AI API Pricing August 2026: What Changed Across 10 Providers

For most of 2026, the story of AI API pricing has been a race downwards. August broke that pattern. In the space of three weeks, one provider cut a model's price by 80%, another quietly cancelled a scheduled increase, and a third announced that prices are going up and floated surge pricing by time of day.

If you set your budget in the spring, at least four of the numbers in it are now wrong. Here is every material change, with the rates as they stand on 12 August 2026.

AI API pricing in August 2026: the short version

• OpenAI cut GPT-5.6 Luna by 80% and Terra by 20% on 30 July. The flagship, Sol, did not move.
• Anthropic made Claude Sonnet 5's introductory $2/$10 rate permanent. The increase to $3/$15 scheduled for 1 September will not happen.
• DeepSeek warned of a significant price rise on 6 August, with no percentage and no date attached. Peak-hour surge pricing is reported but not yet published.
• Google shipped Gemini 3.6 Flash on 21 July at a lower output rate than the model it replaces.
• xAI's Grok 4.5 settled in as flagship at $2/$6, with tool calls billed separately from tokens.

All figures below are US dollars per million tokens, read from vendor pricing pages on 12 August 2026, including Mistral's API pricing page for the open-weight tier. Rate cards move without notice, so re-check the source before you commit a budget.

OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%

The GPT-5.6 family arrived first. Sol, Terra and Luna went generally available on 9 July with a shared context window of just over a million tokens. Three weeks later the price list changed underneath them: on 30 July, Terra's rate fell by 20% and Luna's by 80%, while Sol held at its launch price. Current rates are listed on OpenAI's API pricing page.

That leaves the family spread wide. Sol sits at $5/$30, Terra at $2/$12, and Luna at $0.20/$1.20.

The Luna cut is the aggressive one. An 80% reduction moves a model from "cheap enough for background jobs" to "cheap enough to run on every request". It also lands Luna alongside the budget models from Google and DeepSeek rather than its own siblings.

Two rates worth knowing before you model any of this: cached input bills at 10% of the standard rate, and the Batch API halves both input and output. OpenAI's floor remains GPT-5 nano at $0.05 per million input tokens, and its ceiling is GPT-5.5 Pro at $30/$180, a spread of roughly 600x across a single vendor's catalogue. The full tier list sits on our ChatGPT and OpenAI API pricing page.

Fast mode replaces Priority Processing

Alongside the cuts, OpenAI introduced Fast mode, which takes the place of the older Priority Processing option for latency-sensitive work. If you have Priority Processing referenced anywhere in your configuration or your cost model, that line needs revisiting.

Anthropic cancelled the Claude Sonnet 5 price rise

This is the change most pricing pages have not caught up with.

Claude Sonnet 5 launched with introductory pricing of $2/$10, advertised as running through 31 August 2026. The standard rate of $3/$15 was to apply from 1 September.

On 11 August, Anthropic confirmed to developers and in its own platform documentation that the $2/$10 rate is now the standard price, and that the scheduled increase will not apply. Billing already reflects it. No migration or integration change is required. The note on Anthropic's pricing documentation states it plainly: the previously scheduled increase to $3/$15 on 1 September 2026 will not occur.

In practical terms: if you built a migration plan around a September deadline, you can stand it down. Sonnet 5 at $2/$10 undercuts GPT-5.6 Terra and Gemini 3.1 Pro on output, both sit at $12, while matching them on input. That is a materially different competitive position from the one described in most comparisons written this month.

The rest of the Claude line is unchanged:

• Claude Fable 5, $10/$50
• Claude Mythos 5, $10/$50, limited availability
• Claude Opus 5, $5/$25
• Claude Sonnet 5, $2/$10
• Claude Haiku 4.5, $1/$5

Anthropic's modifiers are among the more generous on the market. Cache hits bill at 10% of standard input, the Batch API takes 50% off both directions, and the full one-million-token context window is charged at standard rates with no long-context surcharge.

Two additions to watch: server-side web search is billed separately at $10 per 1,000 searches, and pinning inference to the US applies a 1.1x multiplier across every token category. See the Claude API pricing overview for the full line-up, or our breakdown of what changed in Claude Opus 5.

Why most pricing pages still show a 1 September increase

Because the confirmation only reached developers on 11 August, as a billing notice and a documentation note rather than a launch announcement. Anything published before that date was accurate when written and is wrong now. Several comparisons from the first week of August still describe the 31 August expiry as a step-up to plan around.

Treat that as a general lesson. The authoritative number is on the vendor's own page, and promotional end-dates are the figures most likely to be stale everywhere else.

