What Does an LLM API Cost in September 2026?
The flagship tier at both OpenAI and Anthropic is $10 per million input tokens and $50 per million output tokens. OpenAI's cheapest current tier is $0.20 input, a 50x spread inside one vendor's own price list.
Every price below was checked on 16 September 2026 against the vendors' own documentation. All figures are USD per million tokens, input first, output second.
| Model | Input | Output | 1M in + 1M out |
|---|---|---|---|
| gpt-6-astra (OpenAI) | $10.00 | $50.00 | $60.00 |
| Claude Fable 5.1 | $10.00 | $50.00 | $60.00 |
| Claude Mythos 5.1 | $10.00 | $50.00 | $60.00 |
| gpt-5.5 | $5.00 | $30.00 | $35.00 |
| Claude Opus 5 | $5.00 | $25.00 | $30.00 |
| gpt-5.6-sol | $4.00 | $20.00 | $24.00 |
| gpt-5.4 | $2.50 | $15.00 | $17.50 |
| gpt-5.6-terra | $2.00 | $12.00 | $14.00 |
| Claude Sonnet 5 | $2.00 | $10.00 | $12.00 |
| Meta Muse Spark 1.3 | $1.25 | $4.25 | $5.50 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $6.00 |
| gpt-5.4-mini | $0.75 | $4.50 | $5.25 |
| gpt-5.4-nano | $0.20 | $1.25 | $1.45 |
| gpt-5.6-luna | $0.20 | $1.20 | $1.40 |
| Meta Muse Spark Contributor | $0.10 | $0.20 | $0.30 |
The right-hand column is derived arithmetic: input plus output for one million tokens of each.
Google Gemini is absent deliberately: its prices were not verified here, and a guess is worse than a gap. The same applies to xAI, covered below. AI Perks tracks $7.7M in credits across 194 companies against these bills.
OpenAI API Pricing and the 10 December Shutdown
OpenAI's current line runs from gpt-6-astra at $10 / $50 down to gpt-5.6-luna at $0.20 / $1.20. The gpt-5 generation has an API shutdown date of 10 December 2026, so nothing new should be built on it.
gpt-6-astra arrived in the first week of September 2026 as the current flagship. Above a 200,000 token context it bills at $20.00 / $75.00, doubling the input rate and raising output by half. That is the only context-window figure verified here.
The Sol, Terra and Luna names replaced the older Instant, Thinking, mini and nano scheme in July 2026.
| Legacy model | Input | Output | Status |
|---|---|---|---|
| gpt-5 | $1.25 | $10.00 | API shutdown 10 December 2026 |
| gpt-5-mini | $0.25 | $2.00 | API shutdown 10 December 2026 |
| gpt-5-nano | $0.05 | $0.40 | API shutdown 10 December 2026 |
There is no current OpenAI model priced like gpt-5-nano. At $0.05 / $0.40 it was the cheapest thing OpenAI sold, and the cheapest current replacement, gpt-5.6-luna, is 4x the input price and 3x the output price. Derived from those list prices, teams moving off nano should budget for a 3x to 4x increase depending on their input and output mix, not a like-for-like swap.
Anthropic API Pricing and the Sonnet 5 Increase That Was Cancelled
Claude Sonnet 5 stays at $2 / $10 per million tokens. Anthropic had scheduled a rise to $3 / $15 on 1 September 2026 and then cancelled it, which is the most consequential pricing change of the quarter and the least reported.
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 announced as introductory through 31 August 2026. It is now simply the standard price.
| Claude model | Input | Output | Note |
|---|---|---|---|
| Claude Fable 5.1 | $10.00 | $50.00 | Top tier, Mythos-class, September 2026 |
| Claude Mythos 5.1 | $10.00 | $50.00 | Limited availability |
| Claude Opus 5 | $5.00 | $25.00 | Launched 24 July 2026 |
| Claude Sonnet 5 | $2.00 | $10.00 | Scheduled increase cancelled |
| Claude Haiku 4.5 | $1.00 | $5.00 | Cheapest Claude tier |
Two things fall out of that table. Opus 5 shipped at exactly the same price as Opus 4.8, so the generation bump cost nothing. And every current Claude model carries an output multiple of precisely 5x input, from Haiku to Fable. OpenAI is not uniform: gpt-6-astra and gpt-5.6-sol are also 5x, but the rest of the 5.x line runs 6x to 6.25x. On output-heavy work that ratio moves the bill more than the input price does. AI Perks tracks credits for both vendors.
Meta Muse and xAI: What Is Published and What Is Not
Meta Muse Spark is $1.25 / $4.25 across versions 1.1, 1.2 and 1.3. A Contributor tier drops that to $0.10 / $0.20, in exchange for Meta training on your prompts and completions. Muse Image is $0.01 per image.
Muse Spark's output multiple is 3.4x, the lowest of any full-price model on this page, and the Contributor tier is lower still at 2x. By arithmetic, that 3.4x creates a clean crossover against Claude Haiku 4.5: Muse is cheaper whenever output exceeds about one third of input volume, and Haiku is cheaper below that.
