Gemini 3.7 Flash Deprecation: Migrating to Gemini 3.8 Flash

API Changes Announced on 2026-10-08

The Gemini developer platform received a series of model retirements and routing updates announced on 2026-10-08 in the Gemini API changelog. The update deprecates gemini-3.7-flash alongside gemini-3.5-flash while establishing a retirement date for the initial preview generation of the managed Deep Research agent.

For standard text and multimodal requests, the backend provides automatic server-side forwarding. Queries directed to gemini-3.7-flash route directly to gemini-3.8-flash. Concurrently, requests specifying gemini-3.5-flash route to gemini-3.6-flash. While existing workloads will not fail immediately due to these routing rules, explicit model migration is necessary to align parameter configurations, adjust billing expectations, and avoid configuration errors.

• Gemini 3.7 Flash (gemini-3.7-flash) is deprecated and automatically routed to Gemini 3.8 Flash (gemini-3.8-flash).
• Gemini 3.5 Flash (gemini-3.5-flash) is deprecated and automatically routed to Gemini 3.6 Flash (gemini-3.6-flash).
• The managed agent deep-research-pro-preview-12-2025 is deprecated with a final shutdown date of October 23, 2026.
• Default agent execution for the Antigravity agent now targets Gemini 3.8 Flash by default.

Gemini 3.8 Flash Capabilities, Pricing, and Token Dynamics

The release of gemini-3.8-flash represents the general availability target for developer workloads requiring rapid reasoning, autonomous tool execution, and long-horizon software engineering. The model features a 1M token context window and accommodates up to 64k max output tokens. It is engineered to perform iterative tool calls, execute multi-file code modifications, and reduce failed agent loops.

Developers monitoring operational costs should note the pricing schedule. An introductory rate is available through December 31, 2026, setting costs at $0.75 per 1M input tokens and $3.75 per 1M output tokens. Standard pricing takes effect January 1, 2027, increasing rates to $1.50 per 1M input tokens and $7.50 per 1M output tokens. For cost estimation across model tiers, teams can refer to /tools/llm-api-cost-calculator/.

Because gemini-3.8-flash verifies intermediate steps and conducts multi-turn tool calling, it can consume higher token volumes on intricate workflows. When tasks do not demand deep reasoning, developers can dial down the thinking effort or direct lightweight traffic to gemini-3.5-flash-lite or gemini-3.6-flash.

Deep Research Agent Sunset: Migrating Before October 23, 2026

Unlike the standard Flash chat models, the deprecated agent endpoint does not route indefinitely. The changelog notes that deep-research-pro-preview-12-2025 will be shut down on October 23, 2026. Any calls submitted after that date will fail. Production pipelines utilizing managed research workflows must switch to one of the newer agent alternatives.

Google provides distinct targets depending on latency and depth requirements. For interactive applications requiring faster turnaround and streamed progress updates to end users, the target is deep-research-preview-04-2026. For automated background jobs requiring deep data gathering and exhaustive synthesis, developers should specify deep-research-max-preview-04-2026.

from google import genai

client = genai.Client()

# Migrate from deep-research-pro-preview-12-2025 before October 23, 2026
interaction = client.interactions.create(
    agent="deep-research-preview-04-2026",
    input="Synthesize industry benchmarks and recent performance analyses."
)

print(interaction.output_text)

Reasoning Levels and the Incompatibility of Minimal Mode

A key architectural change across Gemini 3 models is the formal deprecation of thinking_budget. Setting both thinking_level and thinking_budget returns a 400 error. Developers must configure reasoning depth via the thinking_level string enum, choosing between low, medium, and high.

A critical breaking change awaits teams migrating code from gemini-3.5-flash or smaller Lite variants: the minimal thinking level is not supported on gemini-3.8-flash and returns an error. While gemini-3.5-flash supported minimal for near-zero thought latency, gemini-3.8-flash mandates at least low thinking effort. Applications hardcoded with minimal will break when pointing to gemini-3.8-flash unless the configuration is modified.

{
  "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"
  }
}

In addition to thinking adjustments, Gemini 3 models remove support for traditional sampling parameters. Developers are advised to strip temperature, top_p, and top_k from requests, as the models are calibrated for default settings. Furthermore, candidate_count is unsupported in Gemini 3, and prefilled model turns are not permitted.

Function Calling and Multi-Turn Conversation Rules

Workflows employing structured outputs and external tools must conform to Gemini 3 function calling rules, detailed in guides like /posts/function-and-tool-calling-across-claude-gpt-and-gemini/ and /posts/gemini-structured-output-response-schema-lessons/. Gemini 3 requires exact response matching across calls and results.

• Include call_id and name in every function result payload matching the originating tool trigger.
• Return exactly one function response for every individual function call emitted by the model.
• Embed multimodal content inside the function response payload rather than as adjacent parts.
• Format inline instructions inside function responses using the double newline sequence \n\n.
• Track conversation continuity through server-side previous_interaction_id rather than supplying manual prefilled model turns.

For agent systems built on managed frameworks, Gemini 3.8 Flash now powers the default implementation for the Antigravity agent using agent string antigravity-preview-09-2026. This update provides enhanced reasoning for complex multi-tool workflows, web audits, and environment orchestration.

Production Migration Checklist

To ensure systems remain stable through the deprecation cycle, engineering teams should execute the following verification steps across their Gemini API codebases.

• Update target model strings from gemini-3.7-flash to gemini-3.8-flash in all interaction payloads.
• Update references to gemini-3.5-flash, changing them to gemini-3.6-flash or upgrading to gemini-3.8-flash.
• Replace deep-research-pro-preview-12-2025 with deep-research-preview-04-2026 or deep-research-max-preview-04-2026 prior to the October 23, 2026 deadline.
• Audit all generation_config blocks: eliminate thinking_budget, remove temperature, top_p, top_k, and ensure minimal is not assigned to gemini-3.8-flash.
• Confirm function calling payloads match tool return counts, include call_id and name, and position multimodal data within response payloads.
• Evaluate token usage patterns under thinking_level values of low, medium, or high, and review updated billing schedules at /tools/llm-api-pricing/.