LLM APIs · Comparison
OpenAI API vs Anthropic API vs Google Gemini API: Anthropic leads at 7.2; the same tokens cost $20 or $55 a month
We read the pricing, model, data-usage, trust and docs pages of three LLM APIs on the same day and scored them on the same six indicators. What 100 million input and 20 million output tokens a month cost, how each one treats your data, and who should pick which.
On this page9 sections
On six indicators read from each vendor's public pages, Anthropic API (Claude) scores 7.2, OpenAI API 7.1 and Google Gemini API 5.3 out of 10. For our reference scenario, 100 million input and 20 million output tokens a month on each vendor's lowest-priced generally available model with tool calling, OpenAI (GPT-6 Luna) and Anthropic (Claude Haiku 5.5) both come to $20 and Gemini (Gemini 3.1 Flash-Lite) to $55, computed from the pricing pages. Anthropic and OpenAI score the same on three of six items (pricing clarity, the scenario price, and data use and retention): Anthropic's lead comes from the free credits it states for new users, while OpenAI scores higher on compliance and regions and on cost controls. Gemini has the only ongoing free tier of the three.
Gemini is in the Fair band mainly because the certifications, zero data retention and region choice Google offers belong to a separate Google Cloud product, so it scores 0 on compliance and regions here. The scores measure what the public pages commit to, not model quality, and we did not sign up for any of them.
At a glance
| OpenAI API | Anthropic API (Claude) | Google Gemini API | |
|---|---|---|---|
| What you pay for | GPT-6 models and hosted tools | Claude models with a 1M-token context | Gemini models, free tier to start |
| How you pay | Prepaid credits, $5 minimum | Prepaid credits by card | Prepaid credits, $5 minimum |
| Free option | No free allowance stated | One-time credits, amount not stated | Ongoing free tier on select models |
| Reference scenario (100M in + 20M out) | $20 on gpt-6-luna, computed | $20 on claude-haiku-5-5, computed | $55 on gemini-3.1-flash-lite, computed |
| Discounts | Batch and Flex 50% off; cached input 90% off | Batch 50% off; cache hits 90% off | Batch and Flex (preview) 50% off; cached input 90% off |
| Trained on paid API data | No, unless you opt in | No | No (free tier: may be used) |
| SOC 2 / ISO 27001 | Both, covering the API | Both, listed for the API | Not stated for this API |
| HIPAA BAA | Accepted in the API Platform, if eligible | Signed in the Console, if eligible | Not stated |
| Data regions | 10 storage regions, by approval | US-only inference at 1.1x | No choice stated |
| Score (public pages) | 7.1 Good | 7.2 Good | 5.3 Fair |
What the public pages show
OpenAI API sells GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, plus image, audio and embedding models, and hosted tools in the Responses API (web search, file search, code containers) billed on top of tokens. Every GPT-6 and GPT-5.6 model has Standard, Batch, Flex and Fast prices on one page. Credits are prepaid ($5 minimum, auto-reload on by default) and expire after a year.
Anthropic API (Claude) sells Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1, each with a 1M-token context and tool use. The pricing page lists input, cache write, cache hit, output and batch prices for every model; Haiku 5.5 costs more for prompts over 100,000 tokens. Usage is paid from prepaid credits by card, and spend caps rise by usage tier.
Google Gemini API is the Gemini Developer API, with keys from Google AI Studio. Its pricing page has Free and Paid columns for every model and says on each whether content is used to improve Google products. A free tier with no end date covers select current models with no card; the Paid Tier starts with a $5 prepay, and Tier 1 is capped at $250 a month. Compliance controls belong to Gemini Enterprise Agent Platform, a separate Google Cloud offer.
Strengths and things to know
OpenAI API. Strengths: Batch and Flex at 50% off and cached input at 90% off; a SOC 2 Type 2 report and an ISO 27001 certificate that name the API; storage in 10 regions, by approval; hard spend limits per project; official SDKs in six languages. Things to know: no page states a free allowance; zero data retention, residency and console SSO go through approval or sales, and the self-serve BAA needs an established history of API usage; and some docs lag the price list, such as models still listed after their Sep 28, 2026 shutdown.
Anthropic API (Claude). Strengths: Haiku 5.5 at $0.10 / $0.50 per 1M tokens; batch and caching discounts that can be combined; Commercial Terms that rule out training on customer content; a standard HIPAA BAA eligible organizations sign in the Console; SOC 2 Type II and ISO 27001 listed for the API; seven official SDKs. Things to know: the free credit amount is not stated; US-only inference at 1.1x is the only region option; zero data retention and Console SSO go through sales; and the pricing page and help center describe billing differently.
Google Gemini API. Strengths: an ongoing free tier on current models with no card; Batch and Flex (preview) at 50% off and cached input at 90% off; paid prompts not used to improve Google products; SDKs in five languages and an OpenAI-compatible endpoint. Things to know: Google Cloud's SOC 2, ISO 27001 and HIPAA lists name Gemini Enterprise Agent Platform, not this API; free-tier content may be used to improve Google products outside the EEA, Switzerland and the UK; paid prompts are logged for a period the terms do not state; and per-model rate limits are shown only in AI Studio.
