Gemini API Deprecates 3.7 Flash and Deep Research Agent
Google deprecated gemini-3.7-flash, gemini-3.5-flash, and deep-research-pro-preview-12-2025. Learn routing changes, reasoning levels, and how to migrate.
- Google announced the deprecation of gemini-3.7-flash and gemini-3.5-flash on 2026-10-08, removing them from Google AI Studio and rerouting API traffic automatically.
- gemini-3.7-flash traffic now routes to gemini-3.8-flash, while gemini-3.5-flash routes to gemini-3.6-flash.
- The deep-research-pro-preview-12-2025 agent will be shut down on October 23, 2026, requiring migration to deep-research-preview-04-2026 or deep-research-max-preview-04-2026.
- Gemini 3.8 Flash defaults to medium thinking effort, forbids the minimal thinking level, and drops legacy parameters like thinking_budget, temperature, top_p, and top_k.
Overview of the 2026-10-08 Gemini API Deprecations
On 2026-10-08, Google published an update to the Gemini API changelog introducing immediate deprecations and an upcoming shutdown across both core generation models and managed agents. The update affects two widely used intermediate Flash generation checkpoints and one specialized autonomous research agent. Production workloads pointing to the deprecated identifiers must be audited to prevent configuration errors and unexpected behavior.
Specifically, gemini-3.7-flash and gemini-3.5-flash have been declared deprecated and removed from interactive selection in Google AI Studio. Furthermore, the managed agent deep-research-pro-preview-12-2025 is deprecated and scheduled for full decommission on October 23, 2026. Teams relying on these identifiers must transition to the latest generation identifiers to maintain compatibility with modern platform capabilities.
- gemini-3.7-flash is deprecated and replaced by gemini-3.8-flash.
- gemini-3.5-flash is deprecated and replaced by gemini-3.6-flash.
- deep-research-pro-preview-12-2025 is scheduled for shutdown on October 23, 2026.
Flash Model Automatic Routing and Successor Models
For conversational and completion workloads, Google has implemented server-side rerouting so that incoming API requests do not break immediately. Calls targeting gemini-3.7-flash are automatically routed to gemini-3.8-flash. Similarly, requests addressed to gemini-3.5-flash are transparently redirected to gemini-3.6-flash. While traffic continues to process without immediate drop-offs, explicit migration is recommended across codebase configurations, as documented in our guide on migrating from Gemini 3.7 Flash to Gemini 3.8 Flash.
The primary production target for high-reasoning Flash workloads is now gemini-3.8-flash. Designed for long-horizon software engineering, autonomous agents, and multi-step enterprise workflows, it maintains a 1M token context window and 64k max output tokens. It features iterative planning and autonomous error checking that can take more reasoning steps than predecessor models. Developers seeking a direct equivalent with reduced token overhead can also select gemini-3.6-flash, while high-throughput, low-complexity use cases can evaluate gemini-3.5-flash-lite or gemini-3.1-flash-lite.
| Deprecated Model ID | Automatic Routing Target | Studio Availability | Primary Recommended Replacement |
|---|---|---|---|
| gemini-3.7-flash | gemini-3.8-flash | Removed | gemini-3.8-flash |
| gemini-3.5-flash | gemini-3.6-flash | Removed | gemini-3.6-flash (or gemini-3.8-flash) |
Sunset of Deep Research Agent Preview
The managed research agent checkpoint deep-research-pro-preview-12-2025 has reached formal deprecation and will be permanently shut down on October 23, 2026. Unlike the generation model IDs, which offer transparent traffic rerouting, managed agent endpoints reaching their shutdown date will no longer serve requests once offline. Any integration initiating requests via interactions.create with this agent identifier must be updated prior to the shutdown date.
Google provides two replacement options within the Deep Research framework depending on application latency and depth requirements:
- deep-research-preview-04-2026: Engineered for high speed and execution efficiency, structured for streaming synthesized output back to an active client UI.
- deep-research-max-preview-04-2026: Optimized for maximum comprehensiveness across automated context discovery, exhaustive aggregation, and deep report synthesis.
Migration requires replacing the string assigned to the agent parameter in your interactions.create call payload, as illustrated in the following REST configuration:
{
"agent": "deep-research-preview-04-2026",
"input": "Gather competitive landscape data for enterprise cloud infrastructure.",
"environment": "remote"
}
Reasoning Levels and Parameter Configuration Rules
When moving from older Gemini checkpoints to gemini-3.8-flash, developers must observe key behavioral differences in the reasoning engine. The legacy numerical parameter thinking_budget is no longer supported and must be replaced with the string enum thinking_level. Attempting to supply both thinking_level and thinking_budget in a single request results in a 400 error.
