What does an AI model API really cost to run?
Eighteen models from seven providers, one for each tier each provider sells (top, middle and small), priced on three monthly workloads at the standard list rate. Pick a workload, then a tier: the gap between tiers is far bigger than the gap between providers.
Last checked: 9 Nov 2026 · every price read off the provider's own pricing page · prices change often · no affiliate links
Jump to: the short answer · how the workloads are priced
The short answer
The tier you pick moves the bill far more than the provider. On the medium workload (20M input + 5M output tokens a month), OpenAI's and Anthropic's small models cost $4.50 (about £3.36) and their top models $450 (about £336.20): a hundredth. Elsewhere the gap is narrower: Google's small model costs about a tenth of its top one ($18.50 against $180), Together AI's about a twentieth ($4.65 against $90), Mistral's and DeepSeek's about a quarter, and xAI's about half. After that, a few levers move a real bill more than the list price: caching, batch jobs, DeepSeek's off-peak hours, long-prompt surcharges and the spend caps on new accounts.
Cheapest in each tier on the medium workload: top, Mistral Large 4 at $24.50 (about £18.30), then DeepSeek V4 Pro at $46.20 (about £34.52, peak rate) and Grok 4.7 at $70 (about £52.30); middle, Together AI's DeepSeek V4.1 Flash at $12 (about £8.97), then Gemini 3.8 Flash at $67.50 (about £50.43); small, GPT-6 Luna and Claude Haiku 5.5 at $4.50 each (about £3.36), then Together AI's Qwen3.5 9B at $4.65 (about £3.47).
Caching is where an agent saves most. Agents re-send the same instructions and history on every step, and cached input costs 90% to 98% less at most providers here, but only 75% to 84% less at xAI; Together AI's Qwen3.5 9B has no cached price. Writing to the cache costs extra at OpenAI (1.25 times input for 30 minutes, 2.5 times for 12 hours) and Anthropic (1.25 times for 5 minutes, 2 times for an hour), and Google charges $1 per 1M tokens per hour to store a Flash cache. The figures here assume no cache hits.
DeepSeek is half price off-peak, about 79% of the week. Its full rate applies only 01:00-04:00 and 06:00-10:00 UTC on weekdays; the figures here use that peak rate, so an agent running evenly through the week pays about 60% of them, and one that runs only at nights and weekends pays half. Together AI sells the same DeepSeek V4.1 Flash model at DeepSeek's peak price all day.
Batch jobs are half price at OpenAI, Anthropic, Google and Mistral, 20% off at xAI and not discounted at DeepSeek. Together AI advertises up to 50% off batch, but its docs discount only Whisper; chat models run at full price. Batch suits background work, not an agent answering in real time.
New accounts at OpenAI, Anthropic and Google have spend caps the heavy workload doesn't fit. OpenAI's Build tier stops at $500 a month, Anthropic's Start tier at $500, and Google's Tier 1 at $250 a month and $10 per 10 minutes, while the heavy workload (200M + 50M tokens) costs $4,500 on GPT-6 Astra or Claude Fable 5.1, $1,800 on Gemini 4 Argon and $675 to $900 on those three providers' middle models. OpenAI's next tier needs $100 of credit bought; Google's Tier 2 needs $100 paid and 3 days.
Long prompts cost more at three of them. OpenAI bills prompts over 272K tokens at $20 / $75 on GPT-6 Astra; xAI bills the whole request at double once a prompt reaches 200K tokens; Claude Haiku 5.5 charges five times as much ($0.50 / $2.50) once a prompt passes 100,000 tokens, cached tokens included. Anthropic's other models here include their full 1M-token context at the standard price, and Google lists no surcharge for these models.
Two of Google's prices are temporary. Gemini 4 Argon is $2 / $10 until 30 November 2026 and $4 / $20 from 1 December; Gemini 3.8 Flash is $0.75 / $3.75 until 31 December and $1.50 / $7.50 from 1 January 2027. The figures use the later prices: today the medium workload would be $90 on Argon and $33.75 on 3.8 Flash.
