Token Field Guide
Field guide · researched 23 Sep 2026 · docs, papers, Reddit, GitHub, HN

Spend tokens where thinking pays.

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.

Official vendor docs / announcement Community user-reported, plausible Unverified single source / shaky Myth debunked or outdated

The 8 rules that do most of the work

Ranked by how much they save. The top three are behavioral and dwarf every clever prompt trick on the internet.

1

One topic per chat. Restart with a handoff note.

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.

Biggest leverOfficial
2

Match model and effort to the job

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×.

Biggest leverOfficial
3

Control the output, not just the input

Output costs ~5× input, and thinking is billed as output. Set length, format and language. Ask for Hebrew only for the final text.

HighOfficial
4

One complete prompt beats five corrections

Same task spread over several turns scored 39% worse on average (200k+ simulated chats). Edit your last message instead of stacking "no, I meant…".

HighResearch
5

Send less: excerpts, text, small images

A PDF page costs 1.5–3k text tokens plus an image. Paste the relevant section, convert PDFs to text, crop screenshots.

MediumOfficial
6

Turn off what the task doesn't need

Web search, connectors, deep research, extended thinking, agent mode. Each adds tokens or whole extra loops.

MediumCommunity
7

Reusable material lives in Projects — lean

Project files are cached and count less against limits. But instruction files (CLAUDE.md, AGENTS.md) load every turn: keep them short.

MediumOfficial
8

Don't break the cache mid-session

In coding agents, pick model and effort at the start. Switching, or returning after the cache expired, re-processes everything at full price.

Medium · agentsOfficial

Where your tokens actually go

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.

A 20-turn chat, 1,000 new tokens per turn
Cumulative input tokens processed
What you typed (20k)What was processed (210k)

Caching makes the re-read cheaper (often ~10% of price) but never free — and it does nothing for the quality loss of long context.

~5×

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.

43k

Agents carry baggage. Saying "Shalom" (9 tokens) in Claude Code sent 43,483 input tokens — system prompt, tools, memory files. Mostly cached, still counted.

1.3k

One 1000×1000 image ≈ 1,296 tokens on Claude; a full-HD screenshot ≈ 2,700. Pasted text is usually cheaper and read more reliably.

−39%

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.

Which model, at which effort, for which job

Default to the middle. Go down for anything mechanical, up only when a wrong answer is expensive or the reasoning is genuinely multi-step.

TaskTierEffortIn practice
Quick lookup, definition, rephrase a sentencetranslation of a line, a formula, a fact checkSmall / fastLow / offHaiku 4.5 · ChatGPT Instant · Gemini Fast
Everyday drafting and summarizingemails, notes, summarizing one articleMidLowSonnet 5 · ChatGPT Instant · Gemini 3.8 Flash
Academic writing and synthesislit-review section, structuring a thesis chapter, feedback on a draftMid → TopMediumSonnet 5 or Opus 5.5 at medium · ChatGPT Medium
Statistics, math, debuggingchoosing a test, interpreting output, a bug you can't findTopHighOpus 5.5 high · ChatGPT High · Gemini Pro
High-stakes analysislong case material, decisions with real consequencesTopHighOpus 5.5 · Fable 5.1 or GPT-6 Pro only if top tier fails
Building an app or agentic codingmulti-file, multi-hourTop plans, Mid executesMediumClaude Code: Opus 5.5 medium (default), or opusplan; subagents on Haiku
Research across many sourcesMode choiceMax (rationed)Deep Research only when the question deserves 20+ sources; otherwise normal search
Q&A over a big pile of documentsDifferent toolLowGemini Notebook (ex-NotebookLM) or a Claude Project — not pasting
Bulk extraction / classification (API)CheapestNoneGPT-6 Luna · Gemini Flash-Lite · Haiku, via Batch (−50%)

The effort ladder

Low

Lookups, rewriting, formatting, classification, simple Q&A. The default you should reach for most.

Medium

Writing that needs structure, planning a document, moderate coding. Anthropic's default for Opus 5.5.

High

Multi-step math, statistics, hard debugging, analysis where you'd check the answer anyway.

Extra high / Max

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.

More thinking isn't always better. On "2+3=?" a reasoning model spent 1,953% more tokens than a normal model, going through 13 rounds of checking, and its first round was already right in 92% of cases. Anthropic's own "inverse scaling" study found longer reasoning made models worse on some tasks. Adding "think step by step" gives reasoning models ~3% or less.

Current line-ups, by tier

TierAnthropicOpenAIGoogle
SmallHaiku 4.5$1 / $5GPT-6 Luna$0.10 / $0.50Gemini 3.5 Flash-Lite$0.30 / $2.50
MidSonnet 5$2 / $10GPT-6 Sol$2 / $10Gemini 3.8 Flash$0.75 / $3.75 intro, $1.50 / $7.50 from 2027
TopOpus 5.5 recommended default$4 / $20GPT-6 AstrapremiumGemini 3.1 Pro$2 / $12 (third-party figure)
UltraFable 5.1$10 / $50GPT-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 Hebrew token tax

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.

