Gemini 3.8 Flash

What is Gemini 3.8 Flash?#

Gemini 3.8 Flash is Google's newest Gemini model, released September 2, 2026, three weeks after Gemini 3.7 Flash and the third Flash release in six weeks. It is built on 3.7 Flash and aimed at long-horizon coding and autonomous agents. The shape is unchanged: a 1M-token context window, 64K output tokens, and text, image, audio, and video input at the same speed and price as 3.7 Flash. The reasoning is what moved. It runs on Unifically as google/gemini-3.8-flash, with four reasoning effort levels (low, medium, high, max) and a knowledge cutoff of March 2026. A limited-access sibling, Gemini 3.8 Flash Cyber, exists for security work and is not on Unifically.

What's new in Gemini 3.8 Flash#

Long-horizon coding jumps eight points

On DeepSWE v1.1 it scores 73.7%, up from 65.3% on Gemini 3.7 Flash, ahead of GPT-5.6 Terra at 69.6% and GPT-5.6 Sol at 72.7%, and within a third of a point of Claude Opus 5 at 74.0%. Multi-hour software engineering runs are where the new model separates itself.

Top of its table on Terminal-Bench 2.1

Agentic terminal coding lands at 89.4%, ahead of Claude Opus 5 at 89.1%, GPT-5.6 Sol at 88.8%, and Gemini 3.7 Flash at 85.8%. It is the best Terminal-Bench 2.1 score in the launch comparison, from a Flash model.

More intelligence at the same speed

Artificial Analysis scores it 59 on the Intelligence Index at high effort, three points above Gemini 3.7 Flash, rank #16 of 195. Output speed is 305 tokens per second, rank #3 of 195, and it sits on their intelligence-versus-cost-per-task Pareto frontier.

A fourth reasoning effort level

`reasoning_effort` now takes `low`, `medium`, `high`, and `max`. At `low`, a trivial prompt comes back with no reasoning tokens at all; `max` spends the most. Set it per request so cheap routing calls and hard agent steps share one model.

Best for#

Use cases#

Build coding agents that own a task end to end: repo-wide refactors, failing-test loops, CI triage, and terminal work, where DeepSWE v1.1 and Terminal-Bench 2.1 scores translate into fewer stalled runs. Use it for finance and legal pipelines that extract obligations, reconcile statements, and draft memos; it tops both the Vals Finance Agent v2 and Harvey's Legal Agent Benchmark columns in Google's table. Point it at hour-long recordings, chart-heavy reports, and PDFs: 87.8% on LVBench, 86.2% on CharXiv Reasoning, and 35.0% on GDP.PDF cover video, charts, and documents in one call. For research agents, 54.9% on HLE-Verified and 56.5% on the human-difficult BioMysteryBench split give expert-level answers at Flash latency and cost.

Limitations#

Token use can climb. At higher effort levels the model may spend more tokens to maximize performance, and thinking tokens bill as output tokens. Occasional slowness or timeouts are a known issue at launch, so set client timeouts with headroom on max.

General computer-use agents are not its strength yet. Terminal-Bench 4.0 lands at 19.1% against 51.8% for Claude Opus 5 and 37.3% for GPT-5.6 Sol, OSWorld-2.0 at 59.0% against 75.4% for Claude Opus 5, and GDPVal-AA v2 at 1545 Elo against 1824. For broad desktop automation, a frontier model still wins.

Multilingual safety regressed slightly against Gemini 3.7 Flash in Google's own testing. The knowledge cutoff is March 2026, but some domains only reach January 2025. Most benchmark numbers come from Google's evaluation table; Artificial Analysis is the independent check. The minimal effort level is rejected.

Gemini 3.8 Flash vs Gemini 3.7 Flash#

Every agentic number moves up. DeepSWE v1.1 goes from 65.3% to 73.7%, Terminal-Bench 2.1 from 85.8% to 89.4%, Terminal-Bench 4.0 from 11.2% to 19.1%, OSWorld-2.0 from 50.6% to 59.0%, and the human-difficult BioMysteryBench split from 43.5% to 56.5%. Knowledge-work scores rise more modestly: HLE-Verified 53.6% to 54.9%, GDPVal-AA v2 1482 to 1545 Elo, and the Artificial Analysis Intelligence Index 56 to 59. Price is unchanged, speed stays in the same band, and 3.8 Flash adds the max effort level. Unless a pipeline is regression-locked to 3.7 Flash behavior or depends on its multilingual safety profile, new work belongs on 3.8 Flash.

When to use Gemini 3.8 Flash#

Use Gemini 3.8 Flash as the default for coding agents, finance and legal agents, and long video or document work when you want near-frontier agentic scores at Flash speed and cost. Step up to a frontier model for desktop automation and broad knowledge work, where Claude Opus 5 and GPT-5.6 Sol keep a clear lead. Step down to a smaller model only for pure classification and routing.

API examples#

Call Gemini 3.8 Flash from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/google/gemini-3.8-flash.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "google/gemini-3.8-flash",
    "reasoning_effort": "high",
    "messages": [
      { "role": "user", "content": "Find the failing test in this repo, explain the root cause, and propose a minimal patch." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate, and change reasoning_effort to low for cheap routing calls or max for the hardest steps.