Gemini 3.8 Flash ties Opus 5 on DeepSWE at a price Google doubles on January 1

Google's launch post explains the gains in four words: "3.8 Flash works harder." ^[7] According to the post, it works harder by "executing extra reasoning steps, and calling tools iteratively. At times, the model might use more tokens to maximize performance." ^[7] A per-token rate cannot tell a buyer how many tokens a task will take.

The DeepSWE v1.1 board, as reported in a dev.to comparison of the launches, measures that directly. Datacurve runs 113 tasks through the same mini-swe-agent harness for every model ^[6], so step and token counts compare across models on that board. For the same 74%, Opus 5 costs about five times as much per task as Flash ^[4][2]. Google deserves credit for this. Flash spends many cheap tokens and still finishes the task for less. Artificial Analysis saw the same spending pattern on a different workload. Its Intelligence Index run on Flash used about 1.8 times the median output tokens ^[6], and grading it cost $1,077.95 ^[8].

Google's footnote sets the price change: "Introductory price expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply." ^[3] The introductory rate lasts about four months from the September 2 launch ^[1][7]. At Flash's DeepSWE average, output tokens alone cost about $0.54 per task at today's $3.75 per million and about $1.07 at January's $7.50 ^[2][1]. Input tokens are excluded from those figures, so the full cost is higher. If token use per task holds, and the $2.36 was priced at the introductory rate, a full DeepSWE task goes to about $4.72 ^[3]. Opus 5's $11.84 is still about 2.5 times that ^[5].

Before trusting the $2.36 or the $4.72, I'd log steps and output tokens per finished task on my own harness, against my own tasks. A leaderboard's per-task figure is a measurement of Datacurve's workload.

Meta's Muse Spark 1.3 complicates the per-token view further. Its standard output rate of $4.25 per million sits above Flash's $3.75 today and below Flash's $7.50 from January ^[10][8]. The per-token ranking of the two models flips on a calendar date. The contributor endpoint is about 21 times cheaper than standard on output and 12.5 times cheaper on input ^[4]. The comparison does not include a DeepSWE result for Muse Spark 1.3, so its steps per task are unmeasured.

On Hacker News, user hiddencost wrote of Google's pricing: "you're effectively planning to charge users twice as much for a model that is no longer frontier." ^[12] The thread's top comment was Simon Willison's quick test: "make me a cool thing in html" returned a particle simulation in 13 seconds for 1.8 cents ^[13]. Another commenter noticed that its "60 FPS" counter was hard-coded into the page ^[14].