LLM API Pricing Comparison 2026: What a Prompt Really Costs

Most AI pricing pages list dollars per million tokens and leave you to do the math. This one starts where your bill does: what a single prompt costs, what 1,000 cost, and which model is cheapest for the job you actually have. Every price below comes from OpenAI's, Anthropic's and Google's own pricing pages, checked on October 30, 2026, and the live table switches to announced prices on the day they take effect.

The short version:

• The cheapest capable models are GPT-6 Luna and Claude Haiku 5.5, tied at $0.10 per million input tokens and $0.50 per million output tokens. A typical chat turn costs about four hundredths of a cent.
• The three companies now charge the same four price points, so sticker price no longer separates them. How many tokens a model uses for your task does.
• The answer is most of the bill. Output costs five to eight times more than input, so 70–93% of what you pay for a typical prompt is the reply, depending on how much context you send.
• Google's prices are going up: Gemini 4 Argon doubles on December 1, 2026, and Gemini 3.8 Flash doubles on January 1, 2027.

#The cheapest model at each company for each kind of job

Not every task needs a frontier model. We group models into four levels by what they're good for, then list the cheapest one at each company. List prices are per million tokens, input / output.

On list price alone: Luna and Haiku tie for light work, Google is cheapest for everyday work until December 31, OpenAI's Sol ties Argon for hard work until November 30, and the frontier tier is a tie. After January 1, 2027, Google isn't the cheapest at any level.

Price isn't quality, of course. The cheapest model is only a bargain if it gets the job right the first time; a redo costs more than the price gap. That's the question PromptSaver answers for each prompt.

#What one prompt costs on every model

Here's every current model we track, with the cost of a typical chat turn and of 1,000 of them. A "typical prompt" here is about 1,100 tokens in (your message plus the recent conversation, which chat apps re-send every turn) and 560 tokens out (an answer of roughly 400 words).

The spread is enormous: the same chat turn costs 100 times more on a frontier model than on a light one. At 100,000 prompts a month, that's about $39 on GPT-6 Luna or Claude Haiku 5.5 versus about $3,900 on GPT-6 Astra or Claude Fable 5.1, at the same token count (Claude counts more tokens for the same text, so its real bill runs about a third higher). To run your own numbers, the free AI cost calculator shows the monthly bill for any workload on every model.

#What real tasks cost

A "typical prompt" hides a lot of variety. A code fix sends a whole file; a reasoning task can generate thousands of hidden thinking tokens. Here's what 1,000 runs of six common jobs cost at list prices (input / output tokens in brackets; divide by 1,000 for a single run).

Two things stand out. First, a product description on a light model costs about a sixtieth of a cent; you can write a thousand of them for fifteen cents. Second, the "reasoning task" column is where bills explode, because thinking tokens are billed as output (more on that below).

#Same sticker price, different bill

Something new happened in 2026: the three companies converged on the same prices.

• $0.10 / $0.50: GPT-6 Luna and Claude Haiku 5.5
• $2 / $10: GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon (until November 30)
• $4 / $20: Claude Opus 5.5 and Gemini 4 Argon (from December 1)
• $10 / $50: GPT-6 Astra and Claude Fable 5.1

When the price per token is identical, the real difference is how many tokens each model uses for the same task. Three things drive that:

#Tokenizers count the same words differently

Each company splits text into tokens its own way. Anthropic's pricing documentation notes that its Claude 4.7 and later models produce about 30% more tokens for the same text than older ones. A per-token price cut can be partly eaten by a tokenizer that counts more tokens. Our guide to how AI tokens work covers this in detail.

#Answer length is decided by the model (and your prompt)

Two models given the same question can write answers of very different lengths. Apidog reports a third-party test in which Gemini 4 Argon used far more output tokens per task than GPT-6 Astra, which narrows the price gap you'd expect from the sticker. The single most effective control is in your hands: say how long the answer should be.

#Hidden thinking is billed as output

Reasoning models "think" before they answer, and you pay for that thinking at the output rate. OpenAI's reasoning guide says reasoning can run from a few hundred to tens of thousands of tokens per request; Anthropic and Google also bill thinking as output. A task that needs no reasoning shouldn't be sent to a model with high effort settings.

