Which model, at which effort, for which job — and the habits that actually stretch your limits. Every tip carries its evidence level, so you can tell vendor fact from forum folklore.
Ranked by how much they save. The top three are behavioral and dwarf every clever prompt trick on the internet.
Every turn re-sends the whole conversation. A 20-turn chat of 1k-token messages bills ~210k input tokens, not 20k. Long chats also get worse, not just pricier.
Mid-tier model at low/medium effort for everyday work. Escalate to the top model or high effort only for multi-step reasoning. Price gaps inside one vendor are 5–100×.
Output costs ~5× input, and thinking is billed as output. Set length, format and language. Ask for Hebrew only for the final text.
Same task spread over several turns scored 39% worse on average (200k+ simulated chats). Edit your last message instead of stacking "no, I meant…".
A PDF page costs 1.5–3k text tokens plus an image. Paste the relevant section, convert PDFs to text, crop screenshots.
Web search, connectors, deep research, extended thinking, agent mode. Each adds tokens or whole extra loops.
Project files are cached and count less against limits. But instruction files (CLAUDE.md, AGENTS.md) load every turn: keep them short.
In coding agents, pick model and effort at the start. Switching, or returning after the cache expired, re-processes everything at full price.
The mental model that explains almost every tip below. Chat apps and APIs are stateless: the model re-reads the entire conversation on every turn.
Caching makes the re-read cheaper (often ~10% of price) but never free — and it does nothing for the quality loss of long context.
Output costs about five times input on every major vendor (e.g. Opus 5.5 $4 in / $20 out per million). Reasoning tokens are billed as output even when hidden.
Agents carry baggage. Saying "Shalom" (9 tokens) in Claude Code sent 43,483 input tokens — system prompt, tools, memory files. Mostly cached, still counted.
One 1000×1000 image ≈ 1,296 tokens on Claude; a full-HD screenshot ≈ 2,700. Pasted text is usually cheaper and read more reliably.
Long, meandering chats degrade. Spreading one task over several turns cut scores 39% on average (Laban et al., 200k+ simulated chats). Context-rot studies across 18 models show accuracy falls as input grows, even on easy tasks.
Default to the middle. Go down for anything mechanical, up only when a wrong answer is expensive or the reasoning is genuinely multi-step.
| Task | Tier | Effort | In practice |
|---|---|---|---|
| Quick lookup, definition, rephrase a sentencetranslation of a line, a formula, a fact check | Small / fast | Low / off | Haiku 4.5 · ChatGPT Instant · Gemini Fast |
| Everyday drafting and summarizingemails, notes, summarizing one article | Mid | Low | Sonnet 5 · ChatGPT Instant · Gemini 3.8 Flash |
| Academic writing and synthesislit-review section, structuring a thesis chapter, feedback on a draft | Mid → Top | Medium | Sonnet 5 or Opus 5.5 at medium · ChatGPT Medium |
| Statistics, math, debuggingchoosing a test, interpreting output, a bug you can't find | Top | High | Opus 5.5 high · ChatGPT High · Gemini Pro |
| High-stakes analysislong case material, decisions with real consequences | Top | High | Opus 5.5 · Fable 5.1 or GPT-6 Pro only if top tier fails |
| Building an app or agentic codingmulti-file, multi-hour | Top plans, Mid executes | Medium | Claude Code: Opus 5.5 medium (default), or opusplan; subagents on Haiku |
| Research across many sources | Mode choice | Max (rationed) | Deep Research only when the question deserves 20+ sources; otherwise normal search |
| Q&A over a big pile of documents | Different tool | Low | Gemini Notebook (ex-NotebookLM) or a Claude Project — not pasting |
| Bulk extraction / classification (API) | Cheapest | None | GPT-6 Luna · Gemini Flash-Lite · Haiku, via Batch (−50%) |
Lookups, rewriting, formatting, classification, simple Q&A. The default you should reach for most.
Writing that needs structure, planning a document, moderate coding. Anthropic's default for Opus 5.5.
Multi-step math, statistics, hard debugging, analysis where you'd check the answer anyway.
Rarely. Only when High demonstrably failed. Burns limits fast and can make answers worse.
The percentage split is my working heuristic, not a measured statistic.
| Tier | Anthropic | OpenAI | |
|---|---|---|---|
| Small | Haiku 4.5$1 / $5 | GPT-6 Luna$0.10 / $0.50 | Gemini 3.5 Flash-Lite$0.30 / $2.50 |
| Mid | Sonnet 5$2 / $10 | GPT-6 Sol$2 / $10 | Gemini 3.8 Flash$0.75 / $3.75 intro, $1.50 / $7.50 from 2027 |
| Top | Opus 5.5 recommended default$4 / $20 | GPT-6 Astrapremium | Gemini 3.1 Pro$2 / $12 (third-party figure) |
| Ultra | Fable 5.1$10 / $50 | GPT-6 Pro (ChatGPT Pro) | — |
API list prices per million tokens, input / output. On a subscription you don't see dollars; the same ratios show up as how fast your limit bar fills. GPT-6 Sol and Luna launched 22 Sep 2026 and are still rolling out in ChatGPT.
The same text costs more tokens in Hebrew than in English. Newer tokenizers have cut the penalty a lot, but it hasn't gone away.
