# Gemini 3.8 Flash Narrows the Agent Gap, but Token Use Still Decides Cost

**Gemini 3.8 Flash became generally available on September 2, 2026, with a 73.7% DeepSWE v1.1 score.** Introductory API prices are $0.75 per million input tokens and $3.75 per million output tokens. Those rates look compelling. Google also says the model may consume more tokens on difficult tasks because it reasons in smaller steps, calls tools repeatedly, and checks its work.

The useful buying metric is cost per accepted task. Token price alone cannot tell an engineering team whether a model will lower its production bill.

![Blue abstract identity artwork from the official Gemini 3.8 Flash model page](https://s4.tenten.co/learning/content/images/2026/09/landing-page-1-4.png)

#### Google Built This Flash Release for Long-Running Agents

Google shipped three Flash updates in six weeks. Senior Director of Product Management Tulsee Doshi presented Gemini 3.8 Flash as a workhorse for long-horizon software engineering, autonomous agents, and complex enterprise workflows.

The stable model ID is `gemini-3.8-flash`. It accepts text, images, video, audio, and PDF files. The context limit is 1,048,576 input tokens, while the output limit is 65,536 tokens.

The model supports function calling, code execution, file search, search and Maps grounding, structured outputs, and URL context. Computer use remains in preview. Image generation, audio generation, and the Live API are unavailable.

Google released Gemini 3.8 Flash Cyber on the same day. Gemini Security Lead Raluca Ada Popa co-authored the announcement. That variant targets vulnerability discovery and patching, but access is limited to approved defenders in the Fairwind Program.

#### The Benchmark Gains Are Real, but the Test Conditions Vary

Google's headline table shows a clear improvement over Gemini 3.7 Flash. DeepSWE v1.1 rose from 65.3% to 73.7%. Terminal-Bench 2.1 increased from 85.8% to 89.4%, while HLE-Verified moved from 53.6% to 54.9%.

The model also reached 61.4% on Vals Finance Agent v2. Its 10.0% score on Harvey's Legal Agent Benchmark is the full task-resolution rate, which requires every criterion in a task to pass.

| Evaluation | Gemini 3.8 Flash | Gemini 3.7 Flash | Test context |
|---|---:|---:|---|
| DeepSWE v1.1 | 73.7% | 65.3% | Google computed the 3.8 result with a mini-swe harness and high thinking |
| Terminal-Bench 2.1 | 89.4% | 85.8% | Google computed Gemini results with the default Terminus 2 harness |
| HLE-Verified | 54.9% | 53.6% | Google tested the full 1,811-item verified set |
| Vals Finance Agent v2 | 61.4% | 59.0% | Vals AI runs a shared six-tool agent harness |
| Harvey's Legal Agent Benchmark | 10.0% | 8.8% | The score measures complete task resolution |

![Google DeepMind bar chart showing Gemini 3.8 Flash at 61.4% on Vals Finance Agent v2](https://s4.tenten.co/learning/content/images/2026/09/landing-page-3-2.png)

The methodology deserves as much attention as the scores. Google's comparison mixes internal runs, third-party leaderboards, and provider-reported competitor numbers. DeepSWE uses each model's highest listed thinking level, while Gemini 3.8 Flash runs with high thinking.

Some multimodal conditions also differ. Google's LVBench test sampled 1,024 frames for Gemini and GPT models, but only 300 frames for Claude models because of API limits. HLE-Verified results came from Google's own runs.

There is an unusually candid footnote. Google says it initially reported Claude Opus 5 at 74% on DeepSWE because of rounding. The episode does not erase the benchmark, but it weakens any sweeping claim based on a few tenths of a point.

Use vendor tables to select candidates. Use your own repository, tools, and acceptance tests to select a production model.

#### Cheap Tokens Do Not Guarantee a Cheap Agent

Standard API pricing stays at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. On January 1, 2027, those rates rise to $1.50 and $7.50. Output pricing includes thinking tokens.

Batch requests cost $0.375 per million input tokens and $1.875 per million output tokens during the introductory period. The regular Batch rates will be $0.75 and $3.75.

