The Signal — OpenAI shipping instant model + Codex 5.5 gains outside coding reveal combination of speed, reasoning, and personalization as the real moat. SubQ claims (50x cheaper, 20x faster) expose if incumbents genuinely defensible or hype-dependent.
Consensus: Bullish | Conviction: High
What's Moving
- GPT instant model — Sama pushing users off thinking-model-only; latency + personality hitting threshold where switching costs evaporate. (via @sama)
- Codex 5.5 breadth play — Non-coding performance surprising internal users; positioning as general-purpose beats 5.5 on cost-per-task for most workloads. (via @sama, @bindureddy)
- SubQ vapor claim — 50x speed/20x cost vs Opus 4.7 + Thinking; unverified, but testing whether incumbents' 10-50% efficiency gains are defensible or margin cushion. (via @bindureddy)
- Open source → production hard — DeepSeek v4 Pro closing on GPT 5.5, but non-technical teams + LLM brittleness = Opus/GPT still safer bets for hard problems. (via @bindureddy)
- Gemini Flash waiting — Bindureddy betting Flash beats mid-tier Opus/GPT; Google's agentic speed advantage only converts if Flash ships with real routing/memory stack. (via @bindureddy)
Blind Spot — Consensus underweights personalization & memory as differentiation layer. Sama's "memory/personalization feels like more-than-sum-of-the-parts" is throwaway line, but if ChatGPT's account-level continuity is sticky moat that open-source + SubQ haven't solved, cost arbitrage fails. Also: SubQ's 12M context + benchmarks unverified—no one's stress-tested production failure modes. Hype cycle may be priced in already.
One Actionable Idea — Watch if Anthropic/Google respond to SubQ within 2 weeks with independent evals or shipping; silence = credibility hit; response = admits margin compression real.
Sources: @sama (instant + Codex bullish, speed/personality/memory combo), @bindureddy (open-source gaining but hard problems unsolved; SubQ as credibility test), @svpino (latency/TTS focus)