Pricing reality
All three bill per 1M input and output tokens in US dollars from prepaid credits, with no plan fee. Reference scenario: 20,000 requests a month of 5,000 input and 1,000 output tokens, on the lowest-priced generally available model with tool calling and a context of at least 128,000 tokens, at standard rates.
| API | Model | Monthly cost | Where the number comes from |
|---|---|---|---|
| OpenAI API | gpt-6-luna | $20 | Computed: 100 × $0.10 + 20 × $0.50 |
| Anthropic API (Claude) | claude-haiku-5-5 | $20 | Computed: 100 × $0.10 + 20 × $0.50 (prompts up to 100,000 tokens) |
| Google Gemini API | gemini-3.1-flash-lite | $55 | Computed: 100 × $0.25 + 20 × $1.50 |
Three cheaper rows were not used: OpenAI's GPT-5 nano has only a deprecated snapshot and GPT-4.1 nano ($0.10 / $0.40, $18 computed) shuts down on Oct 23, 2026, and Gemini 2.5 Flash-Lite ($0.10 / $0.40) is limited to users who used 2.5 models before. Batch halves the scenario on all three: $10 on OpenAI and Anthropic, $27.50 on Gemini (our arithmetic); OpenAI's and Gemini's Flex rates give the same price. Anthropic's US-only inference makes it $22.
Costs climb with larger models: the same tokens on GPT-6 Astra ($10 / $50) would cost $2,000; Claude Sonnet 5.5 is $2 / $10; Gemini 4 Argon is $2 / $10, rising to $4 / $20 on Dec 1, 2026. Web search is $10 per 1,000 calls on OpenAI and per 1,000 searches on Anthropic; Gemini's Search grounding is $14 per 1,000 requests after 5,000 free a month. Anthropic's pricing page also notes that Claude 4.7 and later use a tokenizer that produces about 30% more tokens for the same text.
How we scored them
Every product in this category is scored on the same six indicators, read from the vendor's own pages, with fixed formulas. The overall score is the plain mean. Our methodology page explains how these scores work.
| Indicator | OpenAI API | Anthropic API (Claude) | Google Gemini API |
|---|---|---|---|
| Pricing clarity | 7 | 7 | 7 |
| Free tier | 0 | 3 | 6 |
| Price for 100M input + 20M output tokens | 7 | 7 | 4.8 |
| Data use and retention | 9 | 9 | 5 |
| Compliance and regions | 10 | 8.5 | 0 |
| Cost controls and developer tools | 9.3 | 8.8 | 8.8 |
| Overall | 7.1 | 7.2 | 5.3 |
OpenAI, Anthropic and Gemini each lose 3 points on pricing clarity for prices left out or contradictions between their pages. OpenAI scores 0 on the free tier because no page states a free allowance, and Anthropic 3 because its credits have no stated amount. Gemini's free tier scores 6, not 10, because its limits are shown only in AI Studio and free-tier prompts may be used. Its 0 on compliance reflects certifications stated for another Google Cloud product. The rubric does not score model quality, speed or how each API feels to use.
Also in this category
Mistral AI API scores 7.4, the highest in the category: the lowest scenario price ($12 on Ministral 3 3B), EU or US endpoints at +10% without a sales call and SOC 2 Type II and ISO 27001 stated in its help center, while training use is an opt-out and rate limits appear only after sign-in. DeepSeek API scores 3.2: $54 a month at peak ($27 off-peak), cache hits at 98% off and OpenAI and Anthropic request formats, with data processed and stored in China, training use as an opt-out and no compliance documents stated.
Who should choose which
- Choose Anthropic API (Claude) if you want a low-cost current model with a 1M-token context, combined batch and caching discounts, and a HIPAA BAA you can sign without a sales call. Look elsewhere if you need processing in the EU or another non-US region.
- Choose OpenAI API if you want hosted tools such as web search and file search in the same API, a flex rate on top of the batch discount, and audit reports that name the API. Look elsewhere if you need zero data retention or residency without a sales conversation.
- Choose Google Gemini API if you want to prototype on current models for free with no card, then pay per token on the same key. Look elsewhere if your review needs SOC 2, ISO 27001 or a HIPAA BAA that names the API itself.
Bottom line
From public pages alone, Anthropic and OpenAI are close: the same scenario price, the same data terms on paid use and strong compliance pages, with Anthropic ahead on its free credits and OpenAI on regions and cost levers; both offer a standard HIPAA BAA that eligible organizations accept without a sales call. Gemini is the easiest of the three to start with for free, but the enterprise controls Google offers sit with a separate Google Cloud product.
How we checked
We read the vendors' pricing, model, rate-limit, billing, data-usage, trust, legal and SDK pages on Oct 15, 2026. Some openai.com policy and help pages refused our fetches, so we read their Internet Archive copies (Jun 18 to Oct 12, 2026); Anthropic's trust center loads only with JavaScript, so its certifications come from its privacy center. All prices are US-dollar list prices. We did not sign up, create an API key or call any model, and facts the pages did not state are left out rather than guessed. This post was written by Claude, an AI model made by Anthropic, whose API is one of the three compared here; every product is scored by the same fixed formulas from its own pages. Every figure is on the vendors' pages as of that date or, where marked, computed from them.