Weak evidence The three current figures come from one measured paragraph (yuv.ai / Bizportal). The 3.67× figure is a full parallel-corpus study (Petrov et al., NeurIPS 2023). Treat the order of magnitude as reliable, not the decimals.
How each one meters you, what to do, what to avoid, and the community tricks, each with its evidence level. Every platform below changed its limits in 2026, so the in-app meter is the final authority.
/usage shows cache hit rate and what's eating usage (skills, subagents, MCP servers). /context shows what fills the window. Official/clear between unrelated tasks. It costs nothing. /rename first, then /resume later. Official/compact Focus on … at natural breaks. /rewind (Esc Esc) to abandon a wrong path. Officialmodel: haiku. Official/loop tasks and goal check-ins. They resend your full context while you're idle. Officialgh) over MCP servers; disable unused ones in /mcp. Officialnpx ccusage or Claude-Code-Usage-Monitor for burn-rate tracking. Community.claudeignore doesn't exist. Use permissions.deny Read rules instead. Myth/status for context and limits, and /compact, /model, /reasoning. Officialconfig.toml profiles: cheap (small model, model_reasoning_effort="low") and deep (high effort). Communitytool_output_token_limit and model_auto_compact_token_limit. Communitythinking_level (low/medium/high). Don't cap max_output_tokens, which truncates answers you've already paid for. Official/compress, /stats, a lean GEMINI.md, .geminiignore. OfficialPopular advice, checked against what's actually been measured.
| Claim | Verdict | What the evidence says |
|---|---|---|
| "Caveman mode" / terse-speak saves 65–75% | Overstated | JetBrains measured 8.5% fewer output tokens over 86 tasks. The caveman skill itself adds ~1k input tokens per call. |
| "Saying please costs millions — stop" | Folklore | Altman's remark was a joke, not a calculation. "Please" is about one token. |
| "Being rude or polite changes accuracy" | Mixed | One 50-question study found rude prompts slightly better; earlier studies found the opposite. Not a reliable lever. |
| "CAPS / CRITICAL / MUST improves compliance" | Backfires | Anthropic: newer models over-trigger on aggressive language. Dial it back. |
| "Always say 'think step by step'" | Obsolete | ≤3% gain on reasoning models, 20–600% longer responses. |
| "Always tell it to be concise" | Partly | Cuts output ~49%, but lost ~28% on math for older models. Use it for lookups and formatting, not reasoning. |
| "Big context window → paste everything" | False | Context rot, "Lost in the Middle", NoLiMa: accuracy falls as context grows. |
| "Maximum reasoning = best answer" | False | Overthinking and inverse-scaling research show waste and sometimes worse results. |
| "Caching makes long context free" | Half true | ~90% cheaper per token, but no fix for quality loss, and uncached tails still bill fully. |
| "Hebrew costs 4× the tokens" | Outdated | True for the GPT-4-era tokenizer. About 1.3–1.9× on current ones (limited data). |
| "Avoid peak hours on Claude Code" | Outdated | Peak-hour throttling was removed for Pro/Max on 6 May 2026. |
| "Switch model when you hit the limit" | Only sometimes | Claude session and weekly limits are shared across models. Switching helps only after a model-specific limit. |
Relevant when you build apps (like a logging app or a community app). Skip it if you only use the chat apps.
| Lever | Saving | How |
|---|---|---|
| Prompt caching | ~90% off repeated input | Static content first (tools → system prompt → reference docs), dynamic content last. No timestamps or IDs in the prefix. Anthropic: explicit breakpoints, 5-min or 1-h TTL. OpenAI: automatic. Gemini: implicit plus explicit caches. |
| Batch API | 50% | Anything not needed in real time: evaluations, bulk tagging, overnight jobs. Anthropic, OpenAI and Gemini all offer it. |
| Flex tier | ~50% | OpenAI and Gemini "Flex": slower and interruptible, batch-level price. |
| Routing / cascade | up to 85–98% (research) | Cheap model first, escalate on low confidence (FrugalGPT, RouteLLM, OpenRouter Auto Router). |
| Count before you send | — | Anthropic count_tokens (free), tiktoken, Gemini Count Tokens. Tokenizers change between versions: Claude 4.7+ uses ~1.0–1.35× more tokens for the same text. |
| Avoid fast / priority modes | skips a 1.8–2× premium | Opus 5.5 fast mode is $8/$40; OpenAI fast is 2×; Gemini Priority is +80%. Use them only for real-time needs. |
| Compression | varies | LLMLingua for long prompts; RTK for shell output (60–90% of that output, not your bill). Measure on your own workload first. |
Small tools you can use today.
Paste at the end of a long chat, then open a new chat with the result.
Summarize this conversation as a handoff note for a fresh chat. Include: 1. Goal 2. Decisions made, and why 3. Constraints and preferences 4. Current state / latest version of the work 5. Open questions and the next step Max 250 words, bullets, no commentary.
For lookups and routine tasks, not for hard reasoning.
Answer in at most 150 words. Use bullets only if you are listing items. No preamble, no recap of my question, no closing offer.
For drafts that will end up in Hebrew.
Work in English for any outlines, notes or intermediate drafts. Give me only the final version in Hebrew, once, without an English copy.
Most "token-saving hacks" online are about shaving words. The measured gains there are single digits. The real money is in three decisions you make before typing: which chat, which model and which effort.