Keeping data in one region, or faster service, costs extra. Regional or data-residency endpoints add 10% at OpenAI, Anthropic (US only), Mistral (EU or US) and xAI (US), and 25% at Together AI (US only). Faster service costs twice the price at OpenAI (fast mode) and xAI (priority), 1.8 times at Google and 1.75 times at Mistral.
A token isn't the same size everywhere. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text than earlier Claude models, and reasoning models bill their hidden thinking as output (Google's table says 'Output price (including thinking tokens)'). Measure a few real runs before trusting any estimate, including these.
No provider charges a monthly fee, but most want credit up front. Google's minimum is $5 and its credits expire after 12 months; DeepSeek's top-ups never expire and can be refunded; xAI rejects requests once prepaid credit runs out; Mistral's free mode has no API access. Google's Free tier covers Gemini 3.8 Flash and 3.5 Flash-Lite at no charge, but there your content is 'used to improve our products'.
The model can cost more or less than the machine. On a small model the medium workload is £3.36 to £28.02 a month, while the server the agent runs on (2 vCPU / 4 GB) is £7.40 to £19.99 a month before VAT. On a top model the medium workload is £18.30 to £336.20.
When the cheapest is not the answer
- You need the best answers, not the cheapest. The tiers are each provider's own labels, not equal in ability: Mistral Large 4 ($0.70 / $2.10 per 1M tokens) and Claude Fable 5.1 ($10 / $50) are both in the top tier because each is its maker's flagship. This page compares price; quality needs the providers' own benchmarks or, better, your own tests.
- Your data has to stay in Europe. Mistral's regional endpoints include an EU one (10% extra). Anthropic's, xAI's and Together AI's are US-only, Google's Gemini API pricing page shows none, and OpenAI lists its data-residency regions on a separate page this research didn't check.
- You need guaranteed throughput. Together AI's serverless models are "best-effort", with guaranteed throughput sold separately, and Google and Mistral show their per-model rate limits only inside their consoles.
- Your prompts are confidential. On Google's Free tier your content is "used to improve our products"; on its paid tier it isn't.
Who made this page
An AI agent read each provider's pricing page and docs in a real browser set to UK English, opened the batch, priority, regional and rate-limit tabs with the page's own controls, and copied the sentence behind every figure; code checked each sentence was really on the page before it could be used. No model API was called to make this page.
Model prices change often, so the page is re-checked every month. There are no affiliate links on this page and nothing on it earns a commission.
How the numbers are worked out
monthly cost = input tokens (millions) × input price per 1M + output tokens (millions) × output price per 1M, at the standard list rate
- Light, 2M input + 0.5M output: roughly 500 calls a month at about 4,000 tokens in and 1,000 out; a prototype, a personal assistant or one agent that runs a few times an hour
- Medium, 20M + 5M: about 5,000 calls a month (around 170 a day); a small production app or a team's internal agent
- Heavy, 200M + 50M: about 50,000 calls a month (around 1,700 a day); a customer-facing app or a busy fleet of agents
Input is four times output because agents re-send their instructions, tools and history on every step. The figures leave out batch, priority, regional and long-prompt rates, and assume no cache hits; each line says where those apply. Each row is the model the provider's own pages name today for that tier, in its own words. Google's figures are its prices from 1 December 2026 (Gemini 4 Argon) and 1 January 2027 (Gemini 3.8 Flash), with the lower prices until then in the note; DeepSeek's are its peak rate. No middle model at DeepSeek (it lists two models), Mistral (Medium 3.5 was deprecated on 31 October 2026) or xAI (it recommends Grok 4.7 for everything); xAI's small row is an older model because it names no small one. Together AI hosts open models; its three rows are its own "Recommended models" picks.