Tokens for the same text, English = 1.0×
Hebrew relative to English
Englishbaseline
1.00×
GPT (o200k)ChatGPT, current
1.34×
Geminicurrent
1.66×
Claudecurrent
1.90×
GPT-4 era (cl100k)old, full corpus
3.67×

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.

What to do about it

  1. Long reusable material in English: project instructions, templates, system prompts, reference text. It's re-read every turn, so the savings compound.
  2. Hebrew output only for the final text. Output costs ~5×, so the output language matters more than the prompt language.
  3. No round-trips. Don't ask for Hebrew, then an English version, then Hebrew again. Every pass is billed.
  4. Short conversational questions? Just write Hebrew. The saving is tiny and misunderstandings cost more.
  5. Drop niqqud unless you need it. Vowel marks are extra characters. inference
Quality in well-resourced languages is close to English: Claude scores ~92–97% of its English score in Arabic. The Hebrew issue is cost and limits, not intelligence.

Platform by platform

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.

Metered byOne shared pool across chat, Cowork and Claude Code: a 5-hour session limit plus a weekly limit. Cost per message grows with message length, attachments, conversation length, model and features. A single Cowork task can use far more than a chat. Check Settings → Usage.

Do

  • Put documents you reuse in a Project. They're cached and count less against limits. Official
  • Batch related questions into one message, and edit the message instead of sending corrections. Official
  • Use Low/Medium effort for routine work. On Opus 5.5 thinking can't be turned off, so lower the effort. Official
  • Sonnet for everyday work, Opus for complex reasoning, Haiku for lookups. Official

Avoid

  • Marathon chats that mix topics. Each new message pays for the whole history.
  • Uploading a PDF when text would do. Each page goes in as text and an image. Official
  • Leaving web search, research, connectors and memory on for tasks that don't need them. Community
  • Using Cowork for something a plain chat answers.

Tricks

  • Anchor the 5-hour window: send a tiny Haiku message early so the reset lands mid-workday. It only moves the reset; it doesn't add usage. Community
  • Switching models after a session/weekly limit doesn't help, because the pool is shared. It only helps after a model-specific "Opus limit" message. Official
  • Stable facts about you go in Preferences once, not re-explained in every chat. Community

Myths vs evidence

Popular advice, checked against what's actually been measured.

ClaimVerdictWhat the evidence says
"Caveman mode" / terse-speak saves 65–75%OverstatedJetBrains measured 8.5% fewer output tokens over 86 tasks. The caveman skill itself adds ~1k input tokens per call.
"Saying please costs millions — stop"FolkloreAltman's remark was a joke, not a calculation. "Please" is about one token.
"Being rude or polite changes accuracy"MixedOne 50-question study found rude prompts slightly better; earlier studies found the opposite. Not a reliable lever.
"CAPS / CRITICAL / MUST improves compliance"BackfiresAnthropic: 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"PartlyCuts output ~49%, but lost ~28% on math for older models. Use it for lookups and formatting, not reasoning.
"Big context window → paste everything"FalseContext rot, "Lost in the Middle", NoLiMa: accuracy falls as context grows.
"Maximum reasoning = best answer"FalseOverthinking 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"OutdatedTrue for the GPT-4-era tokenizer. About 1.3–1.9× on current ones (limited data).
"Avoid peak hours on Claude Code"OutdatedPeak-hour throttling was removed for Pro/Max on 6 May 2026.
"Switch model when you hit the limit"Only sometimesClaude session and weekly limits are shared across models. Switching helps only after a model-specific limit.

If you build with the API

Relevant when you build apps (like a logging app or a community app). Skip it if you only use the chat apps.

LeverSavingHow
Prompt caching~90% off repeated inputStatic 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 API50%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 / cascadeup 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 modesskips a 1.8–2× premiumOpus 5.5 fast mode is $8/$40; OpenAI fast is 2×; Gemini Priority is +80%. Use them only for real-time needs.
CompressionvariesLLMLingua for long prompts; RTK for shell output (60–90% of that output, not your bill). Measure on your own workload first.

Copy-paste kit

Small tools you can use today.

Handoff note

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.

Output control

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.

Hebrew, efficiently

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.

Start a new chat when…

  • the topic changed
  • the model took a wrong turn twice (research: it rarely recovers)
  • you're repeating context it should already have
  • answers are getting slower, vaguer or contradicting earlier ones
  • you're coming back after hours to a very long thread

Before you hit send

  • Is this the cheapest model that can do it?
  • Does it need thinking or search at all?
  • Did I give all the context in this one message?
  • Am I sending an excerpt, or the whole file?
  • Did I say how long, and in what format?

My professional read

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.

  • Your largest leak is almost certainly conversation length. If you run long, multi-layered threads, and deep reflective or analytical work invites exactly that, you pay for the whole history on every message and get worse answers late in the thread. A handoff note every 15–25 turns is worth more than everything else on this page combined.
  • Hebrew matters, selectively. English for instructions and reference material that's read repeatedly; Hebrew for what you'll actually send or submit.
  • Cowork and agents are the expensive surface. Use them when you need files, tools or autonomy, not for a question a chat can answer.
  • Treat High effort as a tool, not a setting. Default to Medium, raise it for statistics and hard analysis, then lower it again.
Sources (click to expand)