#Why the answer costs more than the prompt

On every current model, output tokens cost five to eight times more than input tokens. Roughly speaking, a model reads your whole prompt in one pass but writes its answer one token at a time, so each answer token takes more computing work.

The result: on a short prompt with a medium-length answer (400 tokens in, 600 out), the answer is 88% of the cost on most models, and 90–93% on the Gemini models with a higher output ratio. Even on a chat turn that re-sends 1,100 tokens of conversation, the answer is still about 72% of the bill.

That changes how you save money. Trimming a few words from your prompt saves little; asking for the length and format you actually need saves the most, along with choosing a model that doesn't need to think hard for an easy job. We walk through every technique, with numbers, in how to reduce your AI costs.

#Every current model, with cached and batch prices

Standard list prices per million tokens, checked October 30, 2026. "Batch" is the price for jobs you can wait up to 24 hours for.

#OpenAI

OpenAI now also charges for writing to its cache (1.25 times the input price for a 30-minute cache, 2.5 times for 12 hours), and its faster "Fast mode" (formerly Priority) costs twice the standard price. Source: OpenAI API pricing.

#Anthropic

Watch the Haiku 5.5 cliff: a prompt over 100,000 input tokens is billed at five times the normal rate for the whole request. A 150,000-token document with a 1,000-token answer costs about $0.078, not the $0.016 the headline price suggests. Source: Anthropic pricing.

#Google

Google also charges for cache storage by the hour, and its output prices include thinking tokens. Source: Gemini API pricing.

#Price changes on the calendar

Prices used to only go down. In 2026, Google scheduled two increases. Here's what each one does to 1,000 short prompts (400 tokens in, 600 out):

#December 1, 2026: Gemini 4 Argon doubles

Argon goes from $2 / $10 to $4 / $20 per million tokens, with batch, priority and cached prices doubling too. The same 1,000 prompts go from $6.80 to $13.60. From that date, GPT-6.1 Sol costs half as much at the same level, and Google's own Gemini 3.1 Pro Preview ($2 / $12) undercuts Argon, though it's still labeled a preview.

#January 1, 2027: Gemini 3.8 Flash and 3.6 Flash double

Both go from $0.75 / $3.75 to $1.50 / $7.50. The same 1,000 prompts go from $2.55 to $5.10, which makes Flash more expensive than GPT-6.1 Terra ($3.40). Google's pricing page lists the new prices and dates but doesn't give a reason.

#Retirements that force a switch

Older models are being shut down, and moving to a successor can change your bill. Claude Sonnet 4.5 retires on November 30, 2026; OpenAI shuts down GPT-5, o3 and related models on December 11, 2026, and GPT-4o and GPT-4.1 on January 14, 2027. Check the OpenAI and Anthropic deprecation pages before you migrate.

#ChatGPT vs Claude vs Gemini API: which is cheapest for what?

• Light work (tagging, short rewrites, extraction, quick answers): OpenAI and Anthropic tie at $0.10 / $0.50. Google's cheapest, Gemini 3.1 Flash-Lite, costs about three times as much.
• Everyday writing and code: Google's Gemini 3.8 Flash is cheapest until December 31; from January 1, OpenAI's GPT-6.1 Terra is. Anthropic's cheapest everyday model, Sonnet 5.5, costs twice Terra's price.
• Hard reasoning: GPT-6.1 Sol and Gemini 4 Argon tie until November 30; after that, Sol is half the price of Argon and of Claude Opus 5.5.
• Frontier: GPT-6 Astra and Claude Fable 5.1 are the same price. Google has no model at this level.

These are API prices, billed per token. Monthly chat subscriptions like ChatGPT Plus or Claude Pro are billed separately and work differently: there, the cost of a long prompt shows up as hitting your message limit sooner. For the chat apps themselves, see ChatGPT vs Claude vs Gemini: plans, prices and which is best at what.

#Ways to pay less

#Batch: half price if you can wait

All three companies take 50% off for requests you submit as a batch and collect within 24 hours. Reports, bulk product descriptions and data cleanup rarely need an instant answer.

