GPT-5.6 Sol Ultra will be in Codex

Inference cost cuts & “compute multipliers”

• Thread connects GPT‑5.6 Sol / Ultra to reports that OpenAI found a way to halve inference costs.
• People speculate on techniques: shared KV/prefix caching, speculative decoding, more efficient architectures, DeepSeek‑style tricks (draft models, DSpark‑like systems).
• Several argue simple response caching or prefix caching won’t help much for long‑tail, long‑context workloads (e.g., coding), where costs matter most.
• “Compute multipliers” are seen as the core trade secrets of labs; keeping them secret is compared to quant trading algorithms. Others criticize this secrecy as anti‑scientific and socially harmful.

Sol Ultra, Pro, subagents & workflows

• Sol Ultra in Codex is described (based on source code) as: same model/“max” effort, plus a prompt nudge to proactively use subagents. No special backend mode exposed.
• GPT‑5.5 Pro is believed to use parallel test‑time compute (multiple passes + a selection/synthesis model), explaining its much higher effective cost.
• Comparisons drawn to Claude Code’s “ultracode” / dynamic workflows: scripts that orchestrate many subagents deterministically, with some disagreement on how mature or “broken” these are.

Model quality, competition & access

• Many see Fable/Mythos as current coding leaders, especially on complex, multi‑step tasks (e.g., reverse‑engineering hardware).
• GPT‑5.5 Pro is praised for quality but criticized as prohibitively expensive in API usage; some hope 5.6 Sol Ultra narrows the gap at subscription prices.
• Some users report 5.5 getting “lazier”/worse recently in Codex, possibly as preparation for 5.6, but this is anecdotal.
• DeepSeek’s open research and low prices are repeatedly cited as pressure on US labs; debate over whether Chinese labs are ahead on efficiency or just structurally forced to optimize.

Enterprise usage, costs & bubble worries

• Corporate users report already seeing 5.6‑Sol Ultra; management initially gamified “most tokens used,” then abruptly pivoted to warnings about high token spend and urging cheaper models.
• Multiple commenters see this as an early sign the economic bubble may deflate: tools are useful but too expensive to replace human work at scale.
• Strategies mentioned: use strong models for planning, smaller ones for “grunt work”; use local models for most tasks and cloud LLMs only when stuck.

Reliability, safety & automation

• Strong skepticism about embedding stochastic LLMs directly in critical, automated business processes: businesses want determinism, auditability, and stable behavior.
• Some argue you can build reliable systems from stochastic parts with feedback and tests; others counter that current LLMs can’t truly learn from deployment‑time feedback and remain opaque.

Branding, naming & UX

• Confusion and annoyance over non‑descriptive names (“Sol”, “Ultra”, “Codex”) versus technical labels; others note marketing and ease‑of‑use often trump descriptive naming.