Skip to content
trending

Guides

OpenAI vs Claude vs Gemini API: Which Should You Choose?

Compare OpenAI, Claude, and Gemini APIs: pricing, free tiers, SDK differences, and real tests of the same question across 6 models with timing and costs.

By 8 min read
OpenAI vs Claude vs Gemini API: Which Should You Choose?

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 120 reads

For beginners, Gemini API offers the cheapest start without a credit card (free tier in Google AI Studio). The cheapest paid model from the big three is GPT-6 Luna from OpenAI at $0.10 / $0.50 per million tokens. Claude from Anthropic is the most expensive when processing Polish text because its tokenizer splits Polish into the most pieces. Below, we compare three APIs on the same questions, with timing, costs, and code differences. This is part of the AI API from scratch learning path.

As of October 4, 2026

  • OpenAI: GPT-6 Luna $0.10 / $0.50, GPT-6.1 Sol $2 / $10, GPT-6 Astra $10 / $50; no free tier.
  • Anthropic: Claude Haiku 4.5 $1 / $5, Sonnet 5.5 $2 / $10, Opus 5.5 $4 / $20, Fable 5.1 $10 / $50; small starter credits.
  • Google: Gemini 3.8 Flash $0.75 / $3.75 (through end of 2026), Flash-Lite from $0.10 / $0.40; free tier for Flash models.
  • Prices per million tokens input / output, excluding VAT.

API Comparison Table

OpenAI API Claude API (Anthropic) Gemini API (Google)
Where to get key platform.openai.com platform.claude.com aistudio.google.com/apikey
Free start no small credits for new accounts yes, free tier for Flash models without card
Cheapest model GPT-6 Luna: $0.10 / $0.50 Haiku 4.5: $1 / $5 Gemini 2.5 Flash-Lite: $0.10 / $0.40
All-purpose model GPT-6.1 Sol: $2 / $10 Sonnet 5.5: $2 / $10 Gemini 3.8 Flash: $0.75 / $3.75
Most powerful available GPT-6 Astra: $10 / $50 Fable 5.1: $10 / $50 Gemini 3.1 Pro (preview): $2 / $12; Gemini 4 Argon in closed program
Context window 1.05M tokens 1M (Haiku 200k) 1.05M
Tokens for our Polish paragraph (439 characters) 126 (Luna) 214 (Sonnet 5.5), 159 (Haiku 4.5) 116 (3.8 Flash)
Python library openai anthropic google-genai
Key variable OPENAI_API_KEY ANTHROPIC_API_KEY GEMINI_API_KEY
Hard spending limit yes, per project (since July 2026) yes, per organization and workspace billing tier limits
Batch mode -50% -50% -50%

Our Test: Same Question to Six Models

We sent one question in Polish to two models from each company (mid-tier and budget), plus DeepSeek for comparison. Everything through a single OpenRouter key, so conditions are equal: same network, same code, measured time for complete response.

"""Same question to OpenAI, Anthropic, Google and DeepSeek models through one OpenRouter key."""
import os, time
from dotenv import load_dotenv
from openai import OpenAI

load_dotenv()
client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key=os.environ["OPENROUTER_API_KEY"])
QUESTION = "Wyjaśnij jedenastolatkowi w trzech zdaniach, co to jest klucz API."

for model in ("openai/gpt-6.1-sol", "anthropic/claude-sonnet-5.5", "google/gemini-3.8-flash",
              "openai/gpt-6-luna", "anthropic/claude-haiku-4.5", "deepseek/deepseek-v4-flash"):
    start = time.time()
    r = client.chat.completions.create(model=model, messages=[{"role": "user", "content": QUESTION}], max_tokens=2000)
    elapsed = time.time() - start
    u = r.usage
    print(f"n### {model}  ({elapsed:.1f} s, {u.prompt_tokens}+{u.completion_tokens} tokens, ${u.cost:.5f})")
    print(r.choices[0].message.content.strip())

Note: We tested this in Polish. The question was: “Wyjaśnij jedenastolatkowi w trzech zdaniach, co to jest klucz API.“ (Explain to an eleven-year-old in three sentences what an API key is.)