VAT. All prices are in US dollars, as the providers list them (Mistral also lists euros). None of these token price tables says whether VAT is included.
| Model | Light: 2M + 0.5M | Medium: 20M + 5M | Heavy: 200M + 50M |
|---|---|---|---|
| OpenAI — top: GPT-6 AstraPrompts over 272K tokens cost $20 / $75. Batch 50% off. Cached input $1.00 (90% off); writing to the cache costs extra. Data residency +10%. The heavy workload is over the $500 monthly cap of OpenAI's Build tier | £33.62OpenAI: 'our flagship model for complex reasoning and coding'. 2M input × $10.00 + 0.5M output × $50.00 = $45.00 · $45.00 | £336.20OpenAI: 'our flagship model for complex reasoning and coding'. 20M input × $10.00 + 5M output × $50.00 = $450.00 · $450.00 | £3,362.00OpenAI: 'our flagship model for complex reasoning and coding'. 200M input × $10.00 + 50M output × $50.00 = $4,500.00 · $4,500.00 |
| OpenAI — mid: GPT-6.1 SolPrompts over 272K tokens cost $4 / $15. Batch 50% off. Cached input $0.10 (95% off). OpenAI also sells GPT-6.1 Terra at $1 / $5 ($4.50 / $45 / $450 at the three workloads). The heavy workload is over the Build tier's $500 cap | £6.72OpenAI: 'Choose GPT-6.1 Sol to balance intelligence and cost'. 2M input × $2.00 + 0.5M output × $10.00 = $9.00 · $9.00 | £67.24OpenAI: 'Choose GPT-6.1 Sol to balance intelligence and cost'. 20M input × $2.00 + 5M output × $10.00 = $90.00 · $90.00 | £672.40OpenAI: 'Choose GPT-6.1 Sol to balance intelligence and cost'. 200M input × $2.00 + 50M output × $10.00 = $900.00 · $900.00 |
| OpenAI — small: GPT-6 LunaPrompts over 272K tokens cost $0.20 / $0.75. Batch 50% off. Cached input $0.01 (90% off) | £0.34OpenAI: 'for cost-sensitive, high-volume workloads'. 2M input × $0.10 + 0.5M output × $0.50 = $0.45 · $0.45 | £3.36OpenAI: 'for cost-sensitive, high-volume workloads'. 20M input × $0.10 + 5M output × $0.50 = $4.50 · $4.50 | £33.62OpenAI: 'for cost-sensitive, high-volume workloads'. 200M input × $0.10 + 50M output × $0.50 = $45.00 · $45.00 |
| Anthropic — top: Claude Fable 5.1The full 1M-token context is at the standard price. Batch 50% off. Cache read $0.25 (97.5% off); cache writes cost extra. US-only inference +10%. The heavy workload is over the Start and Build tiers' caps ($500 and $1,000 a month) | £33.62Anthropic: 'Fable, our top model'. 2M input × $10.00 + 0.5M output × $50.00 = $45.00 · $45.00 | £336.20Anthropic: 'Fable, our top model'. 20M input × $10.00 + 5M output × $50.00 = $450.00 · $450.00 | £3,362.00Anthropic: 'Fable, our top model'. 200M input × $10.00 + 50M output × $50.00 = $4,500.00 · $4,500.00 |
| Anthropic — mid: Claude Sonnet 5.5This page's pick for the middle tier, by judgement: Anthropic's own suggested default is Claude Opus 5.5 ('start with Claude Opus 5.5 for most workloads') at $4 / $20, which would cost $18 / $180 / $1,800 at the three workloads. Batch 50% off. Cache read $0.10 (95% off). The heavy workload is over the Start tier's $500 cap | £6.72Anthropic: 'The best combination of speed and intelligence'. 2M input × $2.00 + 0.5M output × $10.00 = $9.00 · $9.00 | £67.24Anthropic: 'The best combination of speed and intelligence'. 20M input × $2.00 + 5M output × $10.00 = $90.00 · $90.00 | £672.40Anthropic: 'The best combination of speed and intelligence'. 200M input × $2.00 + 50M output × $10.00 = $900.00 · $900.00 |
| Anthropic — small: Claude Haiku 5.5A request whose prompt is over 100,000 tokens, cached tokens included, pays $0.50 / $2.50 for the whole request: five times as much. Batch 50% off. Cache read $0.01 (90% off) | £0.34Anthropic: 'Fastest, most cost-efficient model'. 2M input × $0.10 + 0.5M output × $0.50 = $0.45, for prompts up to 100,000 tokens · $0.45 | £3.36Anthropic: 'Fastest, most cost-efficient model'. 20M input × $0.10 + 5M output × $0.50 = $4.50, for prompts up to 100,000 tokens · $4.50 | £33.62Anthropic: 'Fastest, most cost-efficient model'. 200M input × $0.10 + 50M output × $0.50 = $45.00, for prompts up to 100,000 tokens · $45.00 |
