Gemini 3.6 Flash bills output tokens at five times the input rate

A single ticket does not produce a single request. The example loop in the dev.to writeup reads a ticket from Linear, GitHub or email, searches a repository, opens several source files, runs tests, reads the failures, modifies code, runs tests again, writes a pull request summary, and asks for approval before a merge ^[4]. Nine steps ^[3]. If every one of them sends back the whole conversation, the full repository files and the raw command output, token usage compounds ^[5].

Input runs $1.50 per million tokens and output $7.50 ^[3]. Output therefore bills at five times input ^[1]. Trimming a million tokens of resent context saves $1.50; recovering that same $1.50 on the generation side means cutting 200,000 tokens of output ^[2]. The long implementation plan the writeup names as part of the expensive path is output ^[1].

Google's Gemini API documentation describes tokens as "the smallest unit of text that the Gemini models process," according to the post ^[6]. Billing follows tokens, not messages. One "please fix this bug" request may become many model calls, each carrying thousands of tokens of accumulated context ^[7].

The claim is bounded twice: up to 65%, and on long-horizon engineering tasks ^[2]. The writeup warns against treating it as a blanket promise that every workflow will cost 65% less ^[19], and lists the variables that decide a real result: token usage, the input/output balance, the number of retries, context-caching options, and whether the new model finishes tasks in fewer or more steps ^[8]. It does not identify the baseline or the task set behind the figure ^[20]. For the number to transfer, your retry rate and your input/output split would have to look like whatever was measured.

Routing in the writeup assigns work by risk. Flash-Lite takes high-volume, lower-risk operations ^[10]: converting logs into structured JSON, generating test-case checklists, summarizing a completed tool run, extracting issue IDs, building a concise repository map, reformatting a pull request description ^[11]. Gemini 3.6 Flash covers multi-step reasoning without pricing every step like a premium call ^[12]. The expensive route keeps designing a migration plan, explaining a failing integration test with several possible causes, reviewing a security-sensitive change, comparing two implementations, and recovering from repeated tool failures ^[13]. A coming Gemini 3.5 Pro tier will likely matter for that class of decision, the post says ^[14]. It also argues that a model scoring slightly lower on an isolated benchmark can still cut total operating cost if it handles routine work reliably and escalates only when needed ^[18].

One failure mode is named: a universal fallback that sends every uncertain task to the most expensive model. It feels safe, the post says, and usually hides bad prompt design, missing retrieval, unbounded tool output or weak approval rules ^[15]. Routing every uncertainty upward is the most expensive way to postpone fixing retrieval.

Its cost formula multiplies measured input and output token totals by the published rates, then adds retries and any other services in the pipeline ^[16]. For accounting, the post argues for one figure: the cost of a completed and accepted engineering task, after retries, failures, human review and tool usage ^[17].