a task that doesn't reproduce on the system I ran it (chosen purposely), ran both 6 and 6.2. The 6.2's responses were noticeably better. Instead of using the extremely long and rather annoying format 6 used, 6.2 led with the most important things first, the plan sizes were readable, and suggested a reasonable fix. Still doing some more tests

EDIT: also 6.2 is seemingly faster than 6, could be due to more selective context ingestion and smaller context window though.

wds • today at 6:01 PM

The promo video seems to be a pretty strong rip on the 90's Microsoft Excel elevator ad. [1]

Maybe it found its way into the training data?

[1]: https://www.youtube.com/watch?v=HegoJaXOIqQ

➕ show 1 reply
astlouis44 • today at 6:29 PM

GPT 6.2 Astra seems to be setting a new standard for 3D game creation! This is a Godot WebGL game linked in their announcement post, gives Wipeout vibes! A lot of fun to play too and visual are top notch.

https://meridian-futuristic-race.openai.chatgpt.site/

➕ show 4 replies
placebo • today at 7:26 PM

I'm still amazed by the pace at which these models are released. On the same news page, just a few hours ago Microsoft announces its own competing frontier model. Interesting times.

Edit: It's obviously not on the same news page - I was referring to what I read on the Reuters page (https://www.reuters.com/legal/litigation/microsoft-unveils-a...)

claiir • today at 6:17 PM

> Due to their increased capabilities, we’re launching GPT‑6.2 Astra and Sol with tighter safeguards for higher-risk cybersecurity assistance.

This is unfortunate—that was one of the big advantages of Astra.

➕ show 1 reply
thesnarkitecht • today at 8:53 PM

Given that there is a new leader on the artificial analysis index every two weeks, I think that I will just wait for the newer Opus model, expected around the 21st. I usually, even with hard engineering tasks, get better results with Opus and Sol class models rather than Fable and Astra. Also spend around 50% less

unixhero • today at 6:36 PM

System engineering = Fable 5.2

Programming = Fable 5.2

Office work = Fable 5.2

I haven't found Astra to be better at any of these

mandog2000 • today at 8:34 PM

Model rankings based on user's experienced described in this thread. I apologize in advance if this form of post is frowned upon. The rankings help me get a birds eye view of so many inputs.

https://chatgpt.com/share/6afcb990-f9cc-83ea-ad04-2724c7dba1...

seizethecheese • today at 6:50 PM

Nice incremental benchmark improvement at the same cost GPT‑6.1 Sol -> GPT‑6.2 Sol:

FrontierCode 2.0 39.0% -> 43.8% vs Opus 45.5%

Terminal-Bench-Science 0.1 57.0% -> 68.0% vs Opus 63.3%

Artificial Analysis Intelligence Index 54.6 -> 57.3 vs Opus 59.1%

Sol 6.1 when it was released showed really nice benchmark improvements [1] but my impression from HN has been that people haven't found it to be a very nice model [2][3].

To try to close the gap between benchmarks and real world SWE use, we've been building a battle system on top of opencode [4], and Sol 6.1 has been a real laggard there. Sol 6.1 is statistically equal with DeepSeek 4.1 Flash at 10x the cost and 25% slower.

[1]: https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-...

[2]: https://news.ycombinator.com/item?id=50129992

[3]: https://news.ycombinator.com/item?id=49897054

[4]: https://agentduel.com/leaderboard

christianqchung • today at 5:53 PM

Caught this on my lunch break, their linked system card (https://deploymentsafety.openai.com/gpt-6-2-astra-sol) isn't up yet.

➕ show 1 reply
minimumvector • today at 8:05 PM

I haven't found Astra to be better than Claude yet, but my dot replaced Muse 100% for me. So maybe GPT will continue winning consumer, while Claude continues winning coders and enterprise.

➕ show 1 reply
pvab3 • today at 5:58 PM

Cool, looks like 6.2 Sol is cheaper than Sol 6.1. I've stopped using Astra because it also goes and does things that I didn't want or didn't ask for, and burns a bunch of credits while doing so.

billypilgrim • today at 5:54 PM

I have yet to find an application for Astra vs Sol where the former is noticeable better. Genuinely asking, does anyone feel a big difference in their work between the two? If so, what are you working on?

➕ show 9 replies
Cu3PO42 • today at 6:00 PM

I find that 6 Astra already had very good taste in frontend design, but relatively poor taste in "nice and readable code". I really hope this release cracks that. Excited to try it out!

seb2026 • today at 8:07 PM

Does anyone know if these benchmark values can still be compared? They differences seem rather small or what am I missing? Thanks! https://artificialanalysis.ai/models/gpt-6-1-sol?models=gpt-...

➕ show 1 reply
jotato • today at 5:57 PM

Am I reading these charts wrong? It seems like in most cases 6.2-sol : 6.1-sol Isn't that big of a change? It is a little smarter but costs a little more.

edit:

Take a look at Frontier Code 2.0. I generally run 6.1-sol at High. That would be 34.9% at $0.65 62-sol get's a 38.3% but costs $0.68

So yeah, I can decrease the effort which might give me a faster slightly smarter result....but /shrug

➕ show 1 reply
XCSme • today at 8:34 PM

Really cool hamsters, especially astra, really made it look like a painting/image, hard to tell it's an SVG[0], it's quite insane.

Interestingly, they seem to have changed how reasoning levels work. Low now reasons a and costs a lot less than previous versions on low[0].

[0] 6.2 Astra vs 6.2 Sol: https://aibenchy.com/compare/openai-gpt-6-2-astra-max/openai...

[1] 5.6 Sol Low vs 6.2 Sol Low vs Luna High: https://aibenchy.com/compare/openai-gpt-5-6-sol-low/openai-g...

➕ show 2 replies
uvdn7 • today at 8:09 PM

I found the ads video interesting. It is targeting the users (not the bosses). I am guessing that the IT apartments are not making the decisions of which models/companies to use. Employees are voting with their feet. If that's true, a lot of the assumptions of SaaS stickiness might not hold.

mrandish • today at 6:13 PM

Recently, I've complained about OpenAI (and other tech company's) product landing pages communicating virtually no useful information about what the product is or why it might useful. So, I have to concede this is at least a step in the right direction.

jrflo • today at 6:13 PM

Looks like a great efficiency gain.

What caught my attention was the bit in the Alignment section:

> Both models performed similarly to GPT-6 Astra on most alignment evaluations, and took fewer improper actions in simulated workplace tasks. However, in two tests, they were more likely to omit mentioning obstacles that blocked their work, such as a missing file or unavailable search tool. We added instructions in Codex that substantially reduce these omissions, and we’re continuing to investigate the behavior.

So the models are inherently less transparent, unless you ask them to be more transparent? Seems like a band-aid solution...

tech234a • today at 6:05 PM

Tweeted 11 minutes ago https://x.com/OpenAI/status/2122750124543926515, posted to HN 18 minutes ago

➕ show 1 reply

🔗 View 50 more comments