### openai/gpt-6.1-sol  (3.4 s, 29+90 tokens, $0.00096)
Klucz API to specjalny kod, dzięki któremu aplikacja może korzystać z usług innej aplikacji, na przykład pobierać prognozę pogody. Działa trochę jak przepustka: pokazuje, kto prosi o dostęp i czy ma na to pozwolenie. Trzeba go chronić jak hasło, żeby ktoś obcy nie korzystał z usług na twój koszt.

### anthropic/claude-sonnet-5.5  (4.0 s, 44+188 tokens, $0.00197)
Klucz API to taki tajny „hasło-bilet", który program pokazuje innemu programowi, żeby mógł od niego coś dostać, na przykład prognozę pogody albo mapę. Dzięki niemu ten drugi program wie, kto o to prosi i czy ma na to pozwolenie. To trochę jak karta do szatni lub biblioteki: bez niej cię nie obsłużą, a jeśli ktoś ją ukradnie, może udawać ciebie, dlatego nie wolno jej nikomu pokazywać.

### google/gemini-3.8-flash  (6.3 s, 21+830 tokens, $0.00313)
Klucz API to specjalne, tajne hasło, które działa jak cyfrowa przepustka łącząca dwa różne programy lub gry. 

Dzięki niemu jedna aplikacja wie, że ta druga ma pozwolenie na pobranie od niej potrzebnych informacji, na przykład mapy czy prognozy pogody. 

Można to porównać do biletu w parku rozrywki, który pokazuje obsłudze, że masz prawo wejść na daną atrakcję i z niej skorzystać.

### openai/gpt-6-luna  (2.0 s, 29+111 tokens, $0.00006)
Klucz API to specjalny ciąg znaków, dzięki któremu jedna aplikacja może bezpiecznie korzystać z usług innej, na przykład pobierać prognozę pogody. Działa trochę jak przepustka: usługa rozpoznaje, kto się z nią łączy, i sprawdza, na co mu pozwolić. Trzeba go ukrywać, tak jak hasła, bo ktoś inny mógłby użyć go bez twojej zgody.

### anthropic/claude-haiku-4.5  (2.7 s, 34+138 tokens, $0.00072)
# Klucz API w trzech zdaniach

Klucz API to specjalny kod, który pozwala twojej aplikacji porozmawiać z inną aplikacją lub serwisem internetowym - tak jak hasło, ale do komunikacji między programami. Dzięki niemu serwis wie, że to ty go prosisz o informacje, i może ci je wysłać. Musisz go chronić jak hasło, bo jeśli ktoś go ukradnie, może udawać, że to ty i używać twojego konta.

### deepseek/deepseek-v4-flash  (6.0 s, 29+356 tokens, $0.00002)
Wyobraź sobie, że klucz API to tajny kod, który pozwala twojej grze lub aplikacji rozmawiać z innym programem w internecie, na przykład poprosić o pogodę. Zamiast loginu i hasła, program używa tego kodu, aby udowodnić, że ma pozwolenie na korzystanie z danych - podobnie jak klucz do domku na drzewie. Dlatego nigdy nie wolno go nikomu podawać, bo ktoś obcy mógłby go użyć do kradzieży informacji lub zrobienia czegoś złego w twoim imieniu.
Model Time Output Tokens Cost Our Assessment
GPT-6.1 Sol 3.4 s 90 $0.00096 exactly 3 sentences, correct Polish, good pass analogy
Claude Sonnet 5.5 4.0 s 188 $0.00197 most vivid imagery (coat check, library), one odd phrasing
Gemini 3.8 Flash 6.3 s 830 $0.00313 good, but most expensive: most tokens are thinking
GPT-6 Luna 2.0 s 111 $0.00006 nearly as good as Sol, at 1/16th the price
Claude Haiku 4.5 2.7 s 138 $0.00072 correct, but added unnecessary Markdown heading
DeepSeek V4 Flash 6.0 s 356 $0.00002 cheapest, correct, slightly exaggerated “information theft”

Conclusions from this single question (this is not a benchmark, just a demonstration of how to compare):

  • All six models write correct Polish. Differences are in style and format discipline, not grammar.
  • Cost per query varied 150x (from $0.00002 to $0.00313) with similar answer quality.
  • The “per million tokens” pricing is misleading. Gemini 3.8 Flash has a lower rate than Sonnet 5.5, but came out most expensive because it thinks by default and used 830 output tokens for three sentences.
  • A broader test of models in Polish across many tasks can be found in GPT-6 Astra vs Opus 5.5: 9 tests.

