Gemini 3.8 Flash beats partner-only Argon for anything shipping this quarter

A comparison posted on dev.to recommends shipping on Flash and keeping Argon one configuration switch away ^[24]. It builds that switch as a shell default, `MODEL="${GEMINI_MODEL:-gemini-3.8-flash}"`, interpolated into the `generateContent` URL ^[22]. It also warns against hardcoding a guessed Argon ID, advice that should not need printing ^[21]. The same request body pins `thinkingConfig.thinkingLevel` to `"medium"` and `maxOutputTokens` to 8192 ^[22].

I agree with building on Flash. I would also put the thinking level and the output cap in the same environment as the model name, so one deploy changes all three. Flash's levels are described in Google's thinking documentation ^[25]. Google has not published Argon's model ID, input window, thinking levels or default, how it bills thinking tokens, or a batch price ^{[21][6][8][15][20]}. Argon's evaluations ran at the highest thinking settings ^[8].

The post's daily cost comparison holds token counts equal across the two models ^[17]. On current Gemini models, Flash included, thinking tokens bill as output, and the example's 8,000 output tokens count them ^[15]. The post's Argon figures assume Argon bills the same way ^[15]. Google says higher Flash thinking levels can consume more tokens ^[18]. Suppose Argon runs at the setting its benchmarks used and emits more output tokens per call than the example assumes. Then the per-call gap grows past the 8/3 per-token ratio ^[16][8]. The post's test script already carries the right guard: it asserts `usageMetadata.thoughtsTokenCount` stays below 20,000 ^[23].

Argon's benchmark case rests on two rows. Both are Google-reported, Argon's methodology attributes both to Vals AI, and the two tables appeared a month apart against different comparison models ^[10]. On finance, Argon scores 65.4% to Flash's 61.4% ^[11]. For the legal-agent gap to carry over to someone else's agent, two things have to be true. The task has to resemble the benchmark's, and Argon has to run at the setting that produced the score ^[12][8]. Arena's text leaderboard, a third-party ranking, places Argon (High) first and Flash (High) eleventh on preliminary votes ^[13].

Flash has one more discount. Batch requests are 50% off, at $0.375 input and $1.875 output per million tokens through 2026, according to the Gemini API pricing page as cited in the post ^[20]. On batch, the example workload comes to $52.50 a day ^[3], for jobs that can run as batch work.

For anyone outside Fairwind, access settles it. Argon is open only to Fairwind Program partners ^[5]. Even Flash's cyber sibling, Gemini 3.8 Flash Cyber, is GA only behind an allowlist on the Gemini Enterprise Agent Platform ^[14]. Everyone else can build today on the general Gemini 3.8 Flash ^[1].