DEV Community: Gary Walkers

Prompting techniques and their token cost: system prompts, few-shot, chain-of-thought and injection

Gary Walkers — Sat, 14 Nov 2026 00:17:31 +0000

Every word in a prompt is billed, so each technique has a token cost. Here is what the vendors and papers say, with counts I measured using OpenAI's o200k_base tokenizer.

System prompts are billed on every request

My 225-word support-bot prompt is 275 tokens. At 1,000 requests a day that is 8.25M input tokens a month: $16.50 on Claude Sonnet 5.5 or GPT-6.1 Sol, $0.825 on Haiku 5.5. It will not cache on its own: Anthropic's minimum cacheable prompt for Sonnet 5.5 is 512 tokens and OpenAI's is 1,024 (Anthropic prompt caching, OpenAI prompt caching). To count your own prompts, a browser token counter works without uploading anything.

Few-shot: about 18 tokens per example

In a five-category email classifier, zero-shot is 44 tokens and each example adds about 18, so 10 examples add 181 tokens to every request, or $36.20 per 100,000 requests on Sonnet 5.5. Anthropic recommends 3 to 5 diverse examples (prompting best practices); the original result is Brown et al. (2020). Measure accuracy at 0, 3 and 5 examples before paying for more.

Chain-of-thought is paid as output

Wei et al. (2022) and Kojima et al. (2022) showed step-by-step reasoning helps 2022-era models: MultiArith went from 17.7% to 78.7% with "Let's think step by step" on text-davinci-002. Reasoning models now think internally, and OpenAI's guide says to give them high-level goals. The visible reasoning is output tokens: 1 token for a direct answer versus 58 for a worked one in my example, or $0.01 versus $0.58 per 1,000 answers on Sonnet 5.5. Skip it for classification and extraction.

Prompt injection: design for it

OWASP LLM01 lists prompt injection as the top risk. Indirect injection, hidden instructions in a page or file the model reads, is the harder case, and OWASP says no fool-proof prevention is known. Limit privileges, validate output with deterministic code, require human approval for risky actions, label untrusted content, and test adversarially.

A one-page summary of all of this is in this prompt engineering guide.

Open weight is not open source: how to read an LLM licence, and how Claude, GPT-6 and Gemini 4 compare

Gary Walkers — Sat, 14 Nov 2026 00:09:59 +0000

"Which model should I use?" is easier once you know what a model is and what the labels mean. Prices below were verified against each provider's own pages on 13 November 2026.

What an LLM is

An LLM turns text into tokens, passes them through a neural network and outputs a probability for every possible next token; generating a reply means choosing one, appending it and repeating. Four papers explain the field: Attention Is All You Need (the Transformer, 2017), Scaling Laws for Neural Language Models (2020), Language Models are Few-Shot Learners (2020) and InstructGPT (2022).

Why attention shapes your bill

The Transformer paper's own complexity comparison lists self-attention as costing operations proportional to the square of the sequence length per layer. That is one reason context windows have limits and long prompts are expensive to serve.

Claude vs ChatGPT vs Gemini through their APIs

USD per million tokens, input / output (tier placement is ours): flagship Claude Fable 5.2 $8 / $40, GPT-6 Astra $10 / $50, Gemini 4 Carbon $2 / $10; smallest tier Claude Haiku 5.5 and GPT-6 Luna both $0.10 / $0.50. Gemini 4 Carbon and Argon double on 1 December 2026. Current Claude models list a 1M-token window, GPT-6 1.05M, Gemini 4 1M (Anthropic, OpenAI, Google). For 1,000 uncached chat turns of 1,500 input and 400 output tokens: Sonnet 5.5 $7.00, GPT-6.1 Terra $3.50, Gemini 4 Hydrogen $5.25. A side-by-side with caching and batch rules is in this Claude vs ChatGPT vs Gemini comparison.

Open weight vs open source

Open weight means the parameters are published. Under the Open Source Initiative's Open Source AI Definition 1.0, open source AI also needs sufficiently detailed training-data information and the complete training and running code, all under OSI-approved terms. Publisher examples: gpt-oss is Apache 2.0 and Whisper is MIT. Before building on a model, check the licence for weights and code, commercial limits, acceptable-use rules and hosting cost. A checklist is in this guide to open source vs open weight LLMs.

