AI Tools

What is the cheapest LLM API in 2026?

Published 2026-09-18 · Researched and price-checked on official sources

The cheapest LLM API price verified on 18 September 2026 is TypeSafe's Jev: $0.042 per million input tokens, output billed at zero. Jev returns structured decisions, never text. For ordinary text generation the floor is Google's Gemini 2.5 Flash Lite at $0.10 input and $0.40 output per million tokens.

Key facts

What matters

TypeSafe's Jev quotes $0.042 per million input tokens with output billed at zero, the lowest rate on any public price list checked for this page, and it is still in early access behind a waitlist.
The cheapest general-purpose text model is Google's Gemini 2.5 Flash Lite at $0.10 input and $0.40 output per million tokens, which the Flex/Batch tier halves to $0.05 and $0.20.
DeepSeek V4.1 Flash costs $0.15 input and $0.60 output off-peak and exactly double that in peak hours, which run 01:00-04:00 and 06:00-10:00 UTC Monday to Friday.
OpenAI's cheapest model, GPT-5.6 Luna, is $0.20 input and $1.20 output, while both flagships, GPT-6 Astra and Claude Fable 5.1, list the same $10 input and $50 output.
Cache reads are cheap (10% of the input rate on Astra, 2.5% on Fable 5.1) but cache writes cost 125% of input at $12.50 per million on both, so caching only pays on prompts you resend.
Owning the hardware is the other lever: RackNerd's 8 GB KVM VPS is $119.99 per year, about $10 a month, per its specials page on 18 September 2026.

Cheapest LLM API prices, verified 18 September 2026

Every rate on this page was read from the provider's own pricing page on 18 September 2026: USD per million tokens, standard tier, short context. Four numbers define the cheap end of the market today.

TypeSafe Jev: $0.042 input, output billed at zero. Structured decisions only, and access is still gated.
Google Gemini 2.5 Flash Lite: $0.10 input, $0.01 cached, $0.40 output. The Flex/Batch column halves that to $0.05 and $0.20.
DeepSeek V4.1 Flash: $0.15 input and $0.60 output off-peak, and $0.003 per million on a cache hit.
OpenAI GPT-5.6 Luna: $0.20 input and $1.20 output, with a Batch tier at $0.10 and $0.60.

The top of the market is a tie. GPT-6 Astra and Claude Fable 5.1 both list $10 per million input tokens and $50 per million output tokens, so a flagship decision now turns on caching, batch access and context limits rather than on price.

What TypeSafe's $0.042 per million tokens buys

TypeSafe AI left stealth on 15 September 2026 with a model class it calls System One and a first model named Jev. The launch post was the top Hacker News story of that week at 1,871 points. Jev is not a small language model: it gives up string generation entirely. You define the output schema in advance and it returns typed values with calibrated confidence, which TypeSafe says makes type errors impossible by construction. Published latency is 70ms to 500ms, against 3 to 329 seconds for the frontier text models it compares itself with.

Three caveats belong next to that $0.042 figure. Access is gated: Jev is in early access and the site collects waitlist signups. The service runs from the US West Coast, which shows up as latency everywhere else. And the launch post states plainly that the team cannot prove the pricing is not subsidised and expects it to fall over time, which is a more candid framing than most launch pages offer.

The cheap end of the API market therefore has two different floors. A structured-decision endpoint at $0.042 per million input tokens is not comparable to a text model at $0.10, because it cannot do the job most API buyers are shopping for. If your workload is routing, scoring or checking output at volume, it is the cheapest meter on this page. If you need a paragraph back, ignore it.

Why the input price is not the bill

Vendors advertise input tokens because that is the small number. Three levers decide what you actually pay.

Output tokens. OpenAI and Anthropic both charge five times the input rate for output on their current flagships: $10 in against $50 out on GPT-6 Astra and on Claude Fable 5.1. A workload that writes long answers is priced by the output column, not the headline.
Batch and Flex tiers. OpenAI's Batch table is exactly half the Standard table, so GPT-6 Astra falls to $5 in and $25 out and GPT-5.6 Luna to $0.10 and $0.60. Anthropic advertises the same 50% discount for batch processing, and Google's Flex/Batch column cuts Gemini 2.5 Flash Lite to $0.05 in and $0.20 out.
Caching. A cache read is cheap: $1.00 per million cached input on Astra, which is 10% of the input rate, and $0.25 on Fable 5.1, which is 2.5%. A cache write is not cheap: both charge $12.50 per million, or 125% of the input rate. Caching pays on prompts you resend and costs extra on prompts you send once.

DeepSeek bolts a clock onto the same idea. Its peak rate is exactly double the off-peak rate, and peak hours are 01:00-04:00 plus 06:00-10:00 UTC, Monday to Friday. Everything else, weekends included, bills at half.

Introductory rates that change before spring

Cheap list prices often carry an expiry date, and four of them are dated on the providers' own pages.

