AI Tools

What does a cheap LLM API cost in 2026? Four vendors, verified

Published 2026-10-27 · Researched and price-checked on official sources

DeepSeek's deepseek-flash is the cheap LLM API to beat among the four vendors compared here: $0.15 per million input tokens off-peak and $0.60 per million output, against $1.00 and $5.00 for OpenAI's gpt-6.1-terra. Every figure was read from the vendors' pricing pages on October 27, 2026, including the expiry dates that will reprice two of them.

Key facts

What matters

OpenAI's cheapest text model is gpt-6-luna at $0.10 input and $0.50 output per million tokens, one tenth of gpt-6.1-terra at $1.00 and $5.00.
Claude Haiku 5.5 matches Luna at $0.10 and $0.50 only for prompts up to 100K tokens; above that Anthropic charges five times as much.
Cache reads run 2.5% to 10% of the input rate, while cache writes cost 1.25x input at both OpenAI and Anthropic, so caching only pays when a prefix is reused.
Gemini 4 Argon's $2.00 and $10.00 rates double on December 1, 2026, and Gemini 3.8 Flash doubles on January 1, 2027.
DeepSeek halves every token rate outside 01:00-04:00 and 06:00-10:00 UTC on weekdays, which is the only time-of-day pricing in this comparison.

Cheap LLM API prices as of October 27, 2026

Four vendors publish per-token rates a buyer can compare directly: OpenAI, Anthropic, Google and DeepSeek. The table below covers every model each vendor sells at $10 per million input tokens or less, plus the two flagships that anchor the top of each range. Rates are per one million tokens on the standard on-demand tier, copied from the official pricing pages.

Two rules apply before you read the numbers. Batch processing halves every rate at OpenAI, Anthropic and Google, so an asynchronous job there should be budgeted at 50% of what the table shows. And for any product that resends the same long prompt, the cache column matters more than the input column, a point the next sections work through.

The cheap tier runs from $0.10 to $0.75 per million input tokens

At the bottom of the market, OpenAI's gpt-6-luna and Anthropic's Claude Haiku 5.5 both list $0.10 input and $0.50 output. DeepSeek sits just above them at $0.15 and $0.60 during off-peak hours, with a cache-hit rate of $0.003 per million that is the lowest in this comparison. Google prices Gemini 3.8 Flash at $0.75 input and $3.75 output, five to seven times the small-model rate charged elsewhere.

The price sheet says nothing about whether the cheap models are good enough for your task. gpt-6-luna costs one tenth of gpt-6.1-terra, and the two are not interchangeable for hard reasoning or long agent runs. Treat the cheap tier as the default for classification, extraction, routing, formatting and short replies, and move up a tier when output quality fails on your own evaluation set.

Anthropic adds a context cliff that the headline rate hides. Haiku 5.5 charges $0.10 input and $0.50 output while a prompt stays at or under 100K tokens, then $0.50 and $2.50 above that line, the same fivefold jump as its cache read and cache write rates. The cliff is easy to hit with a pasted transcript or a long tool output: the tokens cost five times as much at 120K context as at 90K, with no change to the model.

Cache reads and cache writes decide the monthly bill

In agentic coding, cache reads account for 96% to 99% of total token spend, because every turn resends the whole conversation, according to Max Woolf's October 25 analysis of the 2026 price cuts. One number then reorders the whole comparison: Anthropic's Fable 5.1 charges $0.25 per million cached input tokens against OpenAI's $1.00 for gpt-6-astra, even though both list the same $10 input and $50 output rate.

Both vendors price a cache write at 1.25x the input rate. That is $1.25 per million for gpt-6.1-terra with 30-minute retention and $12.50 for Fable 5.1. OpenAI charges 2.5x input, or $2.50, for 12-hour retention. Caching therefore costs money up front and only pays back when the same prefix is reused inside the retention window, which is why one-off calls should skip it.

Here is one workload priced six ways: 30 million input tokens per month with 70% served from cache, plus 6 million output tokens.

gpt-6-luna or Claude Haiku 5.5 with prompts under 100K tokens: $4.11 per month
DeepSeek deepseek-flash, off-peak: $5.01 per month
Gemini 3.8 Flash: $30.83 per month
gpt-6.1-terra: $40.05 per month
Claude Opus 5.5: $160.20 per month
Claude Fable 5.1: $395.25 per month

The gap between the cheapest and the most expensive option is roughly 96x on identical token volume. Nothing in the workload changed except the model name, which is the whole argument for matching a model tier to a task rather than defaulting to the flagship.

Four rates that change on a published date

Two of the four vendors are running launch pricing that expires within three months, so today's ranking is not the one you will pay in January.

Gemini 4 Argon: $2.00 input and $10.00 output through November 30, 2026, then $4.00 and $20.00 from December 1, 2026. Cached input moves from $0.10 to $0.20, and Google notes the model supports implicit caching only.
Gemini 3.8 Flash: $0.75 and $3.75 through December 31, 2026, then $1.50 and $7.50 from January 1, 2027, with caching moving from $0.075 to $0.15.
GPT-5.6 Sol: the $4.00 and $20.00 promotional rate is, per OpenAI's pricing page, available at least through November 21, 2026.
Claude Sonnet 5: the $2.00 and $10.00 launch rate was meant to be temporary and became permanent on August 10, 2026.

