APIYI fully supports the official Gemini native format (the /v1beta generateContent endpoint): point the base_url at https://api.apiyi.com and your existing Gemini code and official SDKs migrate seamlessly — no format conversion needed.
This page is based on the official Google documentation (ai.google.dev/gemini-api/docs, as of June 2026). All examples are copy-paste ready.
The OpenAI-compatible format can call Gemini too, but the following are native-only:
- Full thinking controls:
thinking_level (Gemini 3 series) / thinking_budget (2.5 series), thought summaries, thought signatures
- Native multimodal Parts: inline images / audio / video, with
media_resolution cost control — see Multimodal & Code Execution
- Code execution tool:
code_execution runs Python in a sandbox
- Fine-grained usage fields:
thoughts_token_count, cached_content_token_count, and more
For plain text chat, or one codebase across multiple vendors, use OpenAI Compatible Mode instead.
Quick Start
Use Google’s official unified SDK google-genai (the legacy google-generative-ai was sunset on November 30, 2025 (UTC)):
from google import genai
client = genai.Client(
api_key="YOUR_API_KEY", # your APIYI key
http_options={"base_url": "https://api.apiyi.com"}
)
response = client.models.generate_content(
model="gemini-3.5-flash",
contents="Introduce yourself in one sentence"
)
print(response.text)
The base_url is https://api.apiyi.com (without /v1) — different from the OpenAI-compatible format’s https://api.apiyi.com/v1. Use your APIYI key, not a Google AI Studio key.
Streaming
stream = client.models.generate_content_stream(
model="gemini-3.5-flash",
contents="Write a short essay on quantum computing"
)
for chunk in stream:
print(chunk.text, end="", flush=True)
Thinking Controls
Gemini models think by default, and the two generations use different parameters — mixing them raises an error:
| Model series | Parameter | Values |
|---|
| Gemini 3 / 3.1 / 3.5 | thinking_level | minimal (Flash family only) / low / high (default) |
| Gemini 2.5 | thinking_budget | token cap (e.g. 0–8192); model auto-controls if unset |
Passing both thinking_level and thinking_budget to a Gemini 3 series model returns an error — pick one (use thinking_level for the 3 series).
from google.genai import types
# Gemini 3 series: level-based control
response = client.models.generate_content(
model="gemini-3.5-flash",
contents="Prove that the square root of 2 is irrational",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(thinking_level="high")
)
)
# Gemini 2.5 series: token-budget control
response = client.models.generate_content(
model="gemini-2.5-flash",
contents="Simple question, be quick",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(thinking_budget=0) # thinking off
)
)
minimal for low-latency simple tasks (classification, extraction); low for routine chat; high for complex reasoning and code. Thinking tokens bill at the output rate — higher levels cost more.
Thought summaries and thought signatures
- Thought summaries:
include_thoughts=True returns a summary of the reasoning (parts where part.thought is True)
- Thought signatures: encrypted reasoning state introduced with Gemini 3. In multi-turn conversations (especially function calling), pass the
thought_signature from the response back unchanged so the model can continue its reasoning chain. Official SDKs handle this automatically; don’t strip the field in hand-written REST calls — see Function Calling
response = client.models.generate_content(
model="gemini-3.5-flash",
contents="Analyze the time complexity of: def fib(n): return n if n <= 1 else fib(n-1) + fib(n-2)",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(thinking_level="high", include_thoughts=True)
)
)
for part in response.candidates[0].content.parts:
if getattr(part, "thought", False):
print(f"[Thought summary] {part.text}")
else:
print(f"[Final answer] {part.text}")
Common Config Parameters
Passed via config (GenerateContentConfig):
| Parameter | Description |
|---|
system_instruction | System prompt |
temperature | 0–2. Google recommends keeping the default 1.0 for Gemini 3 series — lowering it can hurt reasoning quality |
max_output_tokens | Output cap (includes thinking tokens) |
thinking_config | Thinking controls, see above |
response_mime_type | Set application/json to force JSON output |
response_schema | Schema constraint for structured JSON output |
tools | Function declarations / code_execution and other tools |
media_resolution | Multimodal input cost control, see Multimodal page |
usage = response.usage_metadata
print(f"Input: {usage.prompt_token_count}")
print(f"Output: {usage.candidates_token_count}")
print(f"Thinking: {usage.thoughts_token_count}")
print(f"Cache hits: {usage.cached_content_token_count}")
| Field | Description | Billing |
|---|
prompt_token_count | Input tokens | Input rate |
candidates_token_count | Output tokens | Output rate |
thoughts_token_count | Thinking tokens | Output rate — tune the level to save |
cached_content_token_count | Cached tokens | Official discount, see Cache Billing |
total_token_count | Total | — |
Supported Models and Pricing
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Notes |
|---|
gemini-3.5-flash | $1.50 | $9.00 | Current workhorse — beats 3.1 Pro on several benchmarks, 1M context |
gemini-3.1-pro-preview | $1.80 | $10.80 | Pro flagship |
gemini-3-pro-preview | $1.80 | $10.80 | Previous Pro |
gemini-3-flash-preview | $0.44 | $2.64 | Light and fast |
gemini-3.1-flash-lite | $0.25 | $1.50 | Ultra-budget |
gemini-2.5-pro | $1.25 | $10.00 | 2.5-series Pro |
gemini-2.5-flash | $0.30 | $2.40 | 2.5-series workhorse |
gemini-2.5-flash-lite | $0.10 | $0.40 | Cheapest |
Some models have -thinking / -nothinking alias variants (e.g. gemini-3-flash-preview-nothinking) that pin thinking on/off — handy for clients where you can’t change request parameters. Full list: Models & Pricing. Native vs OpenAI-Compatible
| Feature | Gemini native | OpenAI-compatible |
|---|
| base_url | https://api.apiyi.com | https://api.apiyi.com/v1 |
| SDK | google-genai | openai |
| Thinking control | thinking_level / thinking_budget | reasoning_effort |
| Thought summaries / signatures | ✅ | ❌ |
| Code execution tool | ✅ | ❌ |
| Media input | Native inline Parts (PIL / bytes) | Base64 image_url |
| Cache hit field | cached_content_token_count | cached_tokens |
Notes
- The Files API is not supported (
client.files.upload()); media must be passed inline and each file must stay under 20MB — see Multimodal & Code Execution
- Cache discounts and hit-rate expectations: Cache Billing
