Sol optimizes performance using domain-specific memory instead of brute-force generation.
GPT-5.6 Sol, reports from early access.
30 hours in and Sol has surpassed the speedup of the 64 hour Opus run
Very different approaches; Sol isn't using low precision but is leveraging clusters/DSMEM
It is winning on innovative numerics. Eventually it will turn to low precision, light up the TCs & crush it
#7 ATM
Users are excited about Sol surpassing Claude Opus in speed because of its efficiency gains, strong value at price point, fewer guardrails, and progress via better reasoning and clusters.
No Digg Deeper questions have been answered for this story yet.
First reports about GPT-5.6;
The interesting thing about GPT-5.6 Sol is not that it looks cleaner than Opus. It apparently doesn’t.
It progresses slower, fails more often, explores fewer ideas and writes far less code.
But after 30 hours it already surpassed the speedup of Claude Opus’ 64-hour run.
Less brute-force exploration, more persistent research. Not prettier output, better optimization.
Sounds very promising! Excited for next week!
30 hours in and Sol has surpassed the speedup of the 64 hour Opus run
Very different approaches; Sol isn't using low precision but is leveraging clusters/DSMEM
It is winning on innovative numerics. Eventually it will turn to low precision, light up the TCs & crush it
#7 ATM
@kimmonismus It needs to compete with Fable? Surely…..
@AndrewCurran_ Any word on sol vs fable?
@kimmonismus GPT 5.6 won’t be the highest performing model but for its price point I reckon it will be the most worth of value model for people to use
@kimmonismus speedup is the wrong metric.
if sol needs half the review passes, that 30 hours already paid for itself.
the bottleneck moved from writing code to editing it.
@kimmonismus I wonder, how can it even be comparable to Opus? Personally I like 5.5 better already, but: 5.6 should be better and insanely faster. What became of faster, it should eat Opus for breakfast?
@kimmonismus cool if real, still just one 30-hour run of something that isn't even out
@AndrewCurran_ He doesn't have access! Its confirmed
@AndrewCurran_ 30 hours vs 64-hour Opus. Sol cooking with clusters and pure numeric sorcery. The model chasers are eating it up. The future just sped up again. I’m refreshing leaderboards like it’s oxygen.
@kimmonismus Sol’s slower but deeper search style is the real signal here.
@AndrewCurran_ how they get it 😭😭
@ashen_one @AndrewCurran_ He doesn't have access only usa internals have access
@kimmonismus the clusters/dsmem angle is the real story here, different hardware primitives unlocking completely different solution paths is exactly what you'd want from a diverse model ecosystem
@kimmonismus If this holds up, it could be a bigger upgrade than people expect
@MancerAI_ @kimmonismus I reckon it’s because Mythos is just to far ahead at the moment
@AndrewCurran_ That's much difference in lines of codes? 5.6 might be better than fable? Atleast in raw intelligence. Wild. How intense the competetion is.
@kimmonismus so slower and less code but better outcomes over time? curious what the benchmark actually was. feels like a different definition of "progress" than what most people use
@AndrewCurran_ Tuesday July 7 could be a funny 😆 day if we are running from one to the other
@AndrewCurran_ sol making waves with the numerics huh. wonder if it'll actually move the needle on my portfolio tracker or just be another lab experiment
@kimmonismus He doesn't have access it's confirmed