What an LLM API call really costs: the formula and four levers that move the bill

Gary Walkers — Sat, 14 Nov 2026 00:09:31 +0000

Prices are quoted per million tokens, but the bill depends on how many tokens you send, how many you get back, and which pricing rules apply. Figures below were checked against the providers' own pages on 13 November 2026.

The formula

cost = input_tokens / 1e6 * input_price + output_tokens / 1e6 * output_price, times requests per day, times 30. A support bot sending 1,500 input and 400 output tokens, 1,000 times a day, on Claude Sonnet 5.5 ($2 in, $10 out per million) costs $210.00 a month; GPT-6.1 Sol lists the same prices (Anthropic pricing, OpenAI pricing). You can reproduce it in an LLM API cost calculator.

Lever 1: output costs more

Across 65 models in our price table, output is 1.0x to 10.0x the input price, with a median of 5.0x. On Sonnet 5.5 a chat turn is about 21% output by tokens but 57% of the cost, so trimming replies usually saves more than trimming prompts.

Lever 2: caching

Cached reads cost 0.1x the input price on most Claude models and on GPT-5.6 and later, and 0.05x on models such as Sonnet 5.5 and GPT-6.1 Sol; writing a cache entry costs 1.25x (Anthropic prompt caching). For a 10,000-token fixed prompt at 1,000 requests a day, our worked example falls from $690.00 to $120.72 a month. Minimum prompt lengths apply, so short system prompts will not cache.

Lever 3: batch

Anthropic, OpenAI and Google charge 50% less for jobs that can wait up to a day (OpenAI Batch API). The chatbot above falls from $210 to $105 a month.

Lever 4: very long prompts

OpenAI steps up above 272K input tokens, Gemini 3.1 Pro and Grok above 200K, Claude Haiku 5.5 above 100K. A 300K-token prompt costs about $0.61 on Sonnet 5.5 and $6.08 on GPT-6 Astra (our arithmetic from the published prices).

Compare tiers, not brands

On three reference workloads, Sonnet 5.5 and GPT-6.1 Sol cost the same; the small tier (Haiku 5.5 or GPT-6 Luna) costs $10.50, $9.00 and $8.52 a month, and flagships run 2 to 3 times the mid tier. Run the same 50 to 100 real prompts through each candidate and divide cost by passing answers. The full table is in this OpenAI API pricing breakdown alongside the other providers.

Counting tokens exactly, and why the same sentence costs more in Hindi

Gary Walkers — Sat, 14 Nov 2026 00:03:48 +0000

Every LLM price, context limit and output cap is measured in tokens, so it pays to count them properly. Here is the short version, with the numbers I measured.

What a token is

A token is a chunk of text: a short word, part of a longer word, a few digits or a symbol. OpenAI's rule of thumb is about 4 characters or 0.75 words per token in English (OpenAI API concepts). With tiktoken's o200k_base, "tokenization" is 2 tokens, "2026" splits into 202 + 6, and the German compound "Straßenbahnhaltestelle" splits into 7. The vocabulary comes from byte-pair encoding, applied to subwords as in Sennrich, Haddow and Birch (2015).

Count exactly (OpenAI)

import tiktoken
enc = tiktoken.get_encoding("o200k_base")
print(len(enc.encode("Hello world")))  # 2

tiktoken is OpenAI's open-source tokenizer. If you would rather not run code, this browser token counter reproduces it; I compared it with tiktoken on more than 1,600 strings in both encodings and the ids matched every time. Nothing you paste is uploaded.

Claude and Gemini do not publish downloadable tokenizers, so third-party counts for them are estimates; use each provider's token-counting endpoint for billing. Anthropic says its newer tokenizer (Claude 4.7 and later) produces roughly 30% more tokens for the same text than earlier Claude models (Anthropic models overview).

Words and tokens

About 750 English words per 1,000 tokens. Code, JSON, long numbers and URLs split into many small tokens, so fewer words fit. A tokens to words converter gives the quick table from 1,000 to 1M tokens.

Languages are not equal

Two sentences in 15 languages with OpenAI's exact tokenizers: English 31 tokens, German 37, Russian 43, Hindi 47 on o200k_base but 153 on the older cl100k_base; Chinese used 25. It is a small sample with my own translations, so treat it as direction, not a benchmark. Count a sample of your own text before pricing a non-English workload.

Context windows

As of 13 November 2026, the current Claude 5 models list 1M tokens and 128K max output, GPT-6 lists 1.05M, and Gemini 4 and DeepSeek V4 list 1M, per each provider's model page. A bigger window is not free: you pay for every input token on every request.