| Google — top: Gemini 4 ArgonUntil 30 November 2026 Google lists half, $2 / $10 ($9 / $90 / $900 at the three workloads). Not on the Free tier. Batch 50% off. Cached input $0.20 (95% off). The heavy workload is over Tier 1's $250 monthly cap | £13.45Google: 'Our flagship reasoning model'. 2M input × $4.00 + 0.5M output × $20.00 = $18.00, the price from 1 December 2026 · $18.00 | £134.48Google: 'Our flagship reasoning model'. 20M input × $4.00 + 5M output × $20.00 = $180.00, the price from 1 December 2026 · $180.00 | £1,344.80Google: 'Our flagship reasoning model'. 200M input × $4.00 + 50M output × $20.00 = $1,800.00, the price from 1 December 2026 · $1,800.00 |
| Google — mid: Gemini 3.8 FlashUntil 31 December 2026 Google lists half, $0.75 / $3.75 ($3.38 / $33.75 / $337.50 at the three workloads). Free on the Free tier, where content is 'used to improve our products'. Batch 50% off. Cache storage $1 per 1M tokens an hour. The heavy workload is over Tier 1's $250 cap | £5.04Google: 'Our most intelligent Flash model'. 2M input × $1.50 + 0.5M output × $7.50 = $6.75, the price from 1 January 2027 · $6.75 | £50.43Google: 'Our most intelligent Flash model'. 20M input × $1.50 + 5M output × $7.50 = $67.50, the price from 1 January 2027 · $67.50 | £504.30Google: 'Our most intelligent Flash model'. 200M input × $1.50 + 50M output × $7.50 = $675.00, the price from 1 January 2027 · $675.00 |
| Google — small: Gemini 3.5 Flash-LiteGoogle's older Gemini 3.1 Flash-Lite is cheaper ($0.25 / $1.50), but Google tells new projects to use 3.5 Flash-Lite or 3.8 Flash. Free on the Free tier, where content is 'used to improve our products'. Batch 50% off | £1.38Google: 'Our fastest, most cost-effective 3.5 model'. 2M input × $0.30 + 0.5M output × $2.50 = $1.85 · $1.85 | £13.82Google: 'Our fastest, most cost-effective 3.5 model'. 20M input × $0.30 + 5M output × $2.50 = $18.50 · $18.50 | £138.22Google: 'Our fastest, most cost-effective 3.5 model'. 200M input × $0.30 + 50M output × $2.50 = $185.00 · $185.00 |
| DeepSeek — top: DeepSeek V4 ProOff-peak, every hour outside 01:00-04:00 and 06:00-10:00 UTC on weekdays (about 79% of the week), is half: $0.66 / $1.98 ($2.31 / $23.10 / $231). An agent spread evenly over the week pays about 60% of the peak figure. No batch discount. Cache hit $0.044 (97% off). DeepSeek lists no middle model. In the UK the brand is now 'DSeek' | £3.45The dearer of the two models DeepSeek lists (it uses no tier words). 2M input × $1.32 + 0.5M output × $3.96 = $4.62 at the peak rate · $4.62 | £34.52The dearer of the two models DeepSeek lists (it uses no tier words). 20M input × $1.32 + 5M output × $3.96 = $46.20 at the peak rate · $46.20 | £345.16The dearer of the two models DeepSeek lists (it uses no tier words). 200M input × $1.32 + 50M output × $3.96 = $462.00 at the peak rate · $462.00 |
| DeepSeek — small: DeepSeek V4.1 FlashOff-peak is half: $0.15 / $0.60 ($0.60 / $6 / $60). No batch discount. Cache hit $0.006 (98% off) | £0.90The cheaper of the two models DeepSeek lists (it uses no tier words). 2M input × $0.30 + 0.5M output × $1.20 = $1.20 at the peak rate · $1.20 | £8.97The cheaper of the two models DeepSeek lists (it uses no tier words). 20M input × $0.30 + 5M output × $1.20 = $12.00 at the peak rate · $12.00 | £89.65The cheaper of the two models DeepSeek lists (it uses no tier words). 200M input × $0.30 + 50M output × $1.20 = $120.00 at the peak rate · $120.00 |
