Issue 27 · Model watch

Price per million tokens, October 2026: Claude, Gemini and GPT-6 compared

Input, cached input and output per million tokens for the current Claude, Gemini and OpenAI models, with a source for every row and the footnotes that change the bill.

Mubarak Shajahan10 October 20267 min read

As of 10 October 2026, the three big labs price their mid-tier models almost identically and their cheap tiers nearly so. Claude Sonnet 5.5 and GPT-6 Sol are both $2 in and $10 out per million tokens; Claude Haiku 5.5 and GPT-6 Luna are both $0.10 in and $0.50 out. Gemini 3.8 Flash undercuts both today, and from 1 January its price doubles to $1.50 in and $7.50 out. This post is the table, with a source label on every row and the footnotes that move the real bill.

Method: I read each vendor’s own pricing page on 10 October. Where a row comes from anywhere else it says so. All prices are USD per million tokens (MTok), Standard tier, text, short context, unless a footnote says otherwise.

What do the current models cost?

Source key: A is Anthropic’s pricing page, G is Google’s Gemini API pricing page, O is OpenAI’s developer pricing page. All three are vendor pages.

Model	Input	Cached input	Output	Source
Claude Fable 5.1	$10.00	$0.25	$50.00	A
Claude Mythos 5.1 (limited availability)	$10.00	$0.25	$50.00	A
Claude Opus 5.5	$4.00	$0.20	$20.00	A
Claude Opus 5	$5.00	$0.50	$25.00	A
Claude Sonnet 5.5	$2.00	$0.10	$10.00	A
Claude Sonnet 5	$2.00	$0.20	$10.00	A
Claude Sonnet 4.6	$3.00	$0.30	$15.00	A
Claude Haiku 5.5, prompts up to 100k tokens	$0.10	$0.01	$0.50	A
Claude Haiku 5.5, prompts over 100k tokens	$0.50	$0.05	$2.50	A
Claude Haiku 4.5	$1.00	$0.10	$5.00	A
Gemini 3.1 Pro Preview (up to 200k prompt)	$2.00	$0.20	$12.00	G
Gemini 3.8 Flash (to 31 Dec 2026)	$0.75	$0.075	$3.75	G
Gemini 3.6 Flash (to 31 Dec 2026)	$0.75	$0.075	$3.75	G
Gemini 3.5 Flash-Lite	$0.30	$0.03	$2.50	G
Gemini 3.1 Flash-Lite	$0.25	$0.025	$1.50	G
Gemini 2.5 Pro (up to 200k prompt)	$1.25	$0.125	$10.00	G
Gemini 2.5 Flash	$0.30	$0.03	$2.50	G
Gemini 2.5 Flash-Lite	$0.10	$0.01	$0.40	G
GPT-6 Astra	$10.00	$1.00	$50.00	O
GPT-6 Sol	$2.00	$0.20	$10.00	O
GPT-6.1 Sol	$2.00	$0.10	$10.00	O
GPT-6 Luna	$0.10	$0.01	$0.50	O
GPT-5.6 Sol (promotional)	$4.00	$0.40	$20.00	O
GPT-5.6 Terra	$2.00	$0.20	$12.00	O
GPT-5.6 Luna	$0.20	$0.02	$1.20	O

Third-party cross-check: BenchLM’s table (updated 8 October, hosted on benchlm.ai, not OpenAI) lists the same seven OpenAI rows above with identical numbers. It also lists models the vendor page extract I read did not cover: GPT-5.5 at $5.00 input, $0.50 cached, $30.00 output; GPT-5.4 at $2.50, $0.25, $15.00; GPT-5.4 mini at $0.75, $0.075, $4.50. Those three are third-party only. My read of OpenAI’s pricing page covered Astra, Sol, Luna, 6.1 Sol and the GPT-5.6 models; I did not see GPT-6 Instant priced, so it has no row.

Which footnotes change the bill?

Gemini 3.8 Flash and 3.6 Flash double on 1 January 2027. The Google page prices them “through December 31, 2026” and then, from 1 January 2027, at $1.50 input, $0.15 cached input and $7.50 output. Storage for explicit caches doubles from $0.50 to $1.00 per MTok per hour. A budget built on today’s numbers is wrong by January. 3.5 Flash-Lite, 3.1 Flash-Lite and the 2.5 models carry no such note.

OpenAI long context doubles input. Above 272K input tokens, GPT-6 Sol is $4.00 input, $0.40 cached and $15.00 output; Astra is $20.00, $2.00 and $75.00. The Fast tier is 2x Standard on every GPT-6 and GPT-5.6 model, Ultrafast is 6x on Astra and 6.1 Sol. GPT-5.6 Sol’s price is promotional, “at least through November 21, 2026”. OpenAI also lists cache writes as a separate column; at long context they are 1.25x input (GPT-6 Sol: $5.00 on $4.00 input), which matches the 1.25x write rate in OpenAI’s caching guide for GPT-5.6 and later. Check the page for the short-context write figures before you budget writes.

Claude Haiku 5.5 is priced by prompt length. A request whose prompt exceeds 100,000 tokens pays the higher row, and the count includes cache reads and writes. The 1M context window is available on the other 4.6-and-later models at standard pricing, so a 900k-token request costs the same per token as a 9k one.

