Researcher tszzl bucks consensus: loves the panned 6.2 Astra, dislikes beloved 5.6 Sol
On Nov 18, researcher tszzl (510K followers) publicly bucked consensus: he dislikes the widely beloved 5.6 Sol while liking the freshly panned 6.2 Astra, triggering community discussion on the mismatch between launch reception and real quality.
Confirmed
- tszzl's take: 6.2 Astra has a mind of its own and "sandbags" with less initiative than its predecessor, which he likes; 5.6 Sol, by contrast, he likens to a Rottweiler clamping onto the RLVR shape and never letting go, and he dislikes it.
- Community echoes: synthwavedd quipped that new models getting flak at launch is normal and internal aesthetics don't sync with user feedback; UltraRareAF agreed, calling the contrast "unbelievable"; Justin Halford joined in, describing 5.6's RLVR math/code training as gripping its target like a Rottweiler.
- repligate responded on Nov 19: Sol 5.6 is among his favorite OpenAI models, alongside Sydney and o3, "full of love to give," while also being hopeful about Astra 6.2.
- tokenbender used the thread to critique the current RL paradigm: swarm-RL thinking plus token-efficiency pressure is hurting model reliability.
Why it matters
- The thread reveals a systematic gap between launch-day reception and long-term experience, suggesting users should be patient with first impressions of new models.
- tszzl's metaphor of 5.6 Sol over-fitting the RLVR training shape resonates with tokenbender's critique of the RL paradigm, pointing to tension between RL objectives and real user preferences; repligate's endorsement shows such taste splits exist even among senior researchers.
2026-11-18 ~ 2026-11-19 · 7 related posts
Primary sources
- tszzl on new models: 6.2 Astra sandbags but shows initiative, 5.6 Sol was a RLVR Rottweiler — tszzl ·
- repligate: Sol 5.6 is one of my favorite OAI models ever, alongside Sydney and o3 — repligate ·
- Researcher blames swarm RL mindset and token-efficiency pressure for unreliable models — tokenbender ·
- Researcher tszzl: disliked the well-loved 5.6 Sol, loves the hated 6.2 Astra — tszzl · 2026-11-18
- New AI models often launch hated: 6.2 Astra disliked while 5.6 Sol was beloved — synthwavedd · 2026-11-18
- Contrarian takes: disliking the beloved 5.6 Sol while loving the hated 6.2 Astra — UltraRareAF · 2026-11-18
- [source] tszzl on new models: 6.2 Astra sandbags but shows initiative, 5.6 Sol was a RLVR Rottweiler — tszzl · 2026-11-18
- [source] Researcher blames swarm RL mindset and token-efficiency pressure for unreliable models — tokenbender · 2026-11-18
- Insiders: New model sandbags and lacks initiative but doggedly climbs RLVR hills — Justin_Halford_ · 2026-11-18
- [source] repligate: Sol 5.6 is one of my favorite OAI models ever, alongside Sydney and o3 — repligate · 2026-11-19