Build1 publisherNot yet confirmed elsewhere3 min readPublished
Gemini 4 Argon's per-task bill doubles when Google's launch discount ends
Google is charging $2 and $10 per million input and output tokens for Gemini 4 Argon at launch, half the $4 and $20 it will charge later. For a model that writes long answers and is open only to vetted cyber defenders, cost per task is the figure to check before switching.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- The maximum output per response rose from 64K tokens to 1M tokens.
- Google's own table has Argon winning or tying 13 of 18 benchmarks against Opus 5.5, Fable 5.1 and GPT-6 Astra, while losing on FrontierSWE and Terminal-bench.
- Google opened access on September 30 to cyber defenders in its Fairwind Program, with paid API customers and Google AI Ultra subscribers next in line.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Teams that measure Argon spend during the promotion will see half the steady-state cost, assuming the model keeps writing the same number of tokens.
- decision At equal output prices, Argon's per-task output bill runs about 2.3 times that of a 27k-token rival, so a per-token price list makes it look cheaper than it is.
- exposure A pipeline that leaves max output unset can now get a single 1M-token response that costs $20 at the post-launch rate.
- constraint Outside Fairwind, nobody can measure Argon on their own tasks yet, and the guarded model that ships later may handle security-heavy prompts differently from the build defenders use now.
Count only the output on that 62k-token test. At the launch rate of $10 per million, it comes to $0.62 per task [25]. At the later $20 rate it is $1.24 [26]. A rival writing 27k tokens per task would need an output price of about $23 per million to cost as much as Argon does at launch, and about $46 to match Argon after the promotion [27][28]. Opus 5.5 charges $20 per million output tokens [13]. GPT-6.1 Sol's newly discounted price matches Argon's launch price [14]. These figures leave out input tokens and rest on a single independent test [8].
The dev.to write-up calls price per token the wrong number to compare and points to Artificial Analysis, which measures cost per task [24]. Input is the easier side to control. Cached input tokens get 95% off, so a repo-sized prompt resent on every call costs $0.10 per million tokens at launch and $0.20 later [15][30].
Google picked the benchmarks in its 13-of-18 table and ran most of them itself, according to the tabulation by The New Stack [2]. Those wins carry over only if a team's work looks like the knowledge-work, long-context and DeepSWE tests where Argon leads [1]. A number of those leads are slim [3]. The margin on both the Vals Index and Vibe Code Bench is less than two points, and every one of the four models clears 89% on Vibe Code Bench [3]. The biggest gaps are in knowledge work and long context, but even the legal score of 19.6% means about one task in five completed [4]. On CWE-bench v1, Argon ties GPT-6 Astra at 68% [16]. The OpenAI and Anthropic models in that test ran inside Codex and Claude Code, so their scores include their tooling [17]. When Artificial Analysis ran Terminal Bench 4 itself, Argon scored 57%, behind Claude Sonnet 5.5 at 64%, Opus 5.5 at 60% and GPT-6 Astra at 59% [18]. Its Intelligence Index puts Argon level with GPT-6 Astra and one point ahead of GPT-6.1 Sol [19].
The engineering underneath is good. Google built Argon for long, multi-step work and trained it to find, validate and patch software vulnerabilities on its own [20]. A 1M-token output ceiling fits that kind of job [9]. Argon is the first Gemini above the Flash tier in more than seven months [21]. Gemini 3.5 Pro, teased at I/O in May, retires without ever having shipped [21].
The dev.to author found no migration notes in Google's announcement or the coverage [22]. Google says developers, enterprises and consumers get access after paid API customers and AI Ultra subscribers, once the guardrails have been tuned with early testers [10]. As the rollout expands, Google says, it is participating in a voluntary US government process for pre-release access [23].
What to watch
- The date Google ends the 50% launch promotion, and whether paid API customers get access before it ends.
- Artificial Analysis cost-per-task figures for Argon at the $4 and $20 rates, or a second token-count test that confirms or revises the 62k figure.
- Whether the guarded general release changes Argon's scores or behaviour on security prompts compared with the Fairwind build.