What Is the Cheapest Usable LLM API Right Now?
The cheapest tier from a major lab is gpt-5.6-luna at $0.20 per million input tokens and $1.20 per million output. One million tokens in and one million out costs about $1.40.
Meta undercuts that. Muse Spark Contributor runs $0.10 / $0.20, roughly $0.30 for the same volume, but Meta trains on your prompts and completions. That is the real price of the discount.
Every number here was checked against the vendors' own pricing documentation on 16 September 2026. Several of these models shipped in the last three weeks, so most pages ranking for this query quote pre-launch rumour. AI Perks tracks $7.7M in credits across 194 companies, the other way to lower this bill.
Every Verified Price, Ranked by Real Cost
Ranked by one million input tokens plus one million output tokens, the current lineup runs from $0.30 to $95.00, a 317x spread across models people describe with the same three words.
All figures are USD per million tokens. The blended column is derived from the two listed prices.
| Model | Input | Output | 1M in + 1M out |
|---|---|---|---|
| Meta Muse Spark Contributor | $0.10 | $0.20 | $0.30 |
| gpt-5.6-luna | $0.20 | $1.20 | $1.40 |
| gpt-5.4-nano | $0.20 | $1.25 | $1.45 |
| gpt-5.4-mini | $0.75 | $4.50 | $5.25 |
| Meta Muse Spark 1.1 / 1.2 / 1.3 | $1.25 | $4.25 | $5.50 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $6.00 |
| Claude Sonnet 5 | $2.00 | $10.00 | $12.00 |
| gpt-5.6-terra | $2.00 | $12.00 | $14.00 |
| gpt-5.4 | $2.50 | $15.00 | $17.50 |
| gpt-5.6-sol | $4.00 | $20.00 | $24.00 |
| Claude Opus 5 | $5.00 | $25.00 | $30.00 |
| gpt-5.5 | $5.00 | $30.00 | $35.00 |
| Claude Mythos 5.1 | $10.00 | $50.00 | $60.00 |
| Claude Fable 5.1 | $10.00 | $50.00 | $60.00 |
| gpt-6-astra | $10.00 | $50.00 | $60.00 |
| gpt-6-astra above 200k context | $20.00 | $75.00 | $95.00 |
Two notes on what this table is not. It is a price ranking, not a quality ranking: we have not benchmarked these models and will not publish scores we did not run. And xAI is absent because sources conflict badly on Grok's per-token rates, so check xAI's own developer console.
Meta also prices Muse Image at $0.01 per image. Mythos 5.1 matches Fable's price but is limited availability.
Why Output Tokens Decide the Bill, Not Input
Every model in that table charges more for output than for input, by 2x to 6.25x. Output volume, not context size, is what makes an API bill move.
The multiples cluster by vendor:
- Anthropic runs a flat 5x across Haiku 4.5, Sonnet 5, Opus 5 and Fable 5.1.
- OpenAI's cheap tiers run 6x or worse. gpt-5.4-nano is 6.25x, the steepest on the board; luna, terra, gpt-5.4, gpt-5.4-mini and gpt-5.5 are 6x. Only sol and astra flatten back to 5x.
- Meta Muse Spark is 3.4x, the flattest paid curve here.
This changes which model wins. A summarisation job at 100k in and 2k out is dominated by input price. A coding agent at 5k in and 20k out is dominated by output price, where Muse Spark's 3.4x multiple beats tiers with a cheaper input rate.
Rank by the column matching your ratio, not the input price every comparison page quotes.
Two Price Changes Nobody Covered Properly
Anthropic cancelled the Sonnet 5 increase
Claude Sonnet 5 was scheduled to rise from $2/$10 to $3/$15 per million tokens on 1 September 2026. Anthropic cancelled it.
Verbatim from Anthropic's pricing page: "The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur." The $2/$10 rate was originally announced as introductory through 31 August 2026. It is now simply the standard price.
On a month of 50M input and 10M output tokens that is $200 instead of $300, a 33% saving against the price that was on the calendar (estimate derived from the listed rates). A lot of published content still quotes $3/$15, written before the cancellation.
Legacy OpenAI models shut down on 10 December 2026
gpt-5, gpt-5-mini and gpt-5-nano stop serving API traffic on 10 December 2026. Do not start new builds on them. The bigger issue is that the replacements are not priced the same. Blended figures below are derived from the listed prices.
| Legacy model | Price | Nearest current tier | Price | Blended change |
|---|---|---|---|---|
| gpt-5-nano | $0.05 / $0.40 | gpt-5.6-luna | $0.20 / $1.20 | $0.45 to $1.40, about 3.1x |
| gpt-5-mini | $0.25 / $2.00 | gpt-5.4-mini | $0.75 / $4.50 | $2.25 to $5.25, about 2.3x |
| gpt-5 | $1.25 / $10.00 | gpt-5.6-terra | $2.00 / $12.00 | $11.25 to $14.00, about 1.2x |
The cheap end of OpenAI's line got substantially more expensive. Anyone running volume on gpt-5-nano should model roughly a 3x increase before December, or price Claude Haiku 4.5 and Muse Spark against it now. The tier mapping is our judgement, not an official OpenAI migration path. AI Perks lists credits that absorb the difference.
