LLM API Pricing October 2026: GPT-6, Claude 5.5, Gemini

As of October 2026, frontier LLM APIs cost between $0.10 and $10 per 1M input tokens and between $0.50 and $50 per 1M output tokens. GPT-5.6 Sol and Claude Opus 5.5 both cost $4 / $20, Claude Sonnet 5.5 costs $2 / $10, and Gemini 3.8 Flash costs $0.75 / $3.75 until its price doubles on January 1, 2027.

How much do LLM APIs cost per 1M tokens in October 2026?

The table below lists standard-tier list prices in USD per 1M tokens, prices as of October 2026, taken from each provider's official pricing page on October 1: OpenAI, Anthropic, Google, DeepSeek and xAI. Cached input is what you pay when a prompt prefix is served from cache.

Two notes on the table. The Anthropic cached-input column shows cache hits; cache writes cost 1.25x input for a 5-minute cache and 2x for a 1-hour cache. The OpenAI rows show short-context prices, and long context costs more: GPT-5.6 Sol goes to $8 / $30 above 272K tokens.

Is GPT-5.6 still OpenAI's newest model?

No, and this matters for any price comparison you read. GPT-5.6 launched publicly on July 9, 2026 in three tiers: Sol, Terra and Luna. In September OpenAI shipped GPT-6 Astra and then GPT-6 Sol and GPT-6 Luna on September 22, with API prices 50% below GPT-5.6. The OpenAI models page now recommends GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna as the flagship trio.

GPT-5.6 is still listed and priced, so existing integrations keep working. If you are choosing a model today, compare GPT-6.1 Sol at $2 / $10 with GPT-5.6 Sol at $4 / $20 before you commit. Engadget's launch report listed higher GPT-5.6 prices, so use the official pricing page, not launch coverage.

How do batch, cache and off-peak discounts change the math?

List prices are only the starting point. Every major provider now has at least one way to pay less for the same tokens:

• Batch APIs: OpenAI, Anthropic and Google all take 50% off for asynchronous batch jobs. Claude Opus 5.5 drops to $2 / $10, and Gemini 3.8 Flash to $0.375 / $1.875.
• Prompt caching: Anthropic charges 0.1x input for cache hits on most models, 0.05x on Opus 5.5 and 0.025x on Fable 5.1. OpenAI cached input is one tenth of input on GPT-5.6 and GPT-5.5, and one twentieth on GPT-6.1 Sol.
• Off-peak pricing: DeepSeek halves its rates outside peak hours, which are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.
• Context-length tiers: xAI charges double for Grok 4.7 prompts at or above 200K tokens, Google raises Gemini 3.1 Pro above 200K, and OpenAI raises GPT-5.6 above 272K. Anthropic bills the full 1M context at standard rates on Claude 4.6 and later.

For data-heavy workloads, caching is usually the largest lever. If the same product catalog or policy table goes into thousands of requests, put it first in the prompt and keep it byte-for-byte stable. Our guide to TOON and prompt caching explains why a deterministic serializer matters for cache hits.

Are per-token prices comparable across providers?

Not directly. Each provider counts tokens with its own tokenizer, so the same JSON file produces different token counts on each one. Anthropic says the tokenizer used by Claude 4.7 and later models "produces approximately 30% more tokens for the same text". A Claude prompt can therefore cost more than its list price suggests next to a model that counts fewer tokens for the same text.

The practical rule is to compare cost per task, not cost per token. Run 50 to 100 real requests through each candidate, log the usage fields the API returns, and multiply by the prices above. Our post on Claude's new tokenizer shows how to measure this with the token-counting endpoint.

What does a data-heavy workload cost on each model?

Here is a worked example. Take 100,000 requests a month, each with 5,000 input tokens of JSON records and 500 output tokens. That is 500M input tokens and 50M output tokens per month. The TOON column assumes the records shrink by about 40%, to 300M input tokens. It ignores caching, batch and tokenizer differences, so treat it as a rough comparison. To run the same arithmetic on your own payload, use the TOON token calculator, which estimates cost per 1M calls.

Two things stand out. First, output tokens do not shrink, so the savings get smaller as output gets longer. Second, the Gemini 3.8 Flash line doubles in January. We cover that change in Gemini 3.8 Flash doubles in price in 2027.

How much can TOON save on top of these prices?

TOON (Token-Oriented Object Notation) writes the field names of a uniform array once, in a header, and then lists one row per record. On the current official benchmark (spec v4.1, 5,856 LLM calls across four models), TOON used 42.6% fewer tokens than JSON and scored 72.2% accuracy versus 71.4% for JSON. Here is the price table above as JSON and as TOON, produced by @toon-format/toon 4.1.1:

{"models":[
  {"provider":"OpenAI","model":"gpt-5.6-sol","input":4,"cachedInput":0.4,"output":20},
  {"provider":"Anthropic","model":"claude-opus-5-5","input":4,"cachedInput":0.2,"output":20},
  {"provider":"Anthropic","model":"claude-sonnet-5-5","input":2,"cachedInput":0.2,"output":10},
  {"provider":"Google","model":"gemini-3.8-flash","input":0.75,"cachedInput":0.075,"output":3.75},
  {"provider":"DeepSeek","model":"deepseek-flash","input":0.3,"cachedInput":0.006,"output":1.2}
]}

models[5]{provider,model,input,cachedInput,output}:
  OpenAI,gpt-5.6-sol,4,0.4,20
  Anthropic,claude-opus-5-5,4,0.2,20
  Anthropic,claude-sonnet-5-5,2,0.2,10
  Google,gemini-3.8-flash,0.75,0.075,3.75
  DeepSeek,deepseek-flash,0.3,0.006,1.2

The benchmark number is the best case for one part of the prompt. In production, Halodoc measured 5 to 15% cost savings after switching to TOON, because system prompts, conversation history and output tokens stayed the same. Our Halodoc case study write-up breaks down why. The TOON README also says deeply nested data, purely flat tables (where CSV is smaller) and latency-bound local models are cases where TOON may not help.

Which model should you pick for structured-data prompts?

• Lookups and extraction over clean tables: start with a low-cost tier such as GPT-6 Luna, GPT-5.6 Luna, Gemini 3.8 Flash or DeepSeek Flash. Field retrieval is the easiest question type in the benchmark.
• Aggregation and filtering: TOON and JSON both scored below 50% on these questions in the benchmark. Compute sums and filters in code, then send the results, or move up to a Sol, Sonnet or Opus tier.
• Long agent runs with large shared context: caching dominates. Claude Opus 5.5 cache hits at $0.20 and GPT-6.1 Sol cached input at $0.10 make a big stable prefix cheap.
• Overnight jobs: use the Batch API for half price. See OpenAI Batch API with TOON.

Whichever model you pick, measure your own payload. Paste a real API response into the free JSON to TOON converter to see how many tokens it saves before you change your pipeline.

Frequently Asked Questions