#Prompt caching: pay a fraction for repeated context

If many requests start with the same long instructions or document, caching lets you pay 2.5% to 10% of the normal input price for that repeated part. Example: 1,000 support replies that each share a 3,000-token knowledge-base prefix, plus 300 new tokens in and 300 out. On Claude Sonnet 5.5 that's $9.60 without caching and $3.90 with it, 59% less (before cache-write fees).

#Avoid the extras that quietly raise a bill

Faster processing modes cost about twice as much; data-residency options add about 10%; long-context tiers (like Haiku 5.5 over 100K tokens) can multiply the price; and Google charges for search grounding after 5,000 free requests a month.

#Send each task to the right model

The biggest saving is choosing the cheapest model that does the job well, and asking for an answer that's as long as you need and no longer. That's what PromptSaver does for every prompt: it rewrites it for a better first answer and names the best-fitting model at each company. Browse our prompt library to see the recommended model for 100+ real tasks.

#What changed in 2025–2026

• Small models got much cheaper. Anthropic says Claude Haiku 5.5 is about 75% cheaper overall than Haiku 4.5; GPT-6 Luna costs less than half of what its predecessor did, according to third-party price trackers.
• Middle tiers got cheaper too. Anthropic says Opus 5.5 costs about 40% less than Opus 5 for typical work.
• The top tier got pricier. GPT-6 Astra launched at $10 / $50, up from about $5 / $30 for GPT-5.5 according to third-party trackers. Anthropic's top Fable line has been $10 / $50 since Fable 5.
• Google started raising prices, the only one of the three with announced increases.
• Prices got more complicated: long-context tiers, speed tiers, data-residency surcharges and cache-write fees are now common.

#Frequently asked questions

#What is the cheapest AI API in 2026?

GPT-6 Luna and Claude Haiku 5.5 tie at $0.10 per million input tokens and $0.50 per million output tokens, which works out to about $0.0004 for a typical chat turn. Haiku 5.5 costs five times more for prompts over 100,000 tokens. Google's cheapest current model is Gemini 3.1 Flash-Lite at $0.25 / $1.50.

#Is the Claude API cheaper than the OpenAI API?

They tie at the light level ($0.10 / $0.50 for Haiku 5.5 and Luna) and the frontier level ($10 / $50 for Fable 5.1 and Astra). In between, OpenAI is cheaper: GPT-6.1 Terra ($1 / $5) has no Claude equivalent, and GPT-6.1 Sol ($2 / $10) is half the price of Claude Opus 5.5 ($4 / $20).

#How much does one API call cost?

It depends on the model and on how many tokens go in and come out. A typical chat turn (about 1,100 tokens in and 560 out) costs about $0.0004 on the cheapest models, $0.0078 on mid-to-upper models like Claude Sonnet 5.5, and $0.039 on frontier models like GPT-6 Astra.

#How many words is 1,000 tokens?

About 750 English words, according to OpenAI; Google's guidance is 600 to 800 words. Other languages, code and numbers usually use more tokens per word.

#Do thinking or reasoning tokens cost extra?

Yes. OpenAI, Anthropic and Google all bill a model's hidden reasoning as output tokens, at the output price. A reasoning-heavy request can cost many times more than the visible answer suggests.

#Are Gemini API prices going up?

Yes. Gemini 4 Argon doubles from $2 / $10 to $4 / $20 per million tokens on December 1, 2026, and Gemini 3.8 Flash and 3.6 Flash double from $0.75 / $3.75 to $1.50 / $7.50 on January 1, 2027.

#How much do batch processing and prompt caching save?

Batch saves 50% at all three companies if you can wait up to 24 hours. Cached input costs 2.5% to 10% of the normal input price, which can cut a repetitive workload's bill by more than half.

#Is the API the same as ChatGPT Plus or Claude Pro?

No. Chat subscriptions are a flat monthly price with message limits; the API is billed per token. Long prompts and long answers cost more on the API and use up your limits faster on a subscription.

#How we priced this

All prices are standard pay-per-use list prices for text, taken from OpenAI, Anthropic and Google on October 30, 2026. Costs per prompt and per task are our own calculations from those prices: (input tokens × input price + output tokens × output price) ÷ 1,000,000. Token counts are estimates, because each company's tokenizer counts the same text a little differently. The live table switches to announced prices on the day they take effect, and we update this article when the official pages change.