AI API pricing compared on one dated table

AI API pricing has finally become easy to compare. Every provider bills per million tokens, with input and output listed separately on public price lists. All of them are lined up here, recorded on the same day from the providers' official sources. The range runs from ten cents of input on Claude Haiku 5.5 and GPT-6 Luna to fifty dollars of output on GPT-6 Astra and Claude Fable 5.1. On output alone a factor of about a hundred separates the extremes.

The ground covered goes from the big dated table to a provider-by-provider reading, simulations recalculated on the same job, the hidden costs that skew comparisons and the underlying trend. Value judgments about which model earns its price on which task belong to a sister piece on value for money. The figures that follow stay raw and can be checked line by line on the official pages.

01Key takeaways

• All prices are stated per million tokens, with input and output listed separately, in dollars on the official price lists.
• Today's range runs from $0.10 per million input tokens (Claude Haiku 5.5 and GPT-6 Luna) to $50 per million output tokens (GPT-6 Astra and Claude Fable 5.1). On output alone, from $0.50 to $50, that's a factor of about a hundred.
• The same 1,500-word article costs from about a tenth of a cent to about thirteen cents depending on the model chosen.
• Caching, batch processing and free tiers change the real bill. Comparisons that ignore them mislead.
• Prices per model tier have tended to fall since 2024 while quality keeps rising, so check the official lists before each decision.

02The dated table that compares everything

Prices per million tokens recorded on October 16, 2026 from the official price lists cited in the sources, in US dollars

Model	Input	Output
Mistral Small 4	$0.15	$0.60
GPT-5.4-nano	$0.20	$1.25
DeepSeek V4.1-Flash (peak hours)	$0.30	$1.20
Mistral Large 3	$0.50	$1.50
GPT-5.4-mini	$0.75	$4.50
Gemini 3.6 Flash	$0.75	$3.75
Claude Haiku 4.5	$1	$5
xAI grok-4.3	$1.25	$2.50
DeepSeek V4-Pro (peak hours)	$1.32	$3.96
Mistral Medium 3.5	$1.50	$7.50
GPT-5.4	$2.50	$15
Claude Sonnet 4.6	$3	$15
Claude Opus 4.8	$5	$25
GPT-5.5	$5	$30

Fourteen models from six providers, all recorded on a single date. That's the condition for an honest comparison, because price lists move several times a year and undated comparisons mix eras without saying so. The table is the centerpiece of the pay-per-use AI hub. Everything else, from model choice to budget calculations and saving strategies, is built on these fourteen lines.

03The newer generations released since June

Newer models on the same official price lists, per million tokens, recorded on October 16, 2026

Model	Input	Output
Claude Haiku 5.5	$0.10	$0.50
GPT-6 Luna	$0.10	$0.50
Claude Sonnet 5.5	$2	$10
GPT-6.1 Sol	$2	$10
Gemini 4 Argon	$2	$10
Claude Opus 5.5	$4	$20
GPT-6 Astra	$10	$50
Claude Fable 5.1	$10	$50

Eight more models arrived between June and October 2026, bringing the comparison to twenty-two lines. Two figures need a caveat. Gemini 4 Argon keeps the price shown until November 30, 2026, then moves to $4 and $20, while Claude Haiku 5.5 charges this rate on prompts of up to 100,000 tokens, which covers ordinary content work. The calculations elsewhere on the site still use the main table, because those models remain on sale at unchanged prices.

04How the figures were recorded

Every figure in the tables comes from the provider's official pricing page, consulted on October 16, 2026, with no intermediary or aggregator. The six price lists are cited in the sources and reopen in one click if you want to check. Versions not listed here (reasoning variants, older generations, specialized image or audio models) appear on the same official pages, at the same addresses.

Two choices are deliberate. Prices stay in dollars, the currency of the price lists, since conversion depends on the exchange rate and the VAT of your situation on the day you pay. The models chosen are also the ones a content creator actually meets, leaving obscure references aside. A comparison is only worth as much as its method. Ours fits in these two paragraphs.

05How to read these prices without mistakes

Three keys help you read them. The first is the unit. A million tokens represents about 750,000 words, the equivalent of ten novels. Nobody uses that much per month in content creation, which is why bills come in cents even though price lists are shown in dollars, a contrast that throws every newcomer. If the unit is still unclear, the basics of tokens covers it in ten minutes.

