# AI practitioners prioritize model triage to cut costs > Source: > Published: 2026-06-13 11:21:17.523325+00:00 # AI practitioners prioritize model triage to cut costs Running the same coding task through Claude Fable 5 cost $9, while GPT-5.5 cost $1.50 - a 6x gap on a single job, per The New Stack. Three concurrent signals reinforce the cost-pressure case: Anthropic's free Fable window closes June 22 before reverting to $10/$50 per million tokens; Citadel Securities' Tokenomics report (Frank Flight, June 10) warns that frontier inference costs are concentrating among organizations that can justify the spend; and OpenAI is weighing significant token price cuts per the Wall Street Journal. The convergence makes model triage - routing tasks to cheaper models when frontier capability is not required - an increasingly essential operational discipline. ### The cost comparison The New Stack reports running the same coding task through Claude Fable 5 cost $9, while GPT-5.5 cost $1.50 - a 6x differential that crystallizes a practical question: when is the most capable model actually the right tool? At the API level, Fable 5 lists at $10 per million input tokens and $50 per million output tokens; GPT-5.5 runs at $5 and $30 respectively, per Gizmodo. Both target long-running autonomous tasks like coding, where token counts run into the hundreds of thousands - making small per-token differences compound into large total bills. ### Citadel Securities' Tokenomics report In a June 10 report titled "Tokenomics," Citadel Securities macro strategist Frank Flight argues that the central constraint on AI adoption has shifted from model capability to cost and scarcity. Flight writes that "agentic and complex workflows delivered by frontier models" are "vulnerable to unrealistic expectations of frictionless deployment cost," pointing to Amazon removing its token leaderboard and Microsoft cancelling Claude Code subscriptions as early signals. The report describes "growing signs of a bifurcation in frontier vs everyday AI usage," with frontier inference concentrating among organizations whose operating domains justify the compute cost. A recent decline in the Silicon Data LLM Expenditure Index is cited as consistent with a broader shift toