Claude Opus 5.5 and Sonnet 5.5: Cheaper Data Prompts

Claude Opus 5.5 costs $4 / $20 per 1M input and output tokens and Claude Sonnet 5.5 costs $2 / $10, prices as of October 2026. Opus 5.5 cache hits cost only $0.20 per 1M tokens. For data-heavy prompts, the cheapest setup is to send tables as TOON, put them in a cached prefix, and ask for JSON back with structured outputs.

What do Claude Opus 5.5 and Sonnet 5.5 cost?

These are standard prices in USD per 1M tokens from Anthropic's pricing page, checked on October 1, 2026.

Opus 5.5 is cheaper than the Opus 5 it replaces, and its cache hits are priced at 0.05x input instead of the usual 0.1x. Sonnet 5.5 matches Sonnet 5, whose $2 / $10 launch price became permanent. Anthropic's footnote says the planned rise to $3 / $15 "will not occur".

What else changed in Claude 5.5 for data prompts?

• Model IDs: claude-opus-5-5 and claude-sonnet-5-5. Both have a 1M-token context window and 128K max output, per the models overview, which tells developers to start with Opus 5.5 "for most workloads".
• Long context at flat price: Claude 4.6 and later bill the full 1M window at standard rates, so a 900K-token request costs the same per token as a 9K one.
• Smaller tool overhead: the hidden tool-use system prompt is 286 tokens on Opus 5.5 and Sonnet 5.5, down from 675 on Opus 4.7.
• A hungrier tokenizer: Claude 4.7 and later use a tokenizer that "produces approximately 30% more tokens for the same text". The models page puts 1M tokens at roughly 555K words, down from about 750K on older models.

That last point is why data format matters more on Claude 5.5 than the price table suggests. Our post on Claude's new tokenizer covers how to measure the difference on your own prompts.

How should you format tabular context for Claude 5.5?

Pretty-printed JSON repeats every key on every record and spends tokens on braces, quotes and indentation. TOON writes the keys once in a header. Here are three orders as JSON:

{
  "orders": [
    {
      "id": "ord_1042",
      "customer": "Acme GmbH",
      "country": "DE",
      "status": "shipped",
      "total": 1249.5
    },
    {
      "id": "ord_1043",
      "customer": "Northwind",
      "country": "US",
      "status": "pending",
      "total": 310
    },
    {
      "id": "ord_1044",
      "customer": "Globex",
      "country": "UK",
      "status": "refunded",
      "total": 89.99
    }
  ]
}

And the same data encoded with @toon-format/toon 4.1.1:

orders[3]{id,customer,country,status,total}:
  ord_1042,Acme GmbH,DE,shipped,1249.5
  ord_1043,Northwind,US,pending,310
  ord_1044,Globex,UK,refunded,89.99

The [3] length marker lets the model check that it saw every row. TOON v4 also handles records with a small nested object through nested field groups, so a support-ticket table with an embedded customer stays tabular:

tickets[2]{id,customer{name,plan},priority,ageHours}:
  501,Ada,pro,high,3
  502,Lin,free,low,27

See TOON nested field groups for the rules. On the current official benchmark (spec v4.1, 5,856 calls), TOON used 42.6% fewer tokens than JSON with 72.2% accuracy versus 71.4%. The benchmark counts tokens with OpenAI's o200k_base tokenizer, not Claude's, so measure your own savings with Claude's token-counting endpoint.

How do TOON and prompt caching stack on Claude?

They multiply. TOON shrinks the block, and caching makes every repeat of that block cheap. Put the stable data first, mark it with cache_control, and keep the question at the end:

import Anthropic from "@anthropic-ai/sdk";
import { encode } from "@toon-format/toon";

const client = new Anthropic();
const context = encode({ orders }); // same input always gives the same bytes

const response = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 2048,
  system: [
    {
      type: "text",
      text: "The order data below is TOON: the header lists the fields once, then one record per line.",
    },
    { type: "text", text: context, cache_control: { type: "ephemeral" } },
  ],
  messages: [{ role: "user", content: "Which orders are refunded, and what is their combined total?" }],
});

console.log(response.usage); // check cache_creation_input_tokens and cache_read_input_tokens

Here is the arithmetic for a 200K-token JSON dataset on Opus 5.5. Writing it to the 5-minute cache costs $1.00, and each later read costs $0.04 instead of $0.80 uncached. If TOON cuts the dataset by about 40% to 120K tokens, the write drops to $0.60 and each read to $0.024. On a busy agent loop that reads the same context hundreds of times an hour, both savings add up. To estimate your own reduction instead of assuming 40%, paste a sample into the TOON token calculator. Encoding has to be deterministic, because a single changed byte in the prefix is a cache miss. Our TOON and prompt caching guide covers ordering and key stability.

Should Claude output TOON?

No. Use TOON for input and JSON for output. Claude structured outputs are generally available: you pass a JSON Schema through output_config.format, or set strict: true on a tool, and the response is guaranteed to match. There are limits. Recursive schemas and numeric or string-length constraints are not supported, and additionalProperties must be false. Within those limits, you get validated JSON with no parsing retries, which matters more than a few saved output tokens.

When is TOON not worth it on Claude?

• Small payloads. A single config object gains little, and the format note costs tokens.
• Deeply nested or irregular data. The TOON README says compact JSON often wins here.
• Long multi-turn agents. arXiv 2605.29676 found TOON saved up to 18% of tokens at about a 9-point accuracy cost on open-weight models in agentic benchmarks, and called it "not safe as a default in multi-turn agentic systems".
• Haiku 4.5 workloads with a deadline. Haiku 4.5 is the Claude model in the TOON benchmark (65.6% TOON vs 63.5% JSON), but its retirement is listed as "not sooner than October 15, 2026". Plan your migration tests on Sonnet 5.5.

For a broader Claude integration walkthrough, read TOON with Anthropic Claude, and compare these prices with other providers in LLM API pricing for October 2026. To try the format on your own data, paste it into the JSON to TOON converter.

Frequently Asked Questions