The pricing advantage can disappear if an agent takes more steps. A harder task may need extra reasoning tokens, repeated tool calls, and larger contexts carried across turns. That extra work can still be economical if it prevents retries and human review.

Teams should calculate one number.

`cost per accepted task = total input and output spend / tasks that pass acceptance tests`.

![Google DeepMind chart comparing Gemini 3.8 Flash DeepSWE v1.1 success rate with average cost per task](https://s4.tenten.co/learning/content/images/2026/09/landing-page-2-3.png)

The denominator catches what token-price comparisons miss. First-pass success, tool-call count, P50 and P95 latency, retry frequency, and review time all shape the final cost. Google's DeepSWE chart plots success against average task cost for this reason.

#### Start With Medium Thinking, Then Earn the Upgrade

Gemini 3.8 Flash defaults to `medium` thinking. It also supports `low` and `high`. The `minimal` setting is unsupported and returns an error.

Google recommends low thinking for latency-sensitive chat, drafting, and fast analysis. Medium targets complex code and agent workflows. High targets deep reasoning, math, and difficult multi-step jobs.

A production evaluation should begin with medium. Build a set of 30 to 100 tasks drawn from real traffic, then hold tools, prompts, and acceptance criteria constant. Compare pass rate, output tokens, latency, and retries across all three settings.

High thinking has no purchasing value when medium already clears the quality threshold. Low may be the better default for simple extraction or classification, even if the model's strongest benchmark result used high.

#### Migration Requires More Than a Model-Name Change

Google's migration guide calls for several configuration changes. Update the model ID to `gemini-3.8-flash`. Remove `temperature`, `top_p`, `top_k`, and `candidate_count`, then replace `thinking_budget` with `thinking_level`.

Existing applications should also audit function responses and conversation turns. Multimodal assets must sit inside the response payload. Prefilled model turns are invalid, and the final user turn must contain non-empty text.

These changes can break an established agent even when the new model performs better. Run regression tests before shifting traffic. Keep Gemini 3.7 Flash available as a rollback target because Google continues to support it.

#### Who Should Test Gemini 3.8 Flash First?

The best candidates run multi-tool jobs across repositories or long documents and currently pay a premium for reasoning quality. Software-maintenance agents, enterprise search, financial document work, and research workflows fit Google's target.

High-volume, simple requests may gain less. A stable 3.7 Flash deployment can remain easier to predict when low thinking already meets the service-level target. The Cyber variant is also a separate procurement path with controlled access and governance requirements.

#### Frequently Asked Questions

##### Is Gemini 3.8 Flash generally available?

Yes. Google marked `gemini-3.8-flash` as generally available on September 2, 2026. Developers can use it through the Gemini API and Google AI Studio.

##### What are the context and output limits?

The input limit is 1,048,576 tokens, and the output limit is 65,536 tokens. The model accepts text, image, video, audio, and PDF inputs but returns text only.

##### When will Gemini 3.8 Flash pricing change?

Introductory pricing ends on December 31, 2026. Standard input and output rates double to $1.50 and $7.50 per million tokens on January 1, 2027.

##### Can developers turn thinking off?

No. Gemini 3.8 Flash supports low, medium, and high thinking levels. The minimal setting is unsupported and returns an error.

#### Sources

- [Google — Introducing Gemini 3.8 Flash and 3.8 Flash Cyber](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/)
- [Google DeepMind — Gemini 3.8 Flash](https://deepmind.google/models/gemini/flash/)
- [Google DeepMind — Gemini 3.8 Flash model card](https://deepmind.google/models/model-cards/gemini-3-8-flash/)
- [Google DeepMind — Gemini 3.8 Flash evaluation methodology](https://deepmind.google/models/evals-methodology/gemini-3-8-flash/)
- [Google AI for Developers — What's new in Gemini 3.8 Flash](https://ai.google.dev/gemini-api/docs/generate-content/latest-model)
- [Google AI for Developers — Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing)

#### Author Insight

The most useful choice in Google's launch material is the DeepSWE cost-per-task chart. Agents retry, expand their contexts, and move data between tools. Any evaluation that omits success rate, token use, latency, and human rework is still several steps away from a purchasing decision.