Gemini 3.8 Flash Is GA: What Devs Must Do Before 2027

Google released Gemini 3.8 Flash to general availability on September 2, and it’s live in the Gemini API right now under the model ID gemini-3.8-flash. Two things demand immediate attention: a new thinking_level parameter that replaces deprecated fields you may already be using, and an introductory pricing window that closes December 31 — after which the price doubles. Mark January 1, 2027 on your calendar.

The Pricing Clock Is Already Running

Gemini 3.8 Flash launches at the same introductory price as 3.7 Flash: $0.75 per million input tokens, $3.75 per million output tokens. On January 1, 2027, those rates double to $1.50 and $7.50 respectively. That gives teams roughly four months to evaluate, migrate, and lock in production workloads before the price jump, per Google’s announcement.

There’s a catch, though. Google describes 3.8 Flash as a model that “works harder” — executing extra reasoning steps and calling tools iteratively on complex tasks. That diligence costs tokens. Thinking tokens bill at the output rate and don’t show up in naive cost calculations. A call with 30K input tokens, 800 visible output tokens, and 6K thinking tokens costs $0.048, not the $0.0255 you’d estimate from the listed price alone. Always inspect usage_metadata.total_thought_tokens before you scale anything.

The batch API cuts both rates in half ($0.375/$1.875 per million tokens) for non-real-time workloads, and cached input reads cost $0.075 per million — a 10x discount. If your pipeline processes repetitive content, caching and the batch endpoint should be your first moves.

Gemini 3.8 Flash API Changes: Remove Deprecated Fields First

Swapping model IDs from gemini-3.7-flash to gemini-3.8-flash is the easy part. The breaking changes catch developers off guard. Gemini 3.8 Flash drops temperature, top_p, top_k, thinking_budget, and candidate_count. If your current code sets any of these, the API will reject the request.

The replacement for thinking_budget is thinking_level, which accepts three string values: "low", "medium" (the default), or "high". Don’t try to set "minimal" — it’s not supported and returns a validation error. Here’s the basic Python pattern:

from google import genai
from google.genai import types

client = genai.Client()
resp = client.models.generate_content(
    model="gemini-3.8-flash",
    contents="Your task here",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_level="low"),
        max_output_tokens=1024,
    ),
)
print(resp.usage_metadata)  # Always check total_thought_tokens

Multi-turn apps have two additional requirements: pass previous_interaction_id server-side, and include call_id and name on every FunctionResponse object. Skipping either produces Malformed_Function_Call errors that are difficult to trace back to their source.

Related: Gemini 3.8 Flash Cyber: Restricted Access and What Devs Must Know

Gemini 3.7 Flash vs 3.8 Flash: When to Actually Upgrade

Google itself says developers who care about compute efficiency should continue using 3.7 Flash. That’s unusually candid guidance from a product launch post. Take it seriously.

The case for switching to 3.8 Flash is real for agentic and multi-step workloads. DeepSWE v1.1 improved from 65.3% to 73.7%. OSWorld-2.0 jumped from 50.6% to 59.0%. On professional benchmarks — finance and legal agent tasks — 3.8 Flash beats Claude Opus 5. For long-horizon coding agents and complex reasoning pipelines, the quality gains are meaningful.

However, the case against switching is equally clear. First-token latency worsens: 13.3 seconds versus the 2.99-second median across models, making 3.8 Flash a poor fit for interactive chat or anything user-facing. Pure code generation barely moves (+0.2 points on the Coding Index). Moreover, higher thinking effort at default settings produces larger invoices than the listed price implies, because thinking tokens are invisible until you inspect the usage metadata.

The practical middle ground: keep 3.7 Flash on efficiency-first tasks, route long-horizon agentic work to 3.8 Flash via OpenRouter or directly, and set thinking_level="low" wherever you don’t need deep reasoning. Lowering the thinking level is the single highest-leverage cost control in 3.8 Flash — it directly reduces the most expensive token class. Alternatively, test for free in Google AI Studio before committing production traffic.

Key Takeaways

• Gemini 3.8 Flash is generally available now at introductory pricing ($0.75/$3.75 per million tokens) through December 31, 2026 — prices double on January 1, 2027.
• Remove deprecated fields before migrating: temperature, top_p, top_k, thinking_budget, and candidate_count are gone. Use thinking_level instead.
• Always check usage_metadata.total_thought_tokens — thinking tokens bill at output rates and aren’t visible in naive cost estimates.
• Upgrade for agentic and long-horizon tasks; stay on 3.7 Flash for latency-sensitive or cost-optimized workloads where the difference doesn’t justify the extra tokens.
• Batch API (half price) and caching (10x discount on reads) are the fastest ways to control costs if you do upgrade.