What Each Tier Costs Per Month at Real Volume
At 50 million input and 10 million output tokens a month, the identical job costs $7 on Meta's Contributor tier and $1,000 on gpt-6-astra.
| Model | Estimated monthly cost | Multiple of cheapest |
|---|---|---|
| Meta Muse Spark Contributor | $7 | 1x |
| gpt-5.6-luna | $22 | 3.1x |
| gpt-5.4-nano | $22.50 | 3.2x |
| gpt-5.4-mini | $82.50 | 11.8x |
| Claude Haiku 4.5 | $100 | 14.3x |
| Meta Muse Spark | $105 | 15x |
| Claude Sonnet 5 | $200 | 28.6x |
| gpt-5.6-terra | $220 | 31.4x |
| gpt-5.4 | $275 | 39.3x |
| gpt-5.6-sol | $400 | 57.1x |
| Claude Opus 5 | $500 | 71.4x |
| gpt-5.5 | $550 | 78.6x |
| gpt-6-astra, Fable 5.1, Mythos 5.1 | $1,000 | 142.9x |
Estimates derived from the listed per-million prices at that volume. They exclude caching, batch and committed-use discounts, which we did not verify.
The gap worth staring at is Haiku 4.5 at $100 against the $1,000 flagship row: a 10x premium for capability this page does not measure. Run your own evaluation before paying it.
Which Model Fits Which Workload
Pick by your input/output ratio and whether the work tolerates a cheap tier, not by which flagship has the newest name.
- Classification, routing, extraction, tagging at volume. gpt-5.6-luna at $1.40 blended, or Claude Haiku 4.5 at $6.00 for Anthropic's flatter 5x output curve. Luna is 4.3x cheaper blended.
- Generation-heavy work with long outputs. Meta Muse Spark at $5.50 blended and a 3.4x output multiple, the mildest output penalty on the board.
- General production traffic. Claude Sonnet 5 at $12.00 blended, the tier whose scheduled increase was cancelled outright.
- The top price tiers. Claude Opus 5 at $30.00 blended or gpt-6-astra at $60.00. Opus 5 launched on 24 July 2026 at the same price as Opus 4.8, a straight upgrade for anyone on the older tier.
- Anything you cannot send to a training set. Skip Meta's Contributor tier entirely. The $0.30 blended price exists because Meta trains on your prompts and completions.
- Long context on OpenAI. Above 200k tokens, gpt-6-astra moves to $20 / $75, which is 2x input and 1.5x output over standard. That threshold is the only context-related figure we verified; we publish no context window sizes and no rate limits.
Free Credits That Cover the Bill Entirely
The fastest way to cut an API bill is not a cheaper tier. It is not paying for the tokens at all.
Run the arithmetic against the table above. At the $500 per month Opus 5 workload, $10,000 in credits is twenty months of runway; at the $200 per month Sonnet 5 workload, more than four years. No tier change on this page comes close.
AI Perks tracks $7.7M in credits across 194 companies, covering model APIs and the compute underneath them.
Two things to plan around. Meta has no startup credit program for Muse, so its low list price is the only discount there and you pay cash from day one. And credit terms change often enough that a list checked six months ago is not a list. Current terms are tracked at AI Perks.
Frequently Asked Questions
What is the cheapest LLM API in 2026?
gpt-5.6-luna at $0.20 per million input tokens and $1.20 per million output is the cheapest current tier from OpenAI or Anthropic, about $1.40 for a million in and a million out. Meta Muse Spark Contributor is cheaper at $0.10 / $0.20, but Meta trains on your data. Credits at getaiperks.com.
Did Anthropic raise the price of Claude Sonnet 5?
No. The increase from $2/$10 to $3/$15 per million tokens scheduled for 1 September 2026 was cancelled. Anthropic's pricing page states that the increase "will not occur." The $2/$10 rate, originally announced as introductory through 31 August 2026, is now the standard Sonnet 5 price.
When do the legacy OpenAI models shut down?
gpt-5, gpt-5-mini and gpt-5-nano stop serving API traffic on 10 December 2026. Do not start new builds on them. The nearest replacements cost more: gpt-5-nano's blended $0.45 becomes roughly $1.40 on gpt-5.6-luna, about a 3x increase. Credits can absorb it at getaiperks.com.
How much does the GPT-6 Astra API cost?
$10 per million input tokens and $50 per million output, OpenAI's current flagship rate since its launch on 3-4 September 2026. Above 200,000 tokens of context it rises to $20 / $75. At 50M in and 10M out a month that is roughly $1,000, or about $1,750 on long context, both estimates derived from those rates.
What does the Grok API cost?
Sources conflict badly on xAI's per-token rates, so we will not print one. Pricing varies by model: check xAI's own developer console for current figures. Every other price here was checked against vendor documentation on 16 September 2026.
Which is cheaper, Claude Haiku 4.5 or Claude Sonnet 5?
Haiku 4.5 costs exactly half: $1 / $5 per million tokens against Sonnet 5's $2 / $10, so a million in and a million out runs $6.00 against $12.00. Both sit on Anthropic's flat 5x output multiple. Free credits for either at getaiperks.com.
The cheapest token is the one somebody else already paid for.