Next, look at the gap between columns, since output costs two to six times the input and drives the bill for creative work. Last, go from the table to YOUR bill. That takes one multiplication, laid out in the step-by-step cost formula. The simulations further down give an immediate preview on the site's reference piece, the 1,500-word article.

06OpenAI, the full ladder from nano to flagship

The best-known company lines up the widest range. GPT-5.4-nano sits at $0.20 and $1.25 for everyday jobs, GPT-5.4-mini at $0.75 and $4.50 as the workhorse, GPT-5.4 at $2.50 and $15 for demanding everyday work and GPT-5.5 at $5 and $30 at the top of that generation. The newer GPT-6 family runs from Luna at $0.10 and $0.50 to Astra at $10 and $50. The cache divides the price of input already seen by ten, while batch processing halves the cost of deferred tasks. OpenAI remains the default reference in tools on the market, which makes the first steps simpler.

In practice, the GPT-5.4 and GPT-5.5 ladder alone spans a factor of twenty-four on output, enough to route each task without changing provider. Creating the account and the key is covered in the OpenAI key in five minutes, with prepaid credit to back it.

07Anthropic, three clearly separate tiers

Claude comes in three clean levels. Haiku 4.5 costs $1 and $5 for volume, Sonnet 4.6 costs $3 and $15 as a mid-range model known for writing and Opus 4.8 costs $5 and $25 at the top. The same logic applies to options, with cached input at a fraction of the price and batches at half price for asynchronous work. The 5.5 generation repeats the three tiers at lower prices, with Haiku 5.5 at $0.10 and $0.50, Sonnet 5.5 at $2 and $10 and Opus 5.5 at $4 and $20, while Fable 5.1 sits above them at $10 and $50.

Within the 4.x generation the range is tighter than at OpenAI, a factor of five between the extreme output prices. Its deliberate focus on writing quality makes Sonnet the everyday favorite of many writers. The key setup route, console and credit included, is in the Claude key in practice.

08Google Gemini, the price list and the free tier

Gemini 3.6 Flash, which now also answers the requests sent to Gemini 3.5 Flash, is listed at $0.75 and $3.75 through December 31, 2026, then $1.50 and $7.50. Gemini 4 Argon, the new top model, costs $2 and $10 through November 30, 2026, then $4 and $20, with the full range on Google AI's official price list. The company's distinctive feature lies elsewhere. It offers a lasting free tier with no credit card, the most generous on the market for discovering and learning. The free tier comes with its own data usage terms, which are worth rereading in the console when you sign up.

That double face, free for learning and a public price list for production, makes Google the gentlest entry point in the comparison. The Gemini key and its free tier details the route and the current conditions.

09Mistral, DeepSeek and xAI, the challengers

The challengers sit near the bottom of the main table. Mistral tiers its range with Small 4 at $0.15 and $0.60, Large 3 at $0.50 and $1.50 and Medium 3.5 at $1.50 and $7.50, plus a free experimentation plan with SMS verification and no credit card. DeepSeek now serves its Flash model as V4.1-Flash at $0.30 and $1.20 at peak hours and V4-Pro at $1.32 and $3.96, with every rate halved off-peak (outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays). Its cache brings repeated input below one cent per million tokens, the lowest rate in the table.

xAI completes the group at $1.25 and $2.50 for grok-4.3, the gentlest output-to-input ratio in the table. These low prices have a healthy side effect, because they pull the whole market down. The big companies answered with budget models of their own. Claude Haiku 5.5 and GPT-6 Luna now undercut the challengers outright. Competition works on your bill quarter after quarter, without you having to negotiate anything.

10The second table, options that change the bill

Official pricing options recorded on October 16, 2026, with detailed conditions on the price lists cited

Option	Effect	Where
Cache for repeated input	Up to ÷10, sometimes far more	OpenAI, Anthropic and DeepSeek among others
Batch processing	About 50% off asynchronous jobs	OpenAI and Anthropic among others
Lasting free tier	$0 within quotas	Google AI Studio, Groq, Mistral Experiment
Welcome credit	Varies over time	Depends on the provider, the console is authoritative

These options sometimes weigh more than the price list itself. A production flow with heavy templates at a provider with generous caching can cost less than a floor-priced model with no cache, so the headline gap flips in real use. The first table ranks headline prices. This one is a reminder that the real bill also plays out in the footnotes.

