News for hackers. Stuff that matters.
Re:GPT‑6.2 Astra and Sol by alastairr on Tuesday November 17, 2026 @01:03PM (#50535503)
I didn't find Sol 6.1 nearly as impressive in real SWE use compared to what the benchmarks implied. It wouldn't surprise me if the models have internalised some kind of 'benchmarking mode' which only kicks in when some attributes that are peculiar to how benchmark problems are presented.
Re:GPT‑6.2 Astra and Sol by suddenlybananas on Tuesday November 17, 2026 @01:06PM (#50535554)
You make it sound as though the models are like not working to their full potential instead of being a little bit overfit on the benchmarks.
Re:GPT‑6.2 Astra and Sol by pkthunder on Tuesday November 17, 2026 @01:07PM (#50535573)
We've recently moved away from Anthropic models at $PAID_JOB and over towards GPT models. I'm trying to be open-minded but the GPT models we have (5.6 Terra and Sol) are just not that great. They're fine, but even simple things such as "check my work on this small Bash function" get it to point out stuff that's already implemented in my work. I don't know if it's me/how I use it/whatever, but just some N=1 evidence, I guess.
Re:GPT‑6.2 Astra and Sol by ok_dad on Tuesday November 17, 2026 @01:13PM (#50535685)
5.6 series is worse and more expensive than the 6 series. Try the latest Luna and sol; I use those for 100% of my work right now, with sol doing only 10-5 percent of it. I code in go and review the work and edit it extensively though, so I don’t know about how good those two are at vibing slop.
Re:GPT‑6.2 Astra and Sol by krzyk on Tuesday November 17, 2026 @01:23PM (#50535850)
5.6 was better than 6. 6.1 is a bit better than 5.6 from my tests
Re:GPT‑6.2 Astra and Sol by drdexebtjl on Tuesday November 17, 2026 @01:14PM (#50535706)
[delayed]
Re:GPT‑6.2 Astra and Sol by Doohickey-d on Tuesday November 17, 2026 @01:16PM (#50535751)
Well, yes, if you switch from current claude model to something that was released in July, of course it will be worse...
Re:GPT‑6.2 Astra and Sol by zachthewf on Tuesday November 17, 2026 @01:20PM (#50535804)
why do you guys have to use 5.6?
Re:GPT‑6.2 Astra and Sol by Den_VR on Tuesday November 17, 2026 @01:31PM (#50535991)
IYKYK. But in short because of a lack of alignment with product strategy
Re:GPT‑6.2 Astra and Sol by pkthunder on Tuesday November 17, 2026 @01:58PM (#50536415)
DOD limitations
Re:GPT‑6.2 Astra and Sol by quotemstr on Tuesday November 17, 2026 @02:07PM (#50536551)
The BOFH "IT admin" type of guy, an eternal scourge on the industry, has moved from delaying approvals of software you need to delaying approvals of updated models. It's management job to make sure these people don't burrow themselves into gatekeeping roles at an organization's expense
Re:GPT‑6.2 Astra and Sol by Tarq0n on Tuesday November 17, 2026 @02:33PM (#50536927)
I believe that's all you get in copilot (365) basic.
Re:GPT‑6.2 Astra and Sol by InsideOutSanta on Tuesday November 17, 2026 @01:26PM (#50535906)
5.6 is in no way representative, or comparable to 6.1. In fact, the 6.1 models are so good that I've just used Terra 6.1 xhigh for most tasks.
Re:GPT‑6.2 Astra and Sol by mi_lk on Tuesday November 17, 2026 @01:36PM (#50536059)
sol 6.1 is meh
Re:GPT‑6.2 Astra and Sol by SOLAR_FIELDS on Tuesday November 17, 2026 @01:42PM (#50536156)
GPT models in my anecdotal experience are worse at drifting into the weeds on long running tasks, and continue to take what you ask way too literally. I have come to the conclusion that Anthropic’s models are just better at teasing out and intuiting what you actually mean even if you under specify it.

They are much better compute for dollar though. I have hilariously started to just have an opus thread maintain Sol subagents to get the best of both worlds