Claude Sonnet 5’s price is now permanent. The page says the $2/$10 launch price “is now the standard price” and the rise to $3/$15 planned for 1 September 2026 “will not occur”.

Claude’s tokenizer produces more tokens. The pricing page says Claude 4.7 and later models use a newer tokenizer that produces “approximately 30% more tokens for the same text”. A per-token comparison against Gemini or GPT-6 therefore understates what the same document costs on Sonnet 5.5. I do not have tokenizer ratios for the other two labs, so count tokens on your own text with each vendor’s counter before you compare.

Cached input is not one multiplier. Claude reads at 0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1 and Mythos 5.1 and 0.1x elsewhere. Gemini reads at 0.1x. OpenAI reads at 0.1x, except GPT-6.1 Sol at 0.05x. Writes are priced separately on Claude (1.25x for five minutes, 2x for an hour). The next post in this series takes the mechanics apart.

Batch is half price everywhere I checked. Anthropic’s Batch API gives 50% off input and output. Gemini 3.8 Flash Batch is $0.375 input and $1.875 output. OpenAI’s Batch tier for GPT-6 Sol is $1.00 input, $0.10 cached, $5.00 output. Batch is asynchronous, so it fits evals, backfills and nightly jobs, not user-facing calls.

One secondary-source note: mixed-news reports that OpenAI halved the API price of GPT-6 Sol and Luna as part of the caching announcement. The pricing page shows only current numbers, so I cannot verify the before and after. If you have an old table, discard it.

What does it mean for people shipping products?

Three opinions, as design advice.

The list price is the smallest lever. Sonnet 5.5 and GPT-6 Sol are the same price; the choice between them is quality, latency and your tokenizer. The bigger movers are the ones in the footnotes: whether your prefix caches, whether a long prompt crosses a tier boundary (100k on Haiku 5.5, 200k on Gemini Pro, 272K on OpenAI), and whether a promotional price expires mid-quarter.

Do not budget Gemini 3.8 Flash on the current price past December. If you chose it for cost, run the January numbers now, because at $1.50 and $7.50 it is no longer the cheapest mid-tier option by the same margin.

Price the whole call, not the input. Output is five times input on the Sonnet, Haiku, Opus, Fable, GPT-6 and Gemini Flash rows above. A generation-heavy feature barely benefits from cheap input.

Hands-on: cost of 1,000 requests, uncached

Take a request with 2,000 input tokens and 200 output tokens, no caching, and price 1,000 of them. That is 2M input and 0.2M output tokens.

Model	1,000 requests
Claude Fable 5.1 / GPT-6 Astra	$30.00
Claude Opus 5.5 / GPT-5.6 Sol	$12.00
Claude Sonnet 5.5 / GPT-6 Sol	$6.00
Gemini 3.1 Pro (under 200k) / GPT-5.6 Terra	$6.40
Gemini 3.8 Flash, today	$2.25
Gemini 3.8 Flash, from 1 Jan 2027	$4.50
Gemini 3.5 Flash-Lite	$1.10
Gemini 3.1 Flash-Lite	$0.80
GPT-5.6 Luna	$0.64
Claude Haiku 5.5 / GPT-6 Luna	$0.30

Reproduce it, and swap in your own token counts:

PRICES = {  # USD per MTok: (input, output). From the vendor pages, 10 Oct 2026.
    "claude-sonnet-5-5": (2.00, 10.00),
    "claude-haiku-5-5":  (0.10, 0.50),   # prompts up to 100k tokens
    "gemini-3.8-flash":  (0.75, 3.75),   # until 31 Dec 2026; 1.50 / 7.50 after
    "gpt-6-sol":         (2.00, 10.00),
}

def per_1000(model, in_tokens=2000, out_tokens=200):
    i, o = PRICES[model]
    return (in_tokens * i + out_tokens * o) * 1000 / 1_000_000

for m in PRICES:
    print(f"{m:20s} ${per_1000(m):.2f}")

This is the cost with no caching and a single tokenizer’s token counts. For prefix-heavy workloads, the caching post shows the cached case, where the gaps change shape.

Takeaways

Mid-tier is a tie on list price: Sonnet 5.5 and GPT-6 Sol at $2 and $10. Cheap tier is a tie too: Haiku 5.5 and GPT-6 Luna at $0.10 and $0.50.
Gemini 3.8 Flash is $0.75 and $3.75 only through 31 December 2026; it doubles on 1 January.
Check tier boundaries: 100k on Haiku 5.5, 272K on OpenAI, 200k on Gemini 3.1 Pro. Crossing one changes the rate.
Claude’s newer tokenizer yields about 30% more tokens for the same text; count your own tokens per vendor before comparing.
Trust vendor pages for rows, and treat aggregators as a cross-check. BenchLM matched OpenAI’s page on every row I compared.

Next, cache read pricing across Claude, Gemini and GPT-6 lines up the multipliers, minimums and windows behind the cached column.

Sources

Claude docs: pricing (model table, tiers, batch, tokenizer note) · platform.claude.com
Gemini API pricing (Standard tier, 'through December 31, 2026' notes) · ai.google.dev
OpenAI API docs: pricing (GPT-6 and GPT-5.6, short and long context) · developers.openai.com
BenchLM OpenAI pricing table, updated 8 October 2026 (third-party aggregator) · benchlm.ai
mixed-news: OpenAI GPT-6 prompt caching coverage (secondary) · mixed-news.com