Gemini 3.8 Flash Is Live: API Access Ahead of the Docs

​

Key Points

• Available on the API first: Google put out gemini-3.8-flash on September 2, 2026, roughly three weeks after 3.7 Flash. Google’s own model docs and launch blog do not list this version yet — APIYI has it open for calls
• Priced identically to 3.7 Flash: $0.75 in / $3.75 out per 1M tokens, $0.0750 cached reads. Migrating from 3.7 changes nothing about your costs
• 150 pre-launch test cases: paired against gemini-3.7-flash across both protocols — core capability, reasoning, tool calling and multimodal all line up, with no regression unique to 3.8
• Two endpoints, two groups: both the OpenAI-compatible format and the Gemini native format work; the default and svip groups are both open
• Official numbers pending: third-party outlets report this version targets coding and agents and cuts verbose output, but no official benchmarks or specs have been published, so this article does not relay them

​

Background

​

Details

​

Pre-launch testing

• Function calling closes the loop: single call, result hand-back, parallel and sequential calling all completed; parallel calling returned two calls with intact arguments and IDs across two rounds
• Long context works: on a needle-in-context task with a 14,543-character prefix and a 128-token output cap, both models returned the same correct answer
• Structured output is byte-identical: given the same JSON Schema, both models returned exactly the same JSON
• Multimodal accounting matches: on the same image and the same video clip, IMAGE and VIDEO modality token counts came back digit-for-digit identical to 3.7 Flash
• Thinking tiers work: low / medium / high were all accepted, with thinking tokens scaling across the tiers

​

Specifications

​

Putting It to Work

​

Where it fits

• A flat migration from 3.7 Flash: same price, same protocols, same parameter shapes — change the model name in one place. This is the most direct use of this launch
• Coding and code review: the focus of three consecutive Flash releases, and reportedly pushed further here
• Agents and multi-step workflows: parallel tool calling tested stable, which suits orchestration that needs several calls back at once
• Long documents and multimodal batches: image and video understanding verified, with accounting identical to 3.7, so your existing cost estimates carry over

​

Code examples

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.apiyi.com/v1")

response = client.chat.completions.create(
    model="gemini-3.8-flash",
    messages=[
        {"role": "user", "content": "Write a Python HTTP client with retries and exponential backoff"}
    ]
)
print(response.choices[0].message.content)

curl -X POST "https://api.apiyi.com/v1beta/models/gemini-3.8-flash:generateContent" \
  -H "x-goog-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "Write a Python HTTP client with retries and exponential backoff"}]}]
  }'

​

Best practices

• Migrating from 3.7 means changing the model name only: request shape, parameters and response fields are unchanged; testing found a single field-set difference and zero type changes, so clients need no adaptation
• Match the thinking tier to the task: low for batch classification and extraction where latency matters, high for complex refactoring and multi-step agents
• Ramp gradually: with official specs unpublished, there is no authoritative figure for the context limit. If you lean heavily on long context, run a slice of traffic first and confirm the boundary before cutting over

​

Pricing and Availability

​

Stacking with top-up promotions

​

Bottom Line