Available thinking levels vary depending on the model generation. In gemini-3.5-flash, four levels were supported: minimal, low, medium, and high. However, gemini-3.8-flash completely disallows minimal. Sending minimal to gemini-3.8-flash causes the API to return an explicit error. Production configurations should rely on the default medium setting for agentic programming tasks, low for latency-sensitive pipelines, or high for deep multi-step mathematical problems.
Additionally, Gemini 3.x architectures deprecate standard sampling parameters. Google advises removing temperature, top_p, and top_k completely, as internal reasoning mechanisms are calibrated to default configurations. Prefilled assistant turns and the legacy candidate_count parameter must also be excluded from modern payloads.
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-3.8-flash",
input="Analyze this payment processing pipeline for race conditions during retry attempts and rewrite the transaction locks safely.",
generation_config={
"thinking_level": "medium"
}
)
print(interaction.output_text)
Function Calling and Managed Agent Updates
Engineers employing tool use and function orchestration must adhere to Gemini 3 response validation rules. Every tool execution step requires strict matching between function calls and their corresponding outputs. When interacting with the API, each FunctionResponse or function result step must supply the exact call_id and function name returned during the preceding turn. Failing to supply matched identifiers leads to immediate validation errors or truncated stop events. For broader context on implementing schema-driven workflows, see our guides on function and tool calling across providers and structured output schemas.
Furthermore, managed systems such as the Antigravity agent in Gemini Managed Agents and the Google Antigravity SDK have been updated to run on gemini-3.8-flash by default. Multimodal assets intended for tool responses must be placed directly inside the response payload rather than as separate external parts, and conversational state must be preserved across rounds using server-side interaction tracking.
from google import genai
client = genai.Client()
interaction = client.interactions.create(
agent="antigravity-preview-09-2026",
input="Audit https://web.dev for performance, Core Web Vitals, and SEO.",
environment="remote",
)
print(interaction.output_text)
Pricing Structure and Cost Considerations
The deprecation transition also carries economic implications. gemini-3.8-flash operates on introductory pricing of $0.75 per 1M input tokens and $3.75 per 1M output tokens through December 31, 2026 across Google AI Studio and Gemini Enterprise Agent Platform. On January 1, 2027, standard rates of $1.50 per 1M input tokens and $7.50 per 1M output tokens take effect. For long-term budget modeling, consult our breakdown of the Gemini 3.8 Flash 2027 price adjustment and the LLM API cost calculator.
Because gemini-3.8-flash engages in deeper reasoning, iterative planning, and deterministic tool execution, complex queries may naturally consume more output tokens than equivalent prompts sent to earlier Flash variants. Teams needing lower operational costs for routine extraction tasks can evaluate lightweight variants such as gemini-3.5-flash-lite or benchmark costs across models via our LLM API pricing reference.
| Model Name | Model ID | Input (per 1M tokens) | Cached Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|---|
| Gemini 3.8 Flash (Introductory) | gemini-3.8-flash | $0.75 | $0.075 | $3.75 |
| Gemini 3.5 Flash (Deprecated) | gemini-3.5-flash | $1.50 | $0.15 | $9 |
| Gemini 3.5 Flash-Lite | gemini-3.5-flash-lite | $0.3 | $0.03 | $2.5 |
| Gemini 3.1 Flash-Lite | gemini-3.1-flash-lite | $0.25 | $0.025 | $1.5 |
| Gemini 3.1 Pro Preview | gemini-3.1-pro-preview | $2 | $0.2 | $12 |
Engineering Migration Checklist
To guarantee smooth operational continuity following the 2026-10-08 changelog release, teams should execute the following technical checklist across all Gemini integration layers:
- Update primary Flash endpoints: Switch model strings from gemini-3.7-flash to gemini-3.8-flash, and from gemini-3.5-flash to gemini-3.6-flash or gemini-3.8-flash.
- Replace research agent endpoints: Update interactions.create calls containing deep-research-pro-preview-12-2025 to deep-research-preview-04-2026 or deep-research-max-preview-04-2026 before October 23, 2026.
- Configure thinking levels correctly: Purge thinking_budget numeric fields, adopt thinking_level string enums (low, medium, high), and ensure minimal is not passed to gemini-3.8-flash.
- Scrub deprecated sampling controls: Eliminate temperature, top_p, top_k, candidate_count, and prefilled assistant turns from all request payloads.
- Validate function calling formats: Confirm that all tool execution responses supply exact call_id and name matches, embed multimodal assets inside payloads, and separate instructions with clean breaks.