Gemini 3.8 Flash is $0.75 in and $3.75 out through 31 December 2026, then $1.50 and $7.50 from 1 January 2027.
GPT-5.6 Sol at $4 in and $20 out is promotional pricing that OpenAI's pricing page says is available at least through 21 November 2026.
DeepSeek's model names moved. Legacy deepseek-v4-flash requests are now served by V4.1-Flash at Flash prices, and V4 Pro (V4-Pro-0813) stayed on the API after its 14 September 2026 retirement date with billing unchanged.
Long context is a second price list at OpenAI. GPT-6 Astra is $20 in and $75 out per million tokens once a request moves into the long-context column.

Which cheap LLM API fits your workload

Match the tier to the job rather than to the price column.

Structured decisions at volume (route a ticket, score a lead, check an output): TypeSafe Jev if you can get access, otherwise Gemini 2.5 Flash Lite or DeepSeek V4.1 Flash off-peak.
Bulk offline text: Gemini 2.5 Flash Lite on Flex/Batch at $0.05 in and $0.20 out, OpenAI GPT-5.6 Luna on Batch at $0.10 and $0.60, or Claude Haiku 4.5 with Anthropic's 50% batch discount, which comes to $0.50 in and $2.50 out at published rates.
An interactive product with reasoning: Gemini 3.8 Flash at $0.75 and $3.75 on introductory pricing is the cheapest option here, with OpenAI's GPT-5.6 Terra at $2 and $12 as the alternative.
Agentic coding and long-horizon tasks: Astra and Fable 5.1 cost the same, so compare the cache read rate ($1.00 against $0.25), batch access and context limits instead.

When token spend turns into a line item you manage rather than a bill you pay, the other path is owning the hardware. We priced the cheapest GPU cloud for LLM inference, and for small models a budget VPS running Ollama is a flat monthly cost instead of a meter. RackNerd lists an 8 GB KVM plan at $119.99 per year, which is about $10 a month, on its specials page today.

How these prices were verified

Every number here comes from a provider pricing page fetched on 18 September 2026: OpenAI's API pricing page, Anthropic's pricing page, DeepSeek's Models and Pricing documentation, Google Cloud's Vertex AI pricing page, TypeSafe's launch post, and RackNerd's specials page. Nothing is estimated and no rate is quoted from a review site or a reseller.

Scope limits, stated plainly. These are list prices for short-context requests on standard tiers, and they exclude free tiers, resellers and volume commitments. Google's figures are Vertex AI list prices, so the separate AI Studio free quota is not counted. Anthropic's batch figure comes from its published 50% batch discount rather than a per-model batch table. Providers change list prices without notice, and DeepSeek's own documentation recommends rechecking its pricing page regularly, which is why every rate above carries a date. Several pages currently ranking for this query are stamped May through July 2026 and still quote model names that have since been replaced, so treat any rate without a recent date as history.

At a glance

Provider and model	Input / 1M	Cached input / 1M	Output / 1M	Note
TypeSafe Jev	$0.042	Not published	$0 (billed at zero)	Structured decisions only; early access
Google Gemini 2.5 Flash Lite	$0.10	$0.01	$0.40	Flex/Batch halves to $0.05 and $0.20
DeepSeek V4.1 Flash (off-peak)	$0.15	$0.003	$0.60	Peak doubles to $0.30, $0.006, $1.20
OpenAI GPT-5.6 Luna	$0.20	$0.02	$1.20	Batch tier $0.10 and $0.60
Google Gemini 3.8 Flash (intro)	$0.75	$0.075	$3.75	$1.50 and $7.50 from 1 January 2027
Anthropic Claude Haiku 4.5	$1.00	$0.10	$5.00	Batch discount of 50%
OpenAI GPT-5.6 Terra	$2.00	$0.20	$12.00	Long-context column is 2x input
Anthropic Claude Sonnet 5	$2.00	$0.20	$10.00	Batch discount of 50%
OpenAI GPT-5.6 Sol (promo)	$4.00	$0.40	$20.00	Promotional rate to at least 21 Nov 2026
Anthropic Claude Opus 5	$5.00	$0.50	$25.00	Batch discount of 50%
OpenAI GPT-6 Astra	$10.00	$1.00	$50.00	Long-context column $20 in, $75 out
Anthropic Claude Fable 5.1	$10.00	$0.25	$50.00	Cache read is 2.5% of input

Cheapest LLM API rates, read from vendor pricing pages on 18 September 2026 (USD per million tokens, standard tier, short context)

FAQ

Is TypeSafe's Jev cheap for everything?

No. Jev bills $0.042 per million input tokens with output at zero, but it returns typed structured decisions instead of text and it is still in early access behind a waitlist. Use it for classification, routing, scoring and output checks, and use a text model such as Gemini 2.5 Flash Lite at $0.10 and $0.40 for anything that needs generated prose.

Is Gemini 2.5 Flash Lite cheaper than GPT-5.6 Luna?

Yes, on both columns. Google lists Gemini 2.5 Flash Lite at $0.10 per million input tokens and $0.40 output, against $0.20 and $1.20 for OpenAI's GPT-5.6 Luna. Google's Flex/Batch tier takes Flash Lite down to $0.05 and $0.20, so the gap widens on offline work.

Why do DeepSeek's prices change during the day?

DeepSeek bills peak hours at double its off-peak rates. Peak runs 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, so V4.1 Flash costs $0.30 per million input tokens and $1.20 output in those windows and $0.15 and $0.60 at all other times, including the whole weekend.