The cleanest way to read this release is as a price per marginal index point. Gemini 3.7 Flash delivered 56 at high effort for a measured $0.40 a task, and 3.8 delivers 59 for $0.58 [11][10][2]. The increment works out to $0.18 for three points, or $0.06 a point, against roughly $0.0071 a point that 3.7 averaged across its whole score [3][4]. Marginal intelligence at this end of the curve costs about eight times what the average point costs [5], and a team that books a version bump as a refresh will find a purchase on the invoice.
The rate card explains none of it. Google held the introductory rates at $0.75 per million input tokens and $3.75 per million output through December 31, 2026, and did not change them between the two versions [8][9]. What changed was consumption: output across the evaluation went from 64 million tokens to 120 million, up 87.5 percent [12][2]. Google's model card names the mechanism without dressing it up: "At times, the model might use more tokens to maximize performance, especially at higher effort levels" [19]. Artificial Analysis's per-task figures are weighted averages rather than provider list prices, and two models on identical per-token rates can produce different bills on the same evaluation [18].
The same arithmetic applied upward is what makes this a purchasing question rather than a ranking. Sol's two extra points over Gemini cost about $0.19 each, and Fable's seven cost about $0.44 each [6][7]. Fable's number is also not a clean single-model reading, because its default fallback means the run tested a system [17]. Below Gemini, GLM-5.3 Flash gives up two points at about a sixth of the per-task cost [6][7]. That is four prices for a unit of measured intelligence, and the index offers no basis for deciding which unit a given workload can convert into finished work.
Speed is the one axis where Gemini leads outright in this snapshot, at 304.6 output tokens a second against Sol's 72.4, Fable's 66.4 and GLM's 42.8 [13][14][15][16]. On the full run that is roughly 109 hours of generation for Gemini against about 269 for Sol [9][10], even though Gemini emitted 71 percent more output tokens [11]. For agent work that loops through many turns before returning anything, throughput is the cost a user feels, and it does not appear in the per-task price.
On a $0.58 task, 45 percent looks like rounding error for this quarter [1], and that reading holds up for now. The date that changes it is January 1, 2027, when the introductory rates double [8]. Hold token use where it is now and the same task costs $1.16, which is above what Sol measures today [8]. A team standardizing on 3.8 for agent work is accepting a per-task cost that reprices once, on a known date, by a known multiple.
Cheaper than the two systems above it, faster than everything in the snapshot: that is the board-deck version, and it is missing two things. The comparison ran at high effort while Google's default is medium, so the numbers most teams would actually generate are absent from the record [21]. And the aggregate disagrees with the domains. On Finance Agent v2, 3.8 led three of nine categories while 3.7 still led general quantitative analysis [22], and on the held-out Harvey legal test 3.8 resolved 10.0 percent of complete tasks against Muse Spark 1.2's 25.42 percent [23]. We do not know from this snapshot what 3.8 costs at the setting most production traffic will run on, and that is the number a buyer needs.
