Gemini 3.8 Flash: Frontier Agent Scores at a Flash Price, For Now

Gemini 3.8 Flash is the model Google now points new projects at. Its API models page describes it as the most capable Flash model, aimed at long-horizon coding, autonomous agents and complex enterprise workflows, and recommends it alongside 3.5 Flash-Lite for anyone starting fresh. Requests for the deprecated 3.7 Flash are already routed to it.

The interesting thing about 3.8 Flash is not that it is better than 3.7. It is that on a handful of agentic benchmarks Google's own table puts it level with models that cost five to seven times as much, while on others it is far behind. Knowing which is which, and knowing that it is designed to spend more tokens, is most of what you need to use it well.

Level with the frontier on some rows, far behind on others

Google published the model card on 2 September 2026. It says 3.8 Flash is built on Gemini 3.7 Flash, takes text, images, audio and video with a 1M-token context window, and produces up to 64K tokens of text. The knowledge cutoff is March 2026 for some domains and January 2025 for others. There is no published parameter count.

The benchmark table compares it with Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra. Where it holds its own:

• Terminal-bench 2.1: 89.4%, against 89.1% for Opus 5 and 88.8% for GPT-5.6 Sol
• DeepSWE v1.1: 73.7%, against 74.0% for Opus 5 and 72.7% for Sol
• HLE-Verified: 54.9%, against 54.4% and 54.5%
• Vals Finance Agent v2: 61.4%, top of the table

Where it does not:

• Terminal-bench 4.0: 19.1%, against 51.8% for Opus 5 and 37.3% for Sol
• OSWorld-2.0 (computer use): 59.0%, against 75.4% for Opus 5
• GDPVal-AA v2 (knowledge work): 1545 Elo, against 1824 for Opus 5

The pattern matters more than any single row. On established benchmarks that the field has had time to optimise for, 3.8 Flash is at the frontier. On the newest and hardest revision of Terminal-bench, it scores less than half of Opus 5. That gap is the better guide to what happens on genuinely novel tasks. Older benchmarks tend to flatter every model, because the whole field has tuned against them.

For price context, the same card lists Opus 5 at $5.00 input and $25.00 output per million tokens, and GPT-5.6 Sol at $4.00 and $20.00. 3.8 Flash is listed at $0.75 and $3.75.

It spends tokens on purpose

Google is unusually direct about this. Its 3.8 Flash migration guide says the model "can use more tokens on longer running and complex tasks, by design". It takes smaller reasoning steps, calls tools iteratively and checks its own work. The model card repeats the point under known limitations, alongside occasional slowness or timeouts.

That is a sensible design for an autonomous agent, where a verified answer is worth far more than a fast wrong one. It is a poor default for a chat reply or a one-line classification. The price per token is low; the tokens per task may not be. The token cost of reasoning models covers how to budget for output you never see.

Choosing an effort level

Thinking is set with thinking_level, which replaces the older thinking_budget. Three values are accepted, and minimal returns an error:

• low: for latency-critical work where time to answer matters most
• medium: the default, which Google recommends for complex code and agent use
• high: for deep reasoning, mathematics and difficult multi-step problems

Google's own advice for cutting token use on everyday tasks is to lower the effort. In practice, route by task: low for summaries, lookups and simple edits; medium for agent loops; high only where you have measured that it changes the outcome.

Migrating from older Gemini models

Moving from Gemini 2.5 or an early Gemini 3 model involves more than changing a string. Google's migration notes list:

1. Replace thinking_budget with thinking_level. Sending both is an error.
2. Remove temperature, top_p and top_k, which are deprecated, and candidate_count, which Gemini 3 and later do not support.
3. Remove prefilled model turns, and use the server-side previous_interaction_id for multi-turn state.
4. On generateContent, make sure every function response carries both call_id and name.
5. Update your SDK and preserve thought signatures between turns if you manage history yourself.

If you are coming from 3.7 Flash, you are probably already on 3.8 through the redirect. The Gemini 3.7 Flash guide covers what that silent swap means for your budgets.

Capabilities include function calling, structured outputs, code execution, search and Maps grounding, URL context, caching and computer use in preview, with Batch, Flex and Priority inference available. Google's Antigravity managed agent and SDK now default to 3.8 Flash. A separate 3.8 Flash Cyber variant is restricted to governments and approved partners and is not something you can call.

The price is introductory

Google lists 3.8 Flash at $0.75 per million input tokens and $3.75 output through 31 December 2026, with cached input at $0.075. From 1 January 2027 the rate becomes $1.50 and $7.50. The :batch variant on OpenRouter is the same model on Google's asynchronous Batch API at half the standard rate, $0.375 and $1.875, and Google says those discounted rates double on the same date. Aggregator listings may differ slightly from Google's own prices, so check the rate you are actually billed.

Even at the 2027 rate it is far cheaper than the frontier models it competes with on some benchmarks. But a cost model built in October will be wrong in January unless it accounts for the step.

How to evaluate it on your own work

The claim to test is specific: that 3.8 Flash completes long agentic tasks about as well as a frontier model, for much less. A short protocol:

1. Take ten to twenty real tasks from your history, mixing routine ones with a few that previously needed many turns.
2. Run each at low, medium and high, and through whichever frontier model you use today, on the same harness.
3. Record pass rate, turns, total tokens and wall-clock time, not just pass or fail.
4. Compute cost per completed task at the 2027 rate as well as today's.

How to benchmark LLMs on your own work has the harness detail. If 3.8 Flash at medium lands close to your frontier model on routine work, route that work to it and keep the frontier model for the novel tasks, which is where Terminal-bench 4.0 suggests the gap still lives. For a steadier, cheaper alternative with lower token appetite, the Gemini 3.6 Flash guide covers the previous workhorse.