Capability scores come from the Epoch Capabilities Index (ECI) by Epoch AI, which combines dozens of benchmarks (math, coding, science, reasoning) into one scale. Prices are standard API list prices, checked October 4, 2026.

Models on the green line are the best-value frontier: no other model is both cheaper and more capable. Anything below the line costs more than a frontier model with the same or higher score.

The short version

A chat request is 1,500 input + 400 output tokens of English text. Costs include each model's tokenizer difference: Claude counts about 30% more tokens than GPT for the same text, and that is already in the numbers.

Three things that surprised me

1. The #1 model is not the most expensive. Claude Opus 5.5 (167.3) edges out GPT-6 Astra (166.5), and the two are within each other's margin of error. But Astra lists at \(10 / \)50 per million tokens against Opus's \(4 / \)20, so 10,000 chats cost $350 on Astra and $182 on Opus.

2. "Good enough" is very close to the top. Claude Sonnet 5.5 is 2.2 points under Opus for half the price. Gemini 3.8 Flash and DeepSeek V4 Flash sit 10–13 points below the top at 7–20× lower cost, which is plenty for classification, extraction, summaries and most chatbot traffic.

3. Some popular picks are off the line. Claude Fable 5.1 (\(455 per 10k chats) and GPT-5.5 (\)195) both score below Sonnet 5.5 ($91). Gemini 3.1 Pro ($74) scores below Gemini 3.8 Flash ($25). They may still win on a specific task, but if you use them by default, test the cheaper option first.

What about GPT-6 Sol?

Epoch hasn't scored GPT-6 Sol yet. On a different scale, the Artificial Analysis Intelligence Index (v4.3.2, max reasoning, Sept 30), GPT-6.1 Sol scores 51.8 vs GPT-6 Astra's 52.7, at a fifth of the price. If that holds on ECI, Sol would sit on the frontier next to Claude Sonnet 5.5, which has the same list price. It isn't on the chart because the scales differ.

Live version

The chart and table rebuild every day from Epoch's data and current list prices. You can switch between top capability, best value and cheapest views here:

👉 AI model capability vs price ranking (free, no sign-up)

Full write-up with the "expensive for what you get" table: Best value LLM, October 2026.

Capability scores: Epoch Capabilities Index by Epoch AI, used under CC BY 4.0. Prices: API list prices, no caching or batch discounts.

Short version: row-by-row JSON, pretty-printed, uses about 3× the tokens of CSV. The same data as columnar JSON costs about the same as CSV. It's the repeated keys, not JSON itself.

Update: a reader pointed out I'd left out columnar JSON (one object, one array per field). I measured it on the same table and tokenizer and added it below.

The setup

A product table: 20 rows, 5 fields (id, name, price, in_stock, category). One row in CSV:

1001,Wireless Mouse,9.99,false,electronics

Same data, seven formats, counted with o200k_base, the tokenizer behind GPT-4o and later OpenAI models:

import { getEncoding } from "js-tiktoken";
const enc = getEncoding("o200k_base");
const count = (s) => enc.encode(s).length;

count(csv);                            // 300
count(JSON.stringify(rows));           // 525
count(JSON.stringify(rows, null, 2));  // 884

Results

The older cl100k_base tokenizer gave nearly identical numbers, within 2% for every format.

Why the gap is so big

1. Keys repeat on every row. JSON, YAML and XML write "name":, "price": twenty times. CSV writes them once, in the header, and so does columnar JSON, which is why it lands right next to CSV.
2. Punctuation is tokens. Quotes, braces and colons all take tokens. XML is worst: every value gets an opening and a closing tag.
3. Indentation is for humans. The model doesn't need it. Pretty-printing alone added 359 tokens (+68%) over minified JSON here.

YAML is the odd one: fewer characters than minified JSON, more tokens, because every field gets its own line and its own key.

I measured some code too

Coding agents send a lot of code, so I tried a few things:

• Indentation is nearly free. Modern tokenizers merge runs of spaces into single tokens. Don't reformat code to save tokens.
• Minified code is half the tokens, but don't. You lose names like subtotal and taxRate, which is exactly what helps the model understand the code.
• UUIDs are expensive. If your prompt has a column of long IDs, swap them for row numbers and map back in code.

What I use now

• Flat tables as input: CSV or TSV. TSV if values contain commas, so no quoting.
• If it has to be JSON: go columnar ({"id":[...],"name":[...]}) or header + row arrays. Same data, roughly CSV's cost.
• Data the model returns: minified JSON with a schema. Parse reliability beats a few saved tokens; use structured output if your API supports it.
• Nested data: minified JSON. Flattening it into CSV by hand often costs more than it saves.
• Tables a human will also read: Markdown, only 24% over CSV.
• Avoid pretty JSON and XML in prompts unless a tool requires them.

Dropping unused fields, null fields and extra decimal places helps too.

What it costs

Say you attach a 20-row table to every request, 1,000 requests a day, 30,000 a month, at $2 per million input tokens:

Invisible per request, real for a product, and it scales with bigger tables, RAG results and long API responses.

Caveats

• One table, one tokenizer. Claude and Gemini tokenize differently, so absolute numbers will differ. The ranking (repeated keys and tags cost the most) comes from how the formats are built, so it should hold.
• If the format changes answer quality for your task, test both. A cheaper prompt with worse answers isn't cheaper.

Try it on your own data

I built a free token counter for this. It runs the same o200k tokenizer in your browser, so nothing you paste is uploaded. Paste your data in two formats and compare tokens and cost per model.

Full write-up with more detail: JSON vs YAML vs CSV: which format uses the fewest tokens?

The plans side by side

US prices. OpenAI's help page only says that Pro 200 includes more usage than Pro 100 and Pro 500 the most; the 5×, 10× and 25× figures are the multipliers reported from OpenAI's announcement.