Price a 12-month commitment at the post-expiry rate. An agent on Gemini 4 Argon that costs $2.00 per million input tokens in November costs $4.00 in December, so an integration decision made on the launch number is twice as expensive two months later. The reverse also happens: Anthropic's Sonnet 5 launch discount stuck, which is the only reason that line is not on an expiry date too.

DeepSeek is the only vendor that prices by clock time

DeepSeek charges peak rates from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour, weekends included, bills at half price. For a team in UTC+8 that puts peak from 09:00 to 12:00 and 14:00 to 18:00 local time, which is precisely the working day.

The discount is large enough to schedule around. One hundred million off-peak input tokens on deepseek-flash cost $15, against $30 at peak, and the same halving applies to output tokens. Batch backfills, nightly re-indexing and evaluation runs that no one is waiting on belong in the off-peak window. Its 1M-token context and 384K maximum output also undercut the Western vendors on long-document work, where output tokens drive the bill and DeepSeek charges $0.60 per million against $50.00 for Fable 5.1.

The add-on lines that quietly grow the token bill

Token rates are the headline. Three other line items show up on real invoices.

Speed tiers. OpenAI's Fast mode doubles the standard rate, and Ultrafast multiplies it sixfold: $60.00 input and $300.00 output per million tokens for gpt-6-astra. Anthropic's fast mode for Opus 5.5 serves up to 2.5x faster at twice the standard price.
Tools. Anthropic charges $10 per 1,000 web searches and $0.08 per session-hour for Managed Agents, with code execution at $0.05 per hour per container after 50 free hours per day per organization. Google gives 5,000 Google Search grounding requests per month across Gemini 3 and newer models, then charges $14 per 1,000. DeepSeek bills Search and Fetch at $0.75 per 1,000 successful requests, and those calls stay free through its own harness until 00:00 Beijing time on November 1, 2026.
Regional routing. OpenAI applies a 10% uplift for regional processing and FedRAMP endpoints on models released on or after March 5, 2026. Anthropic charges 1.1x input and output for US-only inference.

A support bot that answers from a search index can spend more on 1,000 searches than on the tokens those searches produce. Model the tool fees for your expected volume before comparing base rates.

Which cheap LLM API fits which job

High-volume classification, extraction and routing: gpt-6-luna or Claude Haiku 5.5 at $0.10 and $0.50, or DeepSeek off-peak at $0.15 and $0.60 when the jobs can be scheduled.
Coding agents and long conversations: cache read price dominates, so compare $0.25 on Fable 5.1 against $1.00 on gpt-6-astra before looking at the headline rate.
Long documents: DeepSeek's 1M context at $0.15 input is the cheapest way to push volume, with Gemini 3.8 Flash at $0.75 as the Western alternative.
Budgets that must hold through Q1 2027: price Gemini at the post-expiry rate, or pick a vendor whose listed rate is not promotional.

The honest limit of this comparison is that price is the only dimension measured. A cheaper model that needs two attempts to answer is not cheaper, and these pricing pages carry no quality signal. Run the same 50 prompts from your own traffic through two candidates before switching a production path.

At a glance

Model	Input	Output	Cache read	Note
gpt-6-luna	$0.10	$0.50	$0.01	OpenAI small model
claude-haiku-5.5	$0.10	$0.50	$0.01	Prompts 100K tokens or less, 5x above
deepseek-flash off-peak	$0.15	$0.60	$0.003	Half price outside peak UTC hours
deepseek-flash peak	$0.30	$1.20	$0.006	01:00-04:00 and 06:00-10:00 UTC, Mon-Fri
gemini-3.8-flash	$0.75	$3.75	$0.075	Doubles January 1, 2027
gpt-6.1-terra	$1.00	$5.00	$0.05	Cache write $1.25 (30m) or $2.50 (12h)
gpt-6.1-sol	$2.00	$10.00	$0.10	GPT-6.1 family flagship
claude-sonnet-5.5	$2.00	$10.00	$0.10	Launch rate made permanent
gemini-4-argon	$2.00	$10.00	$0.10	Doubles December 1, 2026
claude-opus-5.5	$4.00	$20.00	$0.20	Fast mode at 2x price
claude-fable-5.1	$10.00	$50.00	$0.25	Cache write $12.50
gpt-6-astra	$10.00	$50.00	$1.00	Ultrafast at 6x price

Standard on-demand prices per 1M tokens, verified on the vendors' pricing pages on October 27, 2026.

FAQ

Is there a free LLM API?

Google's Gemini API is the one in this comparison with a free tier: Gemini 3.8 Flash lists input and output at no charge, with lower rate limits and with free-tier content used to improve Google's products. The pricing pages for OpenAI, Anthropic and DeepSeek list paid per-token rates.

Is OpenAI cheaper than Anthropic?

At the small tier they are identical: gpt-6-luna and Claude Haiku 5.5 both cost $0.10 input and $0.50 output per million tokens. One tier up OpenAI is half the price, with gpt-6.1-terra at $1.00 and $5.00 against Sonnet 5.5 at $2.00 and $10.00. At the flagship tier both charge $10.00 and $50.00, but Anthropic's cache read is $0.25 against OpenAI's $1.00.

Does Gemini API pricing go up in 2026?

Yes, on a published schedule. Gemini 4 Argon moves from $2.00 input and $10.00 output to $4.00 and $20.00 on December 1, 2026, and Gemini 3.8 Flash moves from $0.75 and $3.75 to $1.50 and $7.50 on January 1, 2027.