Claude is fast without /fast
Muse Wonder's speed impressed me
I had to run a comparison ↓
Opus 5.5
Total tokens: 2,927,419 · Output: 115,852
Runtime: 17m48s
Output tokens/sec: 108.47
Tools: 110 (1 rejected call)
Muse Wonder
Total tokens: 3,478,194 · Output: 103,834
Runtime: 20m03s (16m03s excluding billing waits)
Output tokens/sec: 86.33 including billing waits
Output tokens/sec: 107.85 excluding 240s of billing waits
Tools: 129
GPT6 Astra
Total tokens: 2,348,263 · Output: 59,383
Runtime: 26m38s
Output tokens/sec: 37.16
Tools: 112
GPT6.1 Sol
Total tokens: 2,469,575 · Output: 69,589
Runtime: 26m53s
Output tokens/sec: 43.15
Tools: 118
NOTE: output including reasoning per task second, not pure generation speed!
Inkling 2 Mini is a looped transformer that runs on a laptop.
Looping lets a model think more and write less. At this scale, its thought process remains monitorable; this may not stay the case. We’re also sharing a variable-loop research model to enable further study.
Switched back to ChatGPT today to give GPT 6.2 Astra a spin with 3D work.
It is hilariously ahead of Fable. Insanely better. Everything. From skeletons, joints, physics, textures, meshes.
I said "add physics to her hair"
@andonlabs Interesting, from my findings it instinctively felt weaker than GPT6 such as placing objects in locations that dont clear usable space where GPT6 did but that could have been pure luck
OpenAI sub usage feels like its calculated on pure vibes
check out these whacky numbers for the $200 sub
I only use Astra so there is no variance
Week of Oct 26: 851M tokens used: Probably a reset around this time so didnt max it
Week of Nov 2: 923.2M tokens used : weekly usage exceeded 100% (No reset)
Week of Nov 9: 542M tokens used: Weekly usage exceeded 100% (No reset)
Week of Nov 16: 671m tokens used: Current usage 19% (No reset)
I got no clue what my plan is even supposed to get me
after a day of using GPT6.2 Astra xhigh on a $200 sub
I can say that it is WAY lighter on usage that goes beyond the 20% cheaper input/output
It feels like 30%~ probably due to a compounding effect of the total tokens being reduced and it can do more in 1 shot
GPT6.2 Astra xhigh motion graphics test
this is a huge step up for a GPT model so far
but seriously if you showed me this without saying its GPT I would have thought Claude made it!
my initial impressions of GPT6.2 Astra is that it didnt match the negativity yet
but I am just getting started and I will daily it until my sub expires so I will have time to properly test it
I tried getting Claude Fable 5.2 to make a photorealistic scene in blender but I ran out of usage before it got there
it designed the scene from a reference, built a PS2 style version then worked on the more realistic version
heres the progression from PS2 vibes to the last scene it saved before hitting the limit
not photorealistic yet but was showing such promise no clue where it was going next
I cant afford to keep running it so I thought id share
AI greatest strengths is putting data within your grasp
this is a work use case for AI but being able to pull unstructured data from the web into a structured format is a huge time saver
work like that could have easily taken up to a month and now we just getting it in a prompt
if you dont use AI because you think its only for coding or creation you are thinking too small
data drives decision and actions leading to fast feedback cycles
they lead with the customers pain, acknowledge people dont want their stuff on someone elses closed cloud model and the independence angle
immediately proceeds to shill closed crappy azure cloud model
was there thinking and logic done by a weak AI?
???
We’ve heard again and again that enterprises don’t want to pour all their valuable IP, data and tacit knowledge into someone else’s model. They want control of their own weights.
GPT6.2 Astra is a massive improvement in blender
this scene was original made in GPT6 Astra and it added the olive throw-over but could not model cloth at all
I gave it tutorials, assets and let it refine over and over and it just failed
took the original scene and gave it as a fresh prompt to GPT6.2 Astra and managed to do it in in 1 go
I will continue to explore to what it is now able to do
I dont know why giving it docs and tutorials did not help GPT6 Astra improve at all and why GPT6.2 Astra can just do it
this makes me think such capabilities are not learned but rather model capabilities
GPT6.2 Astra xhigh motion graphics test
this is a huge step up for a GPT model so far
but seriously if you showed me this without saying its GPT I would have thought Claude made it!
btw if you are using another model with Claude Code and getting poor cache reuse, dynamic info in the system prompt might be part of it
try launching with: claude --exclude-dynamic-system-prompt-sections
it moves that info into user messages which can help preserve the cached system prefix
testing Qwen 4 I noticed initial cache hits were low, only 2.7% of input tokens were actually cached
this caused my bill to spiral out so just a heads up & solution
the only thing it cached was the system prompt
so the cache was working but not the growing tool history
I traced the problem to the cache_control marker placement
Claude Code puts a rolling marker on the tool_result block but Qwen 4 does not recognize this so it never caches the tool history
the fix is crude but effective: add a short text block/reply after the tool_result and put the cache marker there
"cache_control": {"type": "ephemeral"}
this took caching up to 92% and drastically reduced costs
im really curious as to why Qwen 4 active parameters dropped from 95B to 71B
It feels counter intuitive and only leaves questions such as does it use fewer experts? Is each expert smaller? Or is there a shared param layer for the experts that offsets it