My Cheatsheets

Scope note: This guide reflects observed behavior and published limits for ChatGPT and major frontier models as of early 2026. Treat numeric limits as indicative; vendors change them without notice.

Core Principles

• LLMs are predictors, not logic engines. They generate plausible continuations, not guaranteed truths.
  • For additional information about how this works, see How LLMs Actually Generate Text (Every Dev Should Know This).
• Think of them as brilliant, distractible assistants. Fast and knowledgeable, but prone to drift, agreement bias, and subtle hallucination.

Tokens & Context

• Context windows continue to expand, but practical reliability still degrades near the ceiling.

  • GPT‑5.2 (ChatGPT): ~256k tokens standard context; Thinking variants support larger effective windows (often ~196k+), tier-dependent.
  • GPT‑4.1 family: up to ~1M tokens in some enterprise/API contexts.
  • Claude 4 (Opus/Sonnet): ~200k–1M tokens depending on plan and variant.
  • Gemini 2.5 Flash / Pro: ~1M tokens input; output typically capped ~65k (Flash documented).
  • Qwen 2.5 (72B Instruct): ~32k tokens (open model cards).
  • Perplexity Sonar: ~128k tokens context reported.

• Context drops still occur when limits are exceeded; older messages are truncated or summarized.

• Best practice: never rely on “infinite” context. Chunk work, summarize checkpoints, and externalize state into documents.

ChatGPT Usage Limits (verified as of early 2026)

Important distinctions

• Manual Thinking requests count toward weekly limits.
• Automatic internal escalation from standard mode does not consume manual Thinking quota.
• UI limits are often lower than API limits.

Model Behavior

What’s Stable

• Focused sessions outperform sprawling threads.
• Explicit structure (steps, constraints, deliverables) improves consistency.
• Repeating key facts or invariants reduces drift.

What’s Changed

• Memory features are now broadly enabled:

  • Preferences and style can persist across sessions.
  • Memory can usually be reviewed, edited, or wiped in settings.

• Safety tuning is stronger and more dynamic:

  • More hedging.
  • More refusals around edge cases.
  • Behavior can change week to week without version bumps.

Confidence vs Accuracy

• Models remain agreeable and can reinforce your mistakes.
• Hedging has increased, but correctness has not improved proportionally.

Rule of thumb:

• Trust for brainstorming, outlining, summarization, code scaffolding.
• Verify for math, logic, legal, medical, financial, or historical facts.

Choosing the Right Model (2026 snapshot)

File & Data Handling

• Uploading many files increases confusion and token pressure.

• Best practice:

  • Work with 2–3 files at a time.
  • Use staged comparisons (A vs B, then fold in C).

• Modality quality matters:

  • OCR errors propagate.
  • Transcripts drop nuance.

Math & Logic (Still Weak Spots)

• Pattern math is strong; symbolic manipulation is fragile.
• Multi‑step reasoning benefits from explicit scaffolding.

Better approach:

• Ask for executable code.
• Use external solvers or calculators for critical results.

Practical Tips

• Specify output shape and depth explicitly.
• Use headings, checkpoints, and acceptance criteria.
• Externalize long‑running work to shared documents.
• Reset threads when contradictions accumulate.

Memory & Personalization

• Memory is active by default in many systems.
• Review stored memory periodically.
• Store preferences and durable context only.
• Prune aggressively.

Considerations

• Cost: long contexts are expensive with diminishing returns.
• Privacy: many providers train on chats unless you opt out.
• Safety: guardrails shift without notice.

Bottom Line

• Keep chats focused and state explicit.
• Know your model’s current limits.
• Verify anything that matters.
• Treat AI as an assistant—not a calculator, database, or oracle.