What a million tokens costs across frontier models, 18 Oct 2026

Every vendor quotes prices per million tokens, but input and output differ by a factor of five or more, cached input differs again, and some prices change on a date. This post collects the published list prices as of 18 October 2026 into one table and shows one way to compare them.

All prices are USD per million tokens, copied from the vendors' own pages fetched on 18 October 2026: Anthropic's pricing documentation, OpenAI's API pricing page, Google's Gemini API pricing page, and Mistral's pricing page. Prices change often, so use the Sources links to re-check before budgeting.

The price table

Notes on how to read it:

• OpenAI prices apply to short context, up to 272K input tokens. Some models have higher long context rates.
• Anthropic lists cache writes separately: 1.25x base input for a 5 minute cache, and 2x for a 1 hour cache. Cache reads are 0.1x base input, except 0.05x on Opus 5.5 and Sonnet 5.5 and 0.025x on Fable 5.1.
• Google charges an hourly storage fee for cached content on most models, not included above.
• Mistral's page lists one example price, not a full model table. Other Mistral models were not priced on the page.
• Model names are as the vendors print them. Some are labelled Preview.

What "a million tokens" costs depends on the mix

A million tokens is not a unit with one price. If all of it is input, you pay the input column. If it is all output, the output column. Real work is a mix. As an illustration, take 80 percent input and 20 percent output. That mix is my assumption, not a vendor figure.

Make your own with this script. Change the mix to match your workload, and add a cached share if your tool caches heavily.

rows = [  # name, input, cached, output (USD per MTok)
    ("Sonnet 5.5", 2.00, 0.10, 10.00),
    ("gpt-6.1-sol", 2.00, 0.10, 10.00),
    ("Gemini 4 Argon (now)", 2.00, 0.10, 10.00),
    ("gpt-5.3-codex", 1.75, 0.175, 14.00),
]
share_in, share_cached, share_out = 0.20, 0.60, 0.20  # must sum to 1
for n, i, c, o in rows:
    cost = share_in*i + share_cached*c + share_out*o
    print(f"{n}: ${cost:.3f} per million tokens")

With 60 percent of the tokens served from cache, 20 percent fresh input and 20 percent output, these four models come out at $2.46, $2.46, $2.46 and $3.255 per million tokens. The cache share moves the answer far more than the model name does, which is why a headline input price alone is a poor guide.

Three things a price table does not tell you

• Quality per token. A cheaper model that needs three attempts can cost more than a dearer one that needs one.
• Tokens per task. Models differ in how many tokens they use to reason or to write code. The table has no such column.
• Discounts. Anthropic lists a 50 percent batch discount, and Mistral says batch is half price. Negotiated enterprise prices are not public.

None of this describes what any provider pays to serve a token. These are list prices charged to customers.

Keeping the bill in your control

Log tokens per request, split by fresh input, cached input and output, and set a monthly alert. If you route work through Safeguard you can bring your own model, so you pick the model and the meter. Safeguard's MCP tool calls are unlimited, so they do not add a second meter.

Quick check

Sources

• Anthropic API pricing
• OpenAI API pricing
• Gemini API pricing
• Mistral pricing