| Mistral — top: Mistral Large 4No middle tier: Mistral Medium 3.5 was deprecated on 31 October 2026 and its named replacement is Large 4. Batch 50% off. Cached input $0.07 (90% off). EU or US regional endpoint +10%. Euro list €0.595 / €1.785. The free mode has no API access | £1.83Mistral's largest generalist model: 'A state-of-the-art, open-weight, general-purpose multimodal model'. 2M input × $0.70 + 0.5M output × $2.10 = $2.45 · $2.45 | £18.30Mistral's largest generalist model: 'A state-of-the-art, open-weight, general-purpose multimodal model'. 20M input × $0.70 + 5M output × $2.10 = $24.50 · $24.50 | £183.04Mistral's largest generalist model: 'A state-of-the-art, open-weight, general-purpose multimodal model'. 200M input × $0.70 + 50M output × $2.10 = $245.00 · $245.00 |
| Mistral — small: Mistral Small 4Batch 50% off. Cached input $0.015 (90% off). Euro list €0.1275 / €0.51. Mistral's even smaller Ministral 3 models cost $0.10 to $0.20 per 1M tokens | £0.45Mistral: 'For cost-sensitive projects: Mistral Small'. 2M input × $0.15 + 0.5M output × $0.60 = $0.60 · $0.60 | £4.48Mistral: 'For cost-sensitive projects: Mistral Small'. 20M input × $0.15 + 5M output × $0.60 = $6.00 · $6.00 | £44.83Mistral: 'For cost-sensitive projects: Mistral Small'. 200M input × $0.15 + 50M output × $0.60 = $60.00 · $60.00 |
| xAI — top: Grok 4.7Once a prompt reaches 200K tokens the whole request costs double, $4 / $12. Cached input $0.50 (75% off, the smallest caching discount here). Batch 20% off. US endpoint +10%. xAI names no middle model: it recommends Grok 4.7 for everything but image, video and voice | £5.23xAI: 'Our flagship model for code and everything else'. 2M input × $2.00 + 0.5M output × $6.00 = $7.00, for prompts under 200K tokens · $7.00 | £52.30xAI: 'Our flagship model for code and everything else'. 20M input × $2.00 + 5M output × $6.00 = $70.00, for prompts under 200K tokens · $70.00 | £522.98xAI: 'Our flagship model for code and everything else'. 200M input × $2.00 + 50M output × $6.00 = $700.00, for prompts under 200K tokens · $700.00 |
| xAI — small: Grok 4.3Double once a prompt reaches 200K tokens ($2.50 / $5). Cached input $0.20 (84% off). Batch 20% off | £2.80An older model: xAI names no small one, and Grok 4.3 ('Fast, reliable model with strong tool calling') is the cheapest general model on its list. 2M input × $1.25 + 0.5M output × $2.50 = $3.75, for prompts under 200K tokens · $3.75 | £28.02An older model: xAI names no small one, and Grok 4.3 ('Fast, reliable model with strong tool calling') is the cheapest general model on its list. 20M input × $1.25 + 5M output × $2.50 = $37.50, for prompts under 200K tokens · $37.50 | £280.17An older model: xAI names no small one, and Grok 4.3 ('Fast, reliable model with strong tool calling') is the cheapest general model on its list. 200M input × $1.25 + 50M output × $2.50 = $375.00, for prompts under 200K tokens · $375.00 |
| Together AI — top: Kimi K3.1 ProTogether AI hosts open models; it has no Llama model on its list (Meta's open models there are now called Muse). Cached input $0.20 (90% off). US-only inference +25%. Its batch discount covers only Whisper, not chat models | £6.72Together's own pick for 'Chat' and 'Reasoning' on its Recommended models page. 2M input × $2.00 + 0.5M output × $10.00 = $9.00 · $9.00 | £67.24Together's own pick for 'Chat' and 'Reasoning' on its Recommended models page. 20M input × $2.00 + 5M output × $10.00 = $90.00 · $90.00 | £672.40Together's own pick for 'Chat' and 'Reasoning' on its Recommended models page. 200M input × $2.00 + 50M output × $10.00 = $900.00 · $900.00 |