Code Differences Between APIs

The three companies have three different API formats. The most important differences you’ll see in your first script:

  • OpenAI recommends Responses API: client.responses.create(model=..., instructions=..., input=...), text in output_text. The older Chat Completions still works and is the market standard.
  • Anthropic: client.messages.create(model=..., max_tokens=..., system=..., messages=[...]). max_tokens is required, and the response is a list of blocks (text, thinking, tool calls).
  • Google: client.models.generate_content(model=..., contents=..., config=...), text in .text, system instruction in config.

Complete, minimal scripts for each company are in the API Key: how to get one article and in the ai-api-starter package (dostawcy/ directory). To be honest: we ran them with fake keys (each company’s servers responded with authorization errors, so the code and libraries work), but we haven’t tested them with real keys from all three companies. An interesting finding from this test: OpenAI and Anthropic respond to bad keys with code 401, while Google uses code 400.

If you don’t want to write three versions of code, use an intermediary that speaks one format: OpenRouter or libraries that support multiple providers at once. That’s what we did in all tests in this learning path.

Polish Language: Tokens and Quality

All three companies understand and write Polish well. The difference shows up in the bill: the same Polish paragraph is 116 tokens in Gemini 3.8 Flash, 126 in GPT-6 Luna, and 214 in Claude Sonnet 5.5. At the same rate per million tokens (Sonnet 5.5 and GPT-6.1 Sol both at $2 / $10), Polish text in Claude therefore costs roughly 1.7 times more. A full measurement across seven models is in the article AI Tokens: what they are and what they cost.

Which API Should You Choose?

  • Learning and don’t want to pay: Gemini API free tier or free models on OpenRouter. Guide: Google AI Studio.
  • Large-scale automation, simple tasks: GPT-6 Luna or Gemini Flash-Lite - pennies for thousands of queries.
  • Text, code, and agents where quality matters: Claude Sonnet 5.5 or GPT-6.1 Sol. Choose after testing on your own examples.
  • Most difficult tasks: Claude Opus 5.5, Fable 5.1, or GPT-6 Astra. Pricing and pitfalls: Claude Opus 5.5 in API, GPT-6 Sol and Luna, GPT-6 Astra.
  • Not sure yet: start with OpenRouter, compare models on your data, then open an account with the winner.

Beyond the big three, it’s worth knowing cheaper alternatives: DeepSeek, Mistral (data in Europe), Kimi, and Grok.

Frequently Asked Questions

Which AI API is cheapest?

From the big three: GPT-6 Luna ($0.10 / $0.50) and Gemini 2.5 Flash-Lite ($0.10 / $0.40). Even cheaper are DeepSeek outside peak hours and open models through OpenRouter. For free: Gemini API free tier.

Is Gemini API free?

In the free tier yes, for Flash and Flash-Lite models, without a card. Google doesn’t publish fixed limits, and in the free tier may use data to improve services (with exceptions for EEA users). Pro models are paid only.

How much does Claude API cost?

From $1 / $5 per million tokens (Haiku 4.5) through $2 / $10 (Sonnet 5.5) and $4 / $20 (Opus 5.5) to $10 / $50 (Fable 5.1). Batch mode gives 50% discount, and cache reads cost 10% of input price.

Can I use one code for all three?

Yes, through an intermediary compatible with OpenAI format, e.g. OpenRouter. Google and Anthropic also provide their own endpoints compatible with OpenAI SDK, but with feature limitations.

Which API writes Polish best?

In our tests, differences in Polish language quality are small, with larger differences in style and format adherence. The difference in cost for Polish text is clear though: Claude uses the most tokens, Gemini the least.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Date

We verified pricing on October 4, 2026. We ran the test the same day on Windows 11, Python 3.13.15, openai 3.24.0, through OpenRouter; total cost of the comparison $0.0069.

Kinetyra

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editor and publisher · Quantslant

I edit and publish Quantslant. AI tools help with research, drafts and translation; choosing topics, running the tests, checking facts and the decision to publish are mine, and so is the responsibility for what appears here.

read nextAI API from Scratch: How It Works, Keys, Costs & First RequestguidesAI API from Scratch: How It Works, Keys, Costs & First RequestLearn how AI APIs work: get your API key, understand token costs, send your first request in Python, and download 18 tested scripts to get started.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map

This page in other languages1

Quantslant in numbers 1,423posts in archive22news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›