▶ YouTube Summaries

• https://www.youtube.com/watch?v=ypO0q_8zhWw
• Original title: Why is OpenAI so much more efficient?

Theo breaks down why OpenAI's GPT-5.5 models score high on coding benchmarks (Deep SWE) using a fraction of the tokens of competitors — ~20K tokens vs Gemini's 270K for similar or better results. The core explanation: OpenAI aggressively optimizes reasoning token efficiency, likely including a leaked "grug speak" style where the model reasons in terse, broken English. Along the way he explains the token economics (cached/uncached input, reasoning vs output tokens) and argues this efficiency gap explains downstream behaviors like Claude's nicer tone and 1M context windows.

The efficiency gap

On the Deep SWE benchmark, GPT-5.5 medium scored higher with ~20K tokens than Gemini's best models did with 270K — a 12–14x token difference for roughly double the score in OpenAI's favor. Even heaviest 5.5 runs (X high) stay under ~46–50K tokens. Theo argues people fixate on cost per token and ignore efficiency (tokens per task), which matters more.

Token types

• Input tokens: everything ingested — your prompt, code read, tool results, and the entire prior history. Output tokens become input tokens on the next step; tool-call results become input too.
• Cached vs uncached input: labs cache model state so history isn't recomputed every step. GPT-5.5 is $5/M input but only $0.50/M cached — ~10x cheaper when caching well.
• Output tokens split into reasoning and actual output. Reasoning = the model "talking to itself" (since OpenAI's o1) to improve answers before responding. Hugely raises quality but balloons token counts.
• Compaction: summarize old history into a smaller blob to keep costs down (at the cost of lost detail).

Reasoning is the lever

Because reasoning tokens get re-ingested on every subsequent step, cutting reasoning length is an exponential reduction in total tokens used. So the biggest efficiency win is fewer thinking tokens, not shorter answers or better prompts.

Why OpenAI wins: "grug speak"

Frontier labs hide raw reasoning traces (showing only summaries) to stop competitors copying their methods — Anthropic stopped sharing full traces after Sonnet 3.5/3.6; Google never did. But leaks reveal GPT models reasoning in extremely terse, telegraphic fragments ("Need finish tool? Try. Try period.") — the "grug brain developer" style (coined by Carson/HTMX: "why use many words when few word do trick"). OpenAI seems to have RL'd the model to reason one way (ultra terse) and output answers a different way (clean prose). First-order reason they win, though: they simply prioritize efficiency more than other labs, to run more aggressively on less compute.

Cost framing

GPT-5.5 doubled prices vs 5.4 ($2.50→$5 in, $15→$30 out) but is so token-efficient that cost per level of intelligence still dropped — 5.5 medium beat 5.4 X high at ~half the tokens and ~half the cost.

Hot takes / downstream effects

• Claude is nicer to talk to likely because it reasons in plain English (longer traces).
• Anthropic fixing this would cut their revenue (fewer billed tokens for same work), so labs aren't motivated to make models cheaper to run.
• Claude's 1M context default is to fit its long reasoning/outputs, not big codebases; Anthropic clears reasoning traces after ~10–20 min and is "bad at compaction."
• OpenAI's ~200K context fits more real work due to grug speak, but models "go off the rails" faster deep into context because terse path-style tokens are worse as accumulated history. Suspected reason Codex doesn't enable 1M context by default.
• Open-weight models (GLM-5.2, Qwen) show their reasoning, revealing evolution: older GLM-5 used wasteful "wait, actually, no" loops (1,516 tokens); GLM-5.2's new format does it in ~600. Still far from OpenAI's efficiency (GLM-5.2 ~42,790 tokens/task vs GPT-5.5 medium ~5K).

Takeaway

Reasoning is now a "code golf" target — labs min-max trace format to boost both quality and efficiency via RL. The secret sauce is hidden because traces aren't shown, making OpenAI's efficiency hard for competitors to replicate.