DeepSeek is raising prices and may add peak-hour surge pricing

On 6 August, DeepSeek told developers that API prices would rise significantly. Its own documentation now states that it plans "to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected", and advises developers to plan usage accordingly. No percentage was given and no effective date was set. The company is also reported to be preparing surge pricing during weekday peak hours, roughly 09:00 to 12:00 and 14:00 to 18:00 Beijing time, though this does not yet appear in DeepSeek's published pricing.

The context makes the reversal more striking. DeepSeek is the provider that triggered the industry-wide collapse in inference prices, by selling frontier-class output at close to zero margin, as we covered when V4 halved inference costs again. Its V4 Flash model still lists at around $0.14/$0.28, with cache hits at $0.0028, and current rates are tracked on our DeepSeek pricing page.

The reported cause is that it worked too well. V4 Flash is said to have processed 8 trillion tokens in a single day on 1 August, demand that outran the available compute. Alongside the price warning, DeepSeek is reported to be funding a large data centre build and raising capital ahead of a possible public listing.

If DeepSeek is in your routing table as the cheap lane, this is the month to price a fallback. An increase with no percentage and no date is not something you can plan around, but a routing layer that can move traffic to another budget-tier model is.

Google and xAI: smaller moves, same direction

Gemini 3.6 Flash lands at $1.50/$7.50

Google's newest Flash model arrived on 21 July at $1.50/$7.50, cutting the output rate by about 17% against the 3.5 Flash it replaces. Gemini 3.1 Pro is unchanged at $2/$12 for prompts up to 200,000 tokens, stepping up to $4/$18 above that threshold according to Google's Gemini API pricing documentation.

Below that, Gemini 3.5 Flash-Lite runs at $0.30/$2.50, and the older Gemini 2.5 Flash-Lite remains Google's cheapest listed model at $0.10/$0.40. Google's pricing documentation carries no retirement date for it, though the 2.0 Flash generation was shut down on 1 June 2026, so treat the 2.5 line as a model with a finite life rather than a permanent foundation. Batch takes 50% off across the line. Full rates are on the Google Gemini pricing page.

Grok 4.5 and the tool-call meter

Grok 4.5 became xAI's flagship on 8 July at $2/$6 with a 500,000-token context window, positioned for coding and agentic work. Grok 4.3 remains the cheaper general-purpose option at $1.25/$2.50, and Grok Build 0.1 covers coding workflows at $1/$2.

Two details matter more than the headline rate here. First, xAI doubles every rate once a prompt reaches 200,000 tokens, and it applies the higher rate to every token in the request, not just the tokens above the threshold. Grok 4.5 therefore becomes $4/$12 on a long prompt. Second, the batch discount is 20% rather than the 50% common elsewhere, covers only Grok 4.3 and the 4.20 variants, and excludes the 4.5 flagship.

Server-side tools also run on their own meter, reported at $5 per 1,000 calls for web search, X search and code execution, $10 per 1,000 for file attachments and $2.50 per 1,000 for collections search. These sit outside xAI's published model pricing table, so confirm them against your own billing before relying on them. For an agent that fires several searches per answer, they are not a rounding error. Current token rates are on our xAI Grok pricing page.

The August 2026 rate card at a glance

USD per million tokens, input/output, standard tier:

Open-weight models sit outside this structure entirely. Llama pricing depends on which host you use, starting from roughly $0.02 per million input tokens, and it moves in both directions: hosted rates on the Llama 3.x line fell this month while the Llama 4 models rose. Check the Meta Llama hosted rates before assuming a direction of travel.

Two current gaps are worth noting. Alibaba's Qwen3.8-Max has API access but no verified public token rate. Mistral's Shieldstral shipped as open weights with no hosted API price. Both are reminders that "available" and "priced" are different states.

Run your own workload through the interactive cost calculator rather than reading a table top to bottom. The ranking changes depending on your input-to-output ratio, and the pricing history explorer holds the month-by-month record behind every figure above.

What the changes look like on a real bill

Rate cards are abstract until you put a workload through them. Take a common shape: an assistant handling 100,000 requests a month, averaging 3,000 input tokens and 500 output tokens per request. That is 300 million input tokens and 50 million output tokens.

• GPT-5.6 Terra, $600 input, $600 output, $1,200
• Claude Sonnet 5, $600 input, $500 output, $1,100
• Grok 4.5, $600 input, $300 output, $900
• Claude Haiku 4.5, $300 input, $250 output, $550
• GPT-5.6 Luna, $60 input, $60 output, $120
• DeepSeek V4 Flash, $42 input, $14 output, $56

Two things fall out of that. The first is the scale of the Luna cut: the identical job would have cost roughly five times as much before 30 July, which is the difference between a line item someone questions and one nobody notices.