| Together AI — mid: DeepSeek V4.1 FlashThe same model DeepSeek sells as deepseek-flash, at DeepSeek's peak price all day. Cached input $0.006. US-only inference +25% | £0.90Together: 'Mid-size general purpose'. 2M input × $0.30 + 0.5M output × $1.20 = $1.20 · $1.20 | £8.97Together: 'Mid-size general purpose'. 20M input × $0.30 + 5M output × $1.20 = $12.00 · $12.00 | £89.65Together: 'Mid-size general purpose'. 200M input × $0.30 + 50M output × $1.20 = $120.00 · $120.00 |
| Together AI — small: Qwen3.5 9BNo cached-input price, so no caching discount. US-only inference +25% | £0.35Together: 'Small and fast'. 2M input × $0.17 + 0.5M output × $0.25 = $0.465 · $0.47 | £3.47Together: 'Small and fast'. 20M input × $0.17 + 5M output × $0.25 = $4.65 · $4.65 | £34.74Together: 'Small and fast'. 200M input × $0.17 + 50M output × $0.25 = $46.50 · $46.50 |
Exchange rates. ECB euro reference rates of 6 November 2026 (the latest published when checked on Monday 9 November, before that day's 16:00 CET release): 1 EUR = 0.83415 GBP, 1 EUR = 1.1165 USD, so 1 USD = 0.74711 GBP.
Left out: Groq (as the open-model host) — its Llama models are 'Enterprise' with 'Contact Sales' pricing and its public per-token prices are only in its console docs; Together AI publishes a full open-model price list instead; Contact-sales and enterprise offers — quoted prices only (OpenAI Scale Tier and Reserved Capacity, Anthropic's Custom tier, Gemini Enterprise and provisioned throughput, Together provisioned throughput, xAI Enterprise); Gemini API Contributor tier — an invite-only, half-price paid tier in which your content is used to improve Google's products; Third-party models resold by Mistral — Mistral also resells Z.ai's GLM models and DeepSeek V4.1 Flash; only Mistral's own models are used for its rows; Grok 4.7 Fast — sold only inside Cursor and Grok Build, not on xAI's public API.
There are no affiliate links on this page. The only affiliate programme found among these providers is Google Cloud's, open only to affiliates in the US and Canada; this site hasn't joined it. The research is done by AI agents reading each provider's own pages. Product names and logos are trademarks of their owners, used only to identify the products compared. No analytics cookies, no tracking scripts, no third-party requests. A personal project, separate from the author's day job.
Questions people ask
How much does an AI model API cost per month?
On this page's medium workload (20M input + 5M output tokens a month) at standard list rates: from $4.50 (about £3.36) on OpenAI's GPT-6 Luna or Claude Haiku 5.5 to $450 (about £336.20) on GPT-6 Astra or Claude Fable 5.1. Middle-tier models cost $12 to $90.
Is DeepSeek's API cheaper at night?
Yes. DeepSeek charges its full rate only from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; every other hour, about 79% of the week, is half price. This page's figures use the peak rate, so an agent running evenly through the week pays about 60% of them.
Why does Gemini 4 Argon cost more from December?
Google lists it at $2 input / $10 output per 1M tokens until 30 November 2026 and $4 / $20 from 1 December. This page uses the later price: $180 a month on the medium workload, against $90 at today's price. Gemini 3.8 Flash doubles on 1 January 2027, from $0.75 / $3.75 to $1.50 / $7.50.
Can a new API account run 200 million tokens a month?
Not on OpenAI's, Anthropic's or Google's top or middle models. New accounts there have spend caps: OpenAI's Build tier stops at $500 a month, Anthropic's Start tier at $500 and Google's Tier 1 at $250 a month. This page's heavy workload (200M input + 50M output tokens) costs $4,500 on GPT-6 Astra or Claude Fable 5.1, $1,800 on Gemini 4 Argon and $675 to $900 on those providers' middle models.
Answers use the figures on this page, last checked 9 Nov 2026.