11The same job at four prices, recalculated simulations

Take the site's yardstick, the 1,500-word article, with about 800 input tokens of brief and 2,500 output tokens. On GPT-5.4-nano it costs 0.33 cents. On GPT-5.4-mini it costs 1.2 cents. On Claude Sonnet 4.6 it costs 4 cents and on GPT-5.5, the most expensive model in the main table, it costs 7.9 cents. On Claude Haiku 5.5 or GPT-6 Luna it drops to 0.13 cents. All calculations were redone from the price lists above and can be checked with a calculator using (800 × input price + 2,500 × output price) ÷ 1,000,000, nothing more.

Stretch that to a typical month of ten articles with one iteration each. The bill then runs from 7 cents on nano to $1.60 on the flagship. Between these two models the gap is a factor of about twenty-four, the same as OpenAI's internal ladder on output. The absolute amounts stay so low that you can choose on quality first. That's exactly the subject of the value-for-money verdict, which judges what these gaps really buy.

12Which model for which need, in brief

For pure volume (product pages, variants, drafts) the floor models fit best, from Claude Haiku 5.5 and GPT-6 Luna to Mistral Small 4 and DeepSeek. The detailed match is in the cheapest model for writing. For everyday balance, the mid-range mini and Haiku models do the job. For the pages that count, use Sonnet, GPT-5.4 and the flagships, dosed content by content. Routing costs only a setting in the tool that you change per piece, with the strategic page on the flagship and drafts on the floor model.

If you're starting from scratch, price isn't the only question, because the free tier, the console and the ecosystem count too. Choosing your first key well crosses all these criteria in a five-minute read. It saves you from picking a provider on the price column alone.

13Three usage profiles, price lists in hand

The frugal blogger comes first, with six articles a month on a mid-range model and a few rewrites on a floor model. Their line in the table is the mini and Haiku column, for a monthly bill of around 30 cents all in. For this profile the hundredfold spread of the comparison is a statistical curiosity. Every route leads to less than one euro, so the choice can rest on writing comfort alone.

The site publisher comes next, with forty pieces a month across several sites. Floor models become their daily routine and options their lever, with caching trimming the cost of repeated templates and batches trimming the cost of night-time generations scheduled in advance. Their bill moves between one and three dollars. The real issue for them is organizing keys and templates, never the price list, which they tamed long ago.

The demanding user comes last, a brand or a writer with a strong voice, who puts flagships on everything published and mid-range models behind the scenes. Ten polished pieces a month cost about a dollar and a half. The flagship premium, two to three times the mid-range, is bought here knowingly, line by line, where the brand is at stake. Three profiles give three readings of the same table. Your statement will tell you which one you resemble.

14Hidden costs that skew comparisons

Four blind spots keep coming back in hasty comparisons. The first is caching. At equal list price, a provider with generous caching costs much less in real use, because your templates repeat at every call. Reasoning comes next. Its thinking tokens, billed as output, send bills soaring on dedicated models, so always compare within the same family, classic against classic.

Access fees come third. Hubs add credit purchase fees on top of the table's prices, while direct providers don't, so compare the all-in figure. The fourth is currency exchange. Price lists speak dollars while your card may pay in another currency at the day's rate, with VAT depending on your status and billing country. None of these factors upsets the ranking, but all of them shift the amounts, sometimes noticeably. That's why the footnotes of a price list deserve as much attention as the list itself before you decide between two close providers.

15The classic mistake of flagship versus budget

Half the debates claiming that one model is overpriced or another unbeatable compare different tiers, pitting one company's flagship against another's budget model. The table shows why that misleads. Each provider tiers its range from floor to top. The INTERNAL gaps (a factor of twenty-four at OpenAI) exceed the gaps between companies at the same tier.

An honest comparison therefore works on three levels, floor against floor, middle against middle and flagship against flagship. At each tier the prices are in fact very close. What really separates them is the quality produced on YOUR content, the options on the price list and the ecosystem. The table helps you discard biased comparisons, never to crown an absolute winner that doesn't exist. Be wary of anyone who names one without stating the tier and the date.