All three Pro plans include the same features: Pro models, Codex, deep research, image creation, memory and file uploads. The only feature difference is Ultrafast, a faster mode for GPT-6 Astra, which is Pro 500 only. Buying extra credits on Pro 100 or Pro 200 does not unlock it.

The part most people miss: there is no bulk discount any more

Divide each price by the usage you get and every plan lands on the same number: $20 per Plus-worth of usage. Pro 500 is not a better deal than Pro 100; it is just more of the same thing, plus speed.

That used to be different. Until this change Pro 200 gave 20× Plus usage, which worked out to $10 per unit, half the price of everything else. New Pro 200 subscribers now get 10×. If you were already on Pro 200, you keep the old 20× allowance until October 29, 2026, then drop to 10× at the same $200.

So the rule is simple: buy the smallest plan you don't hit the limit on. Paying for headroom you never use is the only way to overpay.

Which one should you pick?

• You rarely hit the Plus limit: stay on Plus ($20). None of the Pro plans give you a smarter answer for everyday chat; they give you more of it.

• You hit the Plus limit a few times a week: Pro 100. Five times the usage for five times the price, and the jump from $20 to $100 is the smallest step up.

• You regularly run out on Pro 100: Pro 200. Same price per unit, twice the room.

• You run Codex or agents most of the day, or waiting on output costs you money: Pro 500. It is the only plan with Ultrafast, but note that faster generation uses your allowance faster too.

• You're on the old Pro 200: keep it until October 29; it is the best deal OpenAI sells right now. After that, look at how much you actually used. If you stayed under about a quarter of your old allowance, Pro 100 does the same job for $100 less.

What about just using the API?

If you mostly send short or medium messages, paying per token is often much cheaper than any Pro plan. Here is roughly what a month of chat costs through the API, assuming conversations of 6 messages and English text:

"Normal" = about 150 tokens in and 500 out per message; "long" = about 2,000 in and 700 out. API prices: Sol \(2 / \)10 and Astra \(10 / \)50 per million input / output tokens.

Two things push the numbers up fast. First, every new message re-sends the whole conversation, so long chats cost far more than short ones. Second, languages other than English need more tokens for the same text: Korean about 1.4×, Japanese about 1.8×, so the API bill grows by the same factor.

The takeaway: light and medium users of the everyday model are usually better off on Plus or the API. Pro starts to pay off when you use the top models heavily, work with long documents, or live in Codex.

Check your own numbers

Everyone's usage is different. The Subscription vs API calculator lets you enter how many messages you send, how long they are and which language you write in, and shows what the same month would cost on the API next to ChatGPT, Claude and Gemini plans.

Prices as of October 1, 2026. OpenAI may change allowances again; check chatgpt.com/pricing before you buy.

Originally published on tokensave.app.

Short version: Indian languages got dramatically cheaper. Malayalam went from 10.09× the tokens of English to 1.85×. But the gap is not gone: Punjabi still needs 2.44× the tokens of English, more than any of the 34 languages I measured before.

The prompt

English (34 tokens):

Please summarize the customer email below in three bullet points and suggest a polite reply. The customer says the order arrived two days late and one item was missing from the box.

Hindi version:

नीचे दिए गए ग्राहक के ईमेल को तीन बिंदुओं में सारांशित करें और एक विनम्र जवाब सुझाएँ। ग्राहक का कहना है कि ऑर्डर दो दिन देर से पहुँचा और डिब्बे में एक सामान कम था।

import { getEncoding } from "js-tiktoken";

const now = getEncoding("o200k_base");   // GPT-4o and later
const old = getEncoding("cl100k_base");  // GPT-4 / GPT-3.5

now.encode(hindi).length; // 51
old.encode(hindi).length; // 156

Results

For comparison, Simplified Chinese is 1.03×, Korean 1.44×, Japanese 1.79× and Greek 2.06×. The full 34-language table is here: token cost by language.

What stood out

1. The old tokenizer was brutal for South Indian languages. On cl100k, Malayalam needed 10×, Kannada and Telugu over 9×, Tamil 8.5×. Many characters fell back to raw bytes, so a single letter could cost several tokens.

2. Hindi is now cheaper than Japanese. Hindi dropped from 4.59× to 1.50×. Japanese, which was cheaper than Hindi on the old tokenizer, is now 1.79×.

3. Script matters more than the language family. Hindi and Marathi both use Devanagari and sit at 1.50–1.65×. Punjabi (Gurmukhi) is the outlier at 2.44×, even though it is closely related to Hindi.

What it costs

With a model at $2 per million input tokens, sending this prompt one million times costs:

That is input only. If the model also replies in the same language, the same multiplier applies to output tokens, which usually cost 4–5× more.

How to spend fewer tokens

• Write the system prompt and fixed instructions in English. Keep only the user's message in their language. That part is repeated on every request, so it matters most.
• Ask for intermediate steps in English or JSON (classification, extraction, tool calls) and only the final answer in Hindi, Tamil or whichever language your user reads.
• Cache the fixed part of the prompt (prompt caching). Cached input is often 90% cheaper.

Limitations

• One prompt only. With other text the ratios move by about ±0.1–0.2.
• Translations started from machine translation and were checked, but not by native speakers of every language. If a translation looks off to you, tell me and I will re-measure.
• Claude and Gemini use different tokenizers, so these numbers apply to OpenAI models.

Try it on your own text

I built a free counter that runs this measurement in your browser (nothing is uploaded). It shows how many times more tokens your text uses than English, and one click can clean spaces, translate to English on-device (Chrome) and trim filler: TokenSave in Hindi · in English.

Which Indian language do you use in your prompts, and do your numbers match mine?