Gemini 3.8 Flash: a new worker model for agents, routing, and automation
Google released Gemini 3.8 Flash on September 2 as a production model for long-running coding agents, document work, and multi-tool workflows. The important enterprise change is not a few more benchmark points. It is that the main worker in an agent router is temporarily priced like a lower-cost operating tier while still being capable of consuming substantially more tokens at higher thinking levels.
That makes 3.8 Flash a candidate for work previously escalated to a more expensive model. It should not become an uncontrolled default for every email, invoice, and support ticket.
What is actually new
The model ID is gemini-3.8-flash. It is generally available in the Gemini API with 1,048,576 input tokens, up to 65,536 output tokens, and text, image, audio, video, and PDF input. It supports function calling, structured output, code execution, file search, caching, web and Maps grounding, plus Computer Use in preview. Output is text; Live API and image generation are not supported.
Thinking is configurable as low, medium, or high, with medium as the default. minimal returns an error. Google also says that on long, complex work, 3.8 Flash deliberately takes smaller reasoning steps, calls tools iteratively, and verifies along the way. That can improve completion quality, but it can also increase output and thinking-token usage.
Pricing: a temporary discount, not the permanent baseline
Through December 31, 2026, Gemini Developer API pricing per million tokens is:
| Mode | Input | Cached input | Output including thinking |
|---|---|---|---|
| Standard | $0.75 | $0.075 | $3.75 |
| Batch / Flex | $0.375 | $0.0375 | $1.875 |
From January 1, 2027, Google lists standard pricing at $1.50 input, $0.15 cached input, and $7.50 output. Explicit cache storage is also billed at $0.50 per million tokens per hour through year-end, then $1.
This matters most for workflows that can wait. Overnight support classification, bulk invoice extraction, document reprocessing, and content pipelines belong in Batch or Flex rather than interactive mode simply because that is how the prototype started.
A practical router now
I would split the stack into five layers:
- A deterministic layer checks file type, mandatory fields, PII, duplicates, and known rules without an LLM.
- Flash-Lite handles low-cost classification, simple extraction, and fixed-schema JSON normalisation.
- Gemini 3.8 Flash on low or medium takes multi-document cases, tool use, ambiguous instructions, and longer plans.
- High thinking or a stronger specialist receives only difficult exceptions whose value justifies the cost.
- A human approves payments, contractual decisions, permission changes, publication, and other irreversible actions.
The pattern works in n8n, Make, or custom orchestration. For CRM enrichment, Lite can extract company and contact data, 3.8 Flash can combine the public site, internal history, and business rules, and a person can approve the lead-status change. For invoices, Lite extracts fields, 3.8 Flash handles an unusual attachment, and accounting signs off the posting.
Caching is prompt architecture
Implicit caching is enabled automatically, with a 4,096-token minimum for 3.8 Flash. Improve hit rates by keeping the common prefix stable: system rules, schemas, dictionaries, and shared reference material first; the individual ticket or document afterwards. Track usage.total_cached_tokens or you do not know whether the advertised saving appears in your workload.
There is an API boundary to watch. The newer Interactions API supports implicit caching only. Explicit cache objects require generateContent. Swapping a model ID without checking that distinction can silently change the economics of a long-running agent.
Migrating from 3.6 Flash
First move the model ID and thinking level into configuration rather than embedding them in prompts. Compare 3.6 and 3.8 Flash on 50–100 anonymised production cases. Measure correctness, valid JSON, tool calls, cache hits, latency, output tokens, human corrections, and cost per approved outcome.
Roll out one workflow at a time with a fallback to the previous model. Start where outputs are deterministically testable: code tests, accounting checks for invoices, required CRM fields, or a closed classification set for support. For long agents, enforce a step cap, token budget, persistent state outside the context window, and a stop condition. A smarter model does not replace those controls.
Google positions 3.8 Flash as its strongest workhorse for coding and agents and reports improved vendor benchmarks. That is a reason to evaluate it, not a substitute for an internal eval. The model card still lists hallucinations, occasional slowness or timeouts, and higher token use at greater effort as known limitations.
Sources: Google AI for Developers: Gemini 3.8 Flash and migration overview, current pricing, context-caching rules, product page and vendor benchmarks, model card and limitations, and Marek Bartoš: Czech video signal for the release.