The second is what caching does to the ranking. Say 80% of that input is a stable system prompt or a repeated document. Cache hits at 10% of standard input bring Sonnet 5's bill down to roughly $668, a 39% reduction, with no change of model, provider or output quality.

A caching strategy is frequently worth more than a migration. It is also considerably less work.

Three things the headline rate does not tell you

Long-context thresholds can double your input rate

Comparison tables almost always quote the short-context rate. Gemini 3.1 Pro moves from $2/$12 to $4/$18 above 200,000 tokens. xAI doubles every rate at the same threshold. Anthropic, by contrast, bills the full million-token window at standard rates with no surcharge. If you routinely send long documents or large codebases, that difference outweighs most headline gaps.

Tokenisers are not comparable between providers

A price per million tokens only means something if a token means the same thing everywhere. It does not. Anthropic notes that Claude 4.7 and later models use a newer tokeniser, which produces roughly 30% more tokens for the same text.

So a model that looks 10% cheaper on the rate card can still lose on the invoice. The only reliable comparison is tokens measured on the model you intend to deploy. If the underlying unit is unfamiliar, our explainer on what a token actually is covers how the billing works.

The same model can carry more than one price

The same model can be quoted differently on the same day depending on where you buy it: direct from the vendor, through a cloud marketplace, or via a gateway. Marketplace listings do not reliably surface the vendor's standard rate. Before concluding that a provider is expensive, check which surface the number came from.

How to cut the bill without changing model

Model choice is the biggest lever, but it is not the only one, and the modifiers are consistent enough across vendors to plan around:

• Prompt caching. Cache hits typically bill at 10% of standard input. For any workload with a stable system prompt or a repeated document, this is the single largest saving available without changing anything else.
• Batch processing. A 50% discount on both input and output is standard at OpenAI, Anthropic, Google, Mistral and Azure. xAI is the exception at 20%, and it excludes the flagship. Some providers offer no batch tier at all, including DeepSeek, Perplexity and the hosted Llama models, which matters because those are often the cheapest on headline rate. A mid-tier model at half price can beat a budget model at full price.
• Route, don't standardise. Send volume to a budget model and escalate only the requests that need frontier reasoning. Treat the premium tier as an escalation lane rather than a default.
• Count the non-token line items. Tool calls, web searches, session runtime and per-second video rates sit outside the token meter and are easy to omit from a forecast entirely.

Our cost optimisation FAQ goes through each of these in more detail, and our price changes page shows when a tracked model moves so you are not re-reading vendor pages by hand.

What to watch in September 2026

Three things. Whether DeepSeek attaches a number and a date to its increase, and whether surge pricing by time of day spreads beyond it. Whether Anthropic's decision to make a promotional rate permanent sets a pattern or stays an exception.

And whether the mid-tier cluster holds. Terra, Gemini 3.1 Pro and Sonnet 5 now sit within a couple of dollars of each other, which usually precedes another move.

The broader signal is worth sitting with. For eighteen months the safe assumption was that any given model would be cheaper next quarter. August 2026 is the first month in a long while where that assumption failed in both directions at once.

Frequently asked questions

What is the cheapest AI API in August 2026?

Among first-party APIs, OpenAI's GPT-5 nano is the lowest at $0.05/$0.40, with Ministral 3 (3B) next at $0.10 flat and Gemini 2.5 Flash-Lite at $0.10/$0.40. Open-weight models hosted elsewhere go lower still: Llama 3.1 8B starts around $0.02 per million input tokens, depending on the host. For frontier-adjacent capability at a budget rate, GPT-5.6 Luna at $0.20/$1.20 is the notable new entry after the 30 July cut. Our cheapest AI API ranking sorts these by blended cost rather than headline rate, which changes the order.

Is Claude Sonnet 5 going up to $3/$15 in September?

No. Anthropic has confirmed that the $2/$10 introductory rate is now the standard price and the 1 September increase will not take place. Comparisons published earlier in August may still show the old deadline.

Why is output more expensive than input?

Generating tokens costs more compute than reading them. The multiple is not uniform, though. It ranges from roughly 2x to 8x depending on the model. That is why the input-to-output ratio of your workload changes which model is cheapest for you. A retrieval-heavy job with short answers and a long-form generation job will not rank providers the same way.

How often do these prices change?

More often than most budgets are reviewed. August alone brought two cuts, one cancelled increase, one announced increase and two new models. Anything more than a month old should be treated as indicative rather than current.

Compare the current rates for all 10 providers side by side, or model your own workload in the pricing calculator to see what these changes mean for your monthly spend.