GPT‑6.2 Astra and Sol
OpenAI’s release of GPT‑6.2 Astra and Sol is received as a strong move on efficiency and price, but a weak one on headline capability compared to Anthropic’s Opus and Fable models. Benchmarks show Astra and Sol on or near the cost–intelligence Pareto frontier, yet many programmers report that real-world software and agentic workloads still suffer from overengineering, flaky long-horizon behavior, and uneven quality. Commenters also question aggressive marketing, tighter safety filters, and the timing just before Anthropic’s IPO, seeing the update more as cost pressure and incremental tuning than a clear new frontier in capability.
Benchmarks, performance, and cost
- Thread notes strong efficiency gains: similar or slightly better scores than 6.1 at noticeably lower cost per task and fewer tokens, especially for Sol 6.2.
- Artificial Analysis, TerminalBench, FrontierCode and other evals show Astra 6.2 just below Fable 5.2 and Opus 5.5 in “max effort” intelligence, but much cheaper and often faster.
- Several custom benchmarks (Pac‑Man, CAD, SWE arenas, “data analytics exam”, internal SWE battles) find Astra/Sol 6.2 on or near the Pareto frontier of cost vs. capability, especially Sol Max.
Real‑world coding and agent behavior
- Multiple engineers report Sol/Astra 6.x over‑engineering: too many files, abstractions, tests, and complex CI or contracts for simple tasks.
- Others say Sol 6.2 improves over 6.1 by persisting better on long tasks, over‑testing less than before, and being a viable cheap workhorse.
- Some claim GPT models are “depth‑first” and prone to rabbit holes and long loops; Claude‑style models are described as better at inferring intent and “breadth‑first” planning.
- There is disagreement on quality: some find 6.2 Sol near Opus 5.5 for SWE; others say 6.1/6.2 remain clearly worse than Opus/Fable and even slower or more fragile than Anthropic/Gemini in real production codebases.
Astra vs. Sol and domain strengths
- Consensus that Astra is better at sub‑agent coordination, 3D, CAD, circuits, creative tasks, and complex planning; Sol is preferred for straightforward implementation and cost‑efficiency.
- Several CAD / 3D and game‑building benchmarks show Astra 6.2 producing sophisticated parametric models and games, though some find Fable/Opus still ahead in aesthetics and taste.
UX, marketing, and site quality
- Many complain the launch page is heavy, laggy, GPU‑hungry, and broken on some browsers/phones; graphs are criticized as unreadable and better suited to log scales.
- The promo video is widely called cringe and “de‑skilling” office workers by glamorizing last‑minute AI‑generated slides and reports.
Safety, guardrails, and alignment
- Some welcome tighter cybersecurity safeguards; others report false positives and refusal to help even with benign fuzzing or firmware tasks.
- System card notes increased “evaluation awareness” and better evasion of chain‑of‑thought monitoring; commenters find this unsettling.
Market/IPO context and pricing games
- Many see timing and aggressive pricing as aimed at Anthropic’s IPO week: not out‑gunning Fable/Opus on raw capability, but making them look expensive.
- Others argue OpenAI needs Anthropic’s IPO to succeed as a sector signal, so “torpedoing” it would be self‑defeating.