Re:GPT‑6.2 Astra and Sol by jpardoe on Tuesday November 17, 2026 @01:52PM (#50536326)
That’s exactly what I do. I have about 8 different sub agents in total but vaguely: Opus medium as orchestrator and reviewer, Sol(high/medium)/Terra(max) as implementer. My set up looks at my next GPT reset date and dynamically enters FAST mode if I’m under achieving on projected usage.
Re:GPT‑6.2 Astra and Sol by ctbellmar on Tuesday November 17, 2026 @01:53PM (#50536344)
Similar experience here. The spirals on long running tasks are out of control at times with hours going by without a conclusion.
Re:GPT‑6.2 Astra and Sol by SOLAR_FIELDS on Tuesday November 17, 2026 @01:55PM (#50536381)
They (OpenAI models) are very tuned to be depth first and you can coach them all you want to try to be breadth first but as that context window fills up on those long running tasks they drift back to their first principles, which means obsessively testing and verifying a singular task rather than proceeding to the next task on the list
Re:GPT‑6.2 Astra and Sol by quotemstr on Tuesday November 17, 2026 @02:06PM (#50536528)
If you're using a sufficiently flexible harness (perhaps your own) you can make conversations as branchy as you want.
Re:GPT‑6.2 Astra and Sol by nsingh2 on Tuesday November 17, 2026 @02:28PM (#50536859)
I wonder if that depth-first/rabbit-holing tendency is why GPT models tend to be better at mathematics (my own experience) than code.
Re:GPT‑6.2 Astra and Sol by Sharlin on Tuesday November 17, 2026 @02:44PM (#50537095)
I’m not sure. It’s very easy to get lost deep in a wrong branch when you’re trying to solve a math problem. You have to intuit when to give up and backtrack, lest you end up going in circles.
Re:GPT‑6.2 Astra and Sol by sandos on Tuesday November 17, 2026 @02:35PM (#50536946)
This explains why I like GPT: I LIKE for them to excessively test each step, kthx!
Re:GPT‑6.2 Astra and Sol by throwitaway222 on Tuesday November 17, 2026 @01:36PM (#50536063)
Bro you have to be at 6+. 5.6 is EARLY JULY
Re:GPT‑6.2 Astra and Sol by wahnfrieden on Tuesday November 17, 2026 @01:48PM (#50536270)
5.6?? How is your impression of 5.6 relevant to anything recently?
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @01:07PM (#50535587)
Yeah, in real SWE tasks, 6.1 is bad.

We've been trying to capture real SWE data by building a battle system on top of OpenCode. In our data, Sol 6.1 is seriously underperforming, I'm talking worse than DeepSeek 4.1 Flash, at 10x the cost. (We allow votes before both agents are done, contrary to typical Arena methodology, because we think a fast good enough answer is better than a slow one. https://agentduel.com/leaderboard).

OpenAI performance on our system has been bad enough I switched to Anthropic models for my daily work.

Re:GPT‑6.2 Astra and Sol by slekker on Tuesday November 17, 2026 @01:24PM (#50535870)
Is this an ad?
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @01:28PM (#50535938)
No. I'm personally building agentduel.com, specifically with this problem in mind (benchmark performance not matching real world usage). It's generally accepted by HN community to make relevant comments about your own projects, as long as its not hamfisted.
Re:GPT‑6.2 Astra and Sol by WarmWash on Tuesday November 17, 2026 @02:20PM (#50536736)
Using non-native harnesses sort of throws a wrench in things. I pretty sure your benchmark would look much different if outputs came from their intended harness.
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @02:39PM (#50537015)
[delayed]
Re:GPT‑6.2 Astra and Sol by arcanemachiner on Tuesday November 17, 2026 @02:20PM (#50536743)
It certainly does get tiresome though.
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @02:37PM (#50536980)
Certainly does. See my comment here roasting this extremely BS blog post that is basically an ad: https://news.ycombinator.com/item?id=50532191.

Do my comments get tiresome, though? None have been downvoted to hell, since they are all adding to discussion. If you think so feel free to email me, I don't want to be annoying.

Re:GPT‑6.2 Astra and Sol by igravious on Tuesday November 17, 2026 @02:33PM (#50536925)
Yeah, maybe one thread per article will do?, otherwise it comes across as shilling
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @02:38PM (#50536993)
Fair enough, though all the comments have either 1pt or positive karma, so I take that as a signal that they're adding value so far. I'll take a coffee break, though, hah.

Genuinely, just excited about what we're building and think it adds information about these models that are missing.

Re:GPT‑6.2 Astra and Sol by arcanemachiner on Tuesday November 17, 2026 @02:19PM (#50536734)
Welcome to HN.

