▶ YouTube Summaries

• https://www.youtube.com/watch?v=sKmrLtB47WA
• Original title: This is absolute chaos...

GPT-5.6 Soul is excellent but drains Codex rate limits far faster than 5.5 did, and Theo has watched users (and himself) burn through 5-hour limits repeatedly. He argues much of the pain is OpenAI's fault via Codex defaults, then gives practical, tested advice — avoid Ultra, turn off fast mode, default to high (or medium) reasoning, tame sub-agents, and above all define explicit stopping points in prompts — that he says cut his token use to a fourth or fifth while raising output quality.

Why 5.6 burns so much more than 5.5

5.5 was cheap in practice mainly because it constantly stopped to ask permission and lost track of context, so each message used little (~0.1–2% of a 5-hour limit). 5.6 fixed the stopping and context problems, so it runs much longer per message (up to ~15% of a 5-hour limit on high/XHigh/max), which multiplies usage. Fast mode gives 1.5x speed but burns 2.5x faster — tolerable on 5.5's small messages, brutal on 5.6's long ones (a single message can approach half a 5-hour limit).

Limit structure and recent changes

Subs have a 5-hour limit (resets 5 hours after your first message) and a weekly limit (~4–5 of the 5-hour limits). OpenAI temporarily removed the 5-hour limit for Plus/Business/Pro, leaving only the weekly one. That is convenient but dangerous: previously a runaway Ultra run would be stopped by the 5-hour cap after ~25% of weekly; now one prompt can blow the whole week. Theo notes a trick of running a tiny cron job every 5 hours to keep the rolling window "started," now unnecessary in the interim. Tibo/Teemu also announced ~10% efficiency gains, a usage reset, and that the context limit was reverted from 372K back to 272K after they found it was over-charging usage — so drain should drop.

Practical recommendations

• Ultra: don't touch it for now (dedicated video coming). It is the single biggest cause of runaway burn.
• Fast mode: turn it off. It mattered on 5.5 because you sat watching the model; 5.6 runs so long that inference speed is no longer the bottleneck (tool calls and test runs dominate). Threads routinely run 8+ hours.
• Reasoning level: low, medium, and high are all very good; high is his default because the model self-regulates — simple tasks stop fast, so leaving it on high is cheap. On DeepSWE: low 45% (~$1/task), medium 61% (~$1.86), high 69% (~$3.47, tied with Fable 5's 70% but at a quarter the ~$13 cost). XHigh only reaches 71% for more money and max reaches 73% at ~$8.39 — a >2x cost jump for ~4%. Other benches (Cursor Bench) show different gaps, so it depends on the work, but high sits at the sweet spot of the cost/intelligence curve. /goal is no longer needed with 5.6.
• Sub-agents: both Codex implementations (V1 and V2) are rough; 5.6 was trained to reach for them eagerly and sometimes spins them up needlessly. If you see limit burn plus lots of sub-agents, add to global AGENTS.md: "Only use subagents if the user explicitly requests them." Claude Code and Cursor have meaningfully better sub-agent implementations — Theo has hacked his Codex sub into Claude Code (Tibo blessed it, promising resets to anyone banned).
• Model selection: don't use Luna for coding (it's for programmatic/bulk use over API); Terra is a fine middle ground, but Theo just defaults to Soul on low/medium/high.
• Don't manually shrink the context window / force early compaction — bad advice going around. The model was trained on Codex's specific compaction levels; overriding makes it dumber, triggers costly compaction more often, and saves nothing. Tibo confirmed there is no extra charge above 270K and the threshold is tuned for 5.6 Soul.

The most important tip: define the stopping point

5.6 is extremely eager and won't stop unless told to, so put the stop signs in the prompt yourself. Examples: "write a plan, then stop and ask for feedback before proceeding," or a much longer leash like "build it out, use computer-use to test, keep going until it works, put up a PR, babysit the first round of review comments and address them, then stop — I'll take it from there." Realizing the end is something you define in the prompt (not via tools or harness config) lets you safely let the model run far. Theo urges experimentation over copying others' skills/configs: rewrite your AGENTS.md and claude.md by hand, read your agents' traces, and tune until the model feels right. He notes Open Code's team accidentally ran 5.6 on medium for a month (a reasoning-level key mismatch many tools hit) yet still called it their favorite — a strong endorsement of medium/high as defaults. Sponsors: WorkOS (enterprise + agent auth, Auth.md), G2i (engineer hiring).