Billing is per token, not per task. That is what makes the unchanged rate card misleading. Google's own copy says 3.8 Flash executes extra reasoning steps and calls tools iteratively, and may use more tokens to maximise performance, especially at higher effort levels [4]. Same rate, more units per unit of work.
The config line that proves Google knows this sits in the same post: use lower effort levels to minimise token overhead, or stay on 3.7 Flash, which remains fully supported for efficiency-first workloads [5]. A vendor pointing you at last month's model for cost reasons is doing you a favour. It is also telling you what the new one costs.
Then the date. Both halves of the introductory price double: input from $0.75 to $1.50, output from $3.75 to $7.50 per million tokens, exactly 2.0x each [2][1]. A unit of work that reads a million input tokens and writes a million output tokens costs $4.50 now and $9.00 after the expiry [2]. If the extra diligence lifts token use by half, the post-expiry bill is 3x what you are paying today [3]. That 50% is my illustration, not Google's number; nothing in the announcement quantifies the overhead. Google also says the release cadence is accelerated by long-running agentic loops that recursively evaluate and refine the models [19], which is a reason to expect another version before you finish migrating to this one.
On benchmarks, treat the table as a claim about someone else's repo. Google says 3.8 Flash matches Opus 5 on DeepSWE and beats GPT-5.6 Sol and Sonnet 5 [7]. For that to transfer, your tasks need to look like DeepSWE's: long-horizon engineering problems solved end to end with a verifiable finish. Where it still trails is computer use, 59% against 75.4% for Opus 5 on OSWorld-2.0 [12], and knowledge work, 1545 on GDPVal against 1824 for Opus 5, 279 points back [13][5]. The New Stack notes that Google's comparison set omits Anthropic's Fable 5.1, which shipped a day earlier [9]; on Terminal-Bench 4.0, Flash 3.8 scores 19.1% against Fable's 55.8% and Opus 5's 51.8% [8], a 2.9x gap to Fable [4]. On the coding-focused Terminal-Bench 2.1, Flash 3.8 beats its competitors [10], and The New Stack reads Terminal-Bench 4.0 as the outlier rather than the verdict [11]. The same publication points out that GLM-5.3, DeepSeek v4 Pro and Kimi K3 sit in the same league on DeepSWE 1.1 at better price/performance [20], which is the comparison the price change actually invites.
Fairwind is the more interesting piece of engineering discipline. Google says it invested in vulnerability fixing from the start and prioritised it over offensive capabilities like exploitation [14], and reports 47.2% pass@1 on Collinear's CWE-Bench against 47.8% for a leading frontier model at significantly lower cost [15]. Near-parity on patching at Flash cost, still under 50% first-try, means a human reads every diff. Access runs through about 650 trusted partners including CrowdStrike, Palo Alto Networks, Datadog, Snowflake, Wiz, Accenture and the Center for Internet Security [6]. The 3.5 Cyber equivalent was unnamed and, by The New Stack's reading, ad hoc [16]. Naming the programme and putting 650 partners behind it turns a capability into a procurement category.