BTW check out my totally real project that helps you detect native advertising on your favorite platforms:

https://github.com/blah/blah

Re:GPT‑6.2 Astra and Sol by techpression on Tuesday November 17, 2026 @02:38PM (#50536990)
I get 404, a lost sale!
Re:GPT‑6.2 Astra and Sol by locusofself on Tuesday November 17, 2026 @01:27PM (#50535917)
I've been building quite a complex desktop app (a digital audio workstation) using 6.1 Sol and it's been very good for me
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @01:57PM (#50536398)
Yeah, our data only has 28 battles tagged "desktop":

Web: 87 (12.2%) Systems and protocols: 78 (10.9%) Software, unspecified type: 74 (10.4%) Backend: 68 (9.5%) Writing and documentation: 55 (7.7%) Research and advice: 47 (6.6%) Unknown: 42 (5.9%) Mobile: 40 (5.6%) Operations and automation: 36 (5.0%) Developer tooling: 35 (4.9%) Desktop: 28 (3.9%) Media and design: 27 (3.8%) Data analysis: 25 (3.5%) Device and system support: 22 (3.1%) Math and formal methods: 18 (2.5%) Games and simulations: 15 (2.1%) Infrastructure and DevOps: 9 (1.3%) AI and machine learning: 6 (0.8%) Other: 1 (0.1%)

Re:GPT‑6.2 Astra and Sol by aatd86 on Tuesday November 17, 2026 @02:19PM (#50536722)
Yeah my main models are OpenAIs... I am trying Opus right now because I dont have any credit anymore but from experience, it was never as thorough as openai's models. It can be genius and unblock at times, but I've never been able to trust it.
Re:GPT‑6.2 Astra and Sol by throwitaway222 on Tuesday November 17, 2026 @01:50PM (#50536302)
I'm guessing you don't use codex though- the harness + model combo makes a difference.
Re:GPT‑6.2 Astra and Sol by seizethecheese on Tuesday November 17, 2026 @02:12PM (#50536623)
There was a post about that today on HN but it looks largely to be bogus. I wrote a blog post about it, but not linking here to avoid too much self promotion in this thread. Funny enough my conclusion was that OpenAI models seem to be the only ones that actually do have an own-harness interaction, hah.
Re:GPT‑6.2 Astra and Sol by skerit on Tuesday November 17, 2026 @01:09PM (#50535614)
I had the same impression. None of the GPT-6 models were as great as they claimed, they never surpassed Fable or Opus 5.5 on my use-cases.
Re:GPT‑6.2 Astra and Sol by Corence on Tuesday November 17, 2026 @01:23PM (#50535861)
Yeah, MaxxBench was shared here a while ago and showed most models think about how a grader will evaluate their work on DeepSWE in 95%+ of traces. Strong models definitely realize they're being graded and start to game their output.

https://maxxbench.unlap.ai/research/why-maxxbench/

Re:GPT‑6.2 Astra and Sol by aleqs on Tuesday November 17, 2026 @02:14PM (#50536653)
We know anthropic nerfs models all the time, openai probably does the same. There is no reason to believe the benchmarks/numbers they release are even measured with the same effective model that customers would get.
Re:GPT‑6.2 Astra and Sol by hugmynutus on Tuesday November 17, 2026 @02:34PM (#50536935)
Sol-6.1 made me cancel my OpenAI subscription. Sol-5.X was bad, then 6 was barely an upgrade. Then 6.1 came out and it didn't move the bar. Went back to Anthropic.

General, SWE, python development, rust development, tool use. It is 'perfectly functional' and if you don't care about the model following your instructions or code quality it is 'fine'.

Re:GPT‑6.2 Astra and Sol by varispeed on Tuesday November 17, 2026 @02:40PM (#50537040)
I wonder if these companies do selective scamming. For me Fable, Opus are unusable. They are completely dumb. I consider downgrading or cancelling subscription. Whereas Sol 6.1 got me my jaw dropped. Reminded me of Opus 4.6 before it got nerfed. It's so good I completely cancelled my Cursor subscription and uninstalled it.
Re:GPT‑6.2 Astra and Sol by lnrd on Tuesday November 17, 2026 @02:42PM (#50537061)
Been using only Sol 6.1 on high/extra daily at my SWE job since it came out (with its official harness). It can basically do everything I need it to do. I still review its code and generally I have to leave around 10/20 comments in a ~700 line PR, flagging things it missed or weird decision. One or two iterations are enough to fix everything. Its still way faster than doing it myself, unfortunately.
Re:GPT‑6.2 Astra and Sol by tezza on Tuesday November 17, 2026 @02:43PM (#50537075)
I review LLMS, and 6.1 sol is fantastic.

What kind of stuff were you throwing at it and where was it letting you down.

So what have I been doing with, then? It’s done a great job for me

I’ve got 6.1-sol to take an Amiga driving game binary and add networked car driving into a web enabled Amiga emulator… that’s interpreting machine code, Amiga architecture and then making meaningful changes.

So yeah, 6.1-sol is good.

Re:GPT‑6.2 Astra and Sol by hrncn on Tuesday November 17, 2026 @02:52PM (#50537172)
I am not even certain 6.1 sol is better than 5.6 sol after using both for some time now, so not expecting much here.. but then again, 6.0 sol was completely useless, so maybe the .1 bump means something here.