That Contributor tier is the cheapest verified price here. The trade is your data, so it suits prototypes, not anything under a confidentiality clause.
Meta runs no startup credit program for Muse. List price is your price, unusual in this market.
This page prints no xAI token price. Published sources conflict badly on Grok rates, pricing varies by model, and the only reliable figure is in the xAI console today.
What One Month of Real Usage Costs
On a workload of 50 million input and 10 million output tokens per month, the verified prices produce bills from $7 to $1,000. That is a 143x spread for the same job.
Estimates from the list prices above, at one fixed volume, with no caching or batch discounts.
| Model | Estimated monthly cost |
|---|---|
| gpt-6-astra | $1,000.00 |
| Claude Fable 5.1 | $1,000.00 |
| gpt-5.5 | $550.00 |
| Claude Opus 5 | $500.00 |
| gpt-5.6-sol | $400.00 |
| gpt-5.4 | $275.00 |
| gpt-5.6-terra | $220.00 |
| Claude Sonnet 5 | $200.00 |
| Meta Muse Spark 1.3 | $105.00 |
| Claude Haiku 4.5 | $100.00 |
| gpt-5.4-mini | $82.50 |
| gpt-5.4-nano | $22.50 |
| gpt-5.6-luna | $22.00 |
| Meta Muse Spark Contributor | $7.00 |
The cancelled Sonnet 5 increase is worth $100 a month on this exact workload: $3 / $15 would have produced $300 instead of $200. That is a 33% saving, roughly $1,200 a year, from a decision most teams never noticed.
The December shutdown cuts the other way: the same workload costs about $6.50 on gpt-5-nano and about $22.00 on gpt-5.6-luna, so the cheapest seat in OpenAI's house gets more expensive this winter. Credits at getaiperks.com are the hedge.
Which Tier Should You Actually Pick?
Pick the cheapest tier that passes your own evaluation. This page makes no capability claims: capability was not part of this verification, and a pricing page that ranks models on tests it did not run is guessing.
What price alone does tell you:
Output-heavy work, such as agents that write a lot: the output multiple dominates. Claude is a flat 5x, OpenAI runs 5x to 6.25x, Muse Spark 3.4x.
Input-heavy work, such as retrieval over long documents: the input rate dominates, and crossing OpenAI's 200,000 token threshold takes gpt-6-astra from $10 to $20 input.
High volume, low complexity, such as classification or routing: the bottom tiers put you at $22 a month rather than $1,000, a 45x difference on identical work.
Anything shipping after December: do not build on gpt-5, gpt-5-mini or gpt-5-nano. The API shuts down on 10 December 2026.
How Free Credits Change the Math
A team running Claude Sonnet 5 at the volume above spends $2,400 a year. Credits routinely cover multiples of that, which makes the choice between a $200 tier and a $500 tier far less urgent.
AI Perks tracks $7.7M in credits across 194 companies, an average of roughly $39,700 per company, though the distribution skews toward a handful of large grants.
Against the estimates above, $10,000 in credits is about 50 months of that Sonnet 5 workload, or 10 months of gpt-6-astra. Secure credits first, then choose a tier on capability rather than on the invoice.
Size the bill against the tables above, then check which of those vendors has a live offer running against it.
Frequently Asked Questions
How much does the OpenAI API cost in 2026?
OpenAI's current flagship, gpt-6-astra, costs $10 per million input tokens and $50 per million output tokens, rising to $20 and $75 above a 200,000 token context. The cheapest current tier, gpt-5.6-luna, is $0.20 and $1.20. Credits that offset those bills are tracked at getaiperks.com.
Is Claude cheaper than GPT?
At the top and middle of both lines the prices are close. Flagships match exactly at $10 and $50. Claude Sonnet 5 at $2 and $10 undercuts gpt-5.6-terra at $2 and $12 on output. At the bottom OpenAI is cheaper, with gpt-5.6-luna at $0.20 against Claude Haiku 4.5 at $1.00.
Did Anthropic raise Claude Sonnet 5 pricing?
No. Anthropic had scheduled an increase to $3 and $15 per million input and output tokens on 1 September 2026, then cancelled it. Sonnet 5 stays at $2 and $10, which was originally announced as an introductory rate and is now the standard price.
What is the cheapest LLM API?
Among verified prices here, Meta's Muse Spark Contributor tier at $0.10 and $0.20 per million tokens is cheapest, but Meta trains on your prompts and completions. The cheapest tier without that trade is gpt-5.6-luna at $0.20 and $1.20. Free credits at getaiperks.com beat both.
What does the xAI Grok API cost?
xAI pricing varies by model, and published third-party figures conflict badly with one another. No token price is printed here because none could be verified against xAI's own documentation. Check the xAI console for current rates rather than trusting a secondhand table.
How do I get free OpenAI and Anthropic API credits?
Both vendors and several cloud partners run credit programs, and the amounts vary widely by company profile and stage. AI Perks tracks $7.7M in credits across 194 companies in one place, including which vendors currently have live offers and what each is worth. Start at getaiperks.com.
Know the price. Then get someone else to pay it.