16The underlying trend is falling prices

The structural movement works in your favor. Prices at equal or higher capability have tended to fall since 2024, driven by competition and efficiency gains. The table goes out of date in the right direction, downward, though not evenly. DeepSeek and Mistral Small 4 raised their rates between June and October 2026, while Anthropic and OpenAI released cheaper generations. Check the official price lists again at every purchase decision, since they are authoritative to the day and the source links lead straight to them.

The strategic consequence is to lock nothing in. The strength of BYOK is precisely that it follows these drops the same day, with no renegotiation. When prices stall, the levers that cut the bill take over, starting with caching and batches, which stack without changing your work habits.

17How to use the comparison over time

Date your decisions. A model choice made on the June price list gets reviewed on the December one, in five minutes. Price drops and new models arrive in waves. A review twice a year is enough to catch them, with no daily monitoring and no fear of missing the deal of the century.

Another habit is to test by tier. When a new floor-priced challenger comes out, compare it with your current floor model on three real pieces, never with your flagship. The test costs fractions of a cent and settles the question better than general rankings, which are calibrated on tasks that aren't yours. BYOK makes these tests free in practice, since plugging in one more key needs no migration.

Finally, bring everything back to your reference piece. Thinking in terms of what one article costs on your current setup turns every new price list into an instant comparison. That's the unit everyone understands, far better than the millions of tokens in official announcements. The profiles above give the starting orders of magnitude.

18From the table to your key

With the comparison read, two steps remain. First, place your total budget, which is the job of the full pay-per-use budget. Then create the key for the provider you picked through all the other key guides, five minutes each from account to first billed call. The AI API pricing in the table then becomes YOUR pricing, readable on your statement and checkable against the table at every bill, which is the whole difference from opaque pricing.

Plug any of these keys into OpenTahr and you pay the provider's list price, since the account is free. The table above is exactly what you'll pay, with no in-house credit and no margin on your generations. That's the whole promise of the setup. You can check it on every monthly statement.

19Frequently asked questions

Which AI API is the cheapest?

On the day of recording, Claude Haiku 5.5 and GPT-6 Luna share the floor for input ($0.10 per million) and output ($0.50), ahead of Mistral Small 4 and DeepSeek V4.1-Flash off-peak. Being cheapest says nothing about the quality produced on your content. Test the floor models on your real tasks before making them your daily tool, since the full test costs fractions of a cent and settles it in an evening.

OpenAI or Claude, which costs less?

At a comparable tier the price lists look alike. Mini against Haiku and GPT-5.4 against Sonnet stay within a few tens of cents per million, while Haiku 5.5 and GPT-6 Luna share the exact same price. The gap plays out more on caching, batches and the perceived quality on your content than on the raw price list. Compare at the same tier and never flagship against budget, for the reasons given in the section on that classic mistake.

Why are prices per million tokens?

It's the common unit of the market, large enough for readable prices and standard enough to compare providers. A million tokens represents about 750,000 words. Your real content is therefore priced in thousandths of the listed rate, which is why bills come in cents even on models listed at several dollars.

Will the prices in the comparison change?

They probably will. The direction since 2024 has mostly been downward. That's why every table on the site shows its recording date and points to the official price lists. Those lists are authoritative on the day you make your decision.

Should I pick a single provider?

Nothing forces you to, which is the advantage of keys. Many people combine a floor model for volume with a high-end one for strategic pages, two keys in the same tool. A multi-provider key ring follows the table instead of suffering it. It takes twenty minutes to set up, one key after another. It also reorganizes just as quickly when price lists move.

Transparency note. All prices were recorded on October 16, 2026 from the providers' official price lists, cited in the sources. The simulations are recalculated from those lists and can be checked with a calculator. None of the providers named here pays us.

20Sources

1. OpenAI, official API price list
2. Anthropic, official Claude API price list
3. Mistral, official API price list
4. DeepSeek, official API price list
5. Google, official Gemini API price list
6. xAI, official models and pricing page
7. Anthropic, Claude API pricing documentation (all models)
8. Mistral, API price list per model