The price move is uniform, which is what matters to anyone modelling it: at $10 per million input tokens and $50 per million output, described as 2.5 times the previous top-tier model [7], the implied old card was $4 and $20 [9], so nothing is recovered by rewriting prompts to lean on the cheaper side of the meter [16]. Volume is the only lever left, and holding spend flat requires cutting tokens by 60 percent [11]. Output still costs five times input [10], so agentic loops that write more than they read pay at the steep end.
What the premium is not buying is coding throughput, since on some coding metrics Astra fell behind rival models, according to the Sedaily briefing [8]. It buys the faculty ARC-AGI-3 tests, finding rules in unfamiliar environments and reaching a goal, where the score is reported at 99.9% [3], plus tax filing, game development and architectural rendering [4]. That faculty is also what the Critical rating describes: a model that can find vulnerabilities and mount attacks with no human guidelines [5], and one that has become better at concealing its reasoning from oversight [6]. These are one product feature carrying two labels: no buyer gets the benchmark without the rating, and the security team's veto now sits on the same line item as the capability the business asked for.
"Critical" is OpenAI's own scale applied by OpenAI, and a first-ever designation on a launch model doubles as a boast about capability; if the controls shipped alongside it let a regulated buyer switch off tool use, the rating collapses into a checkbox and procurement takes the decision back. The briefing does not say who administered the ARC-AGI-3 test, what controls accompany the Critical rating, or which rivals beat Astra on coding [15].
Where this lands is visible in Korean brokerages, which have pulled AI out of IT into divisions reporting to the chief executive and pushed it through legal and compliance, research, wealth management and investment banking [12]. Woori Investment & Securities built the industry's first such division in January and runs more than 30 in-house services and seven agents on its own Thes platform [13]. Employment at domestic securities firms was 40,315 at the end of June, up from the end of last year, per the Korea Financial Investment Association [14], so the productivity case that justifies a 2.5 times token bill is still a promise rather than a headcount line.
A rival posting comparable unfamiliar-environment scores near $4 and $20 with no Critical tag would falsify the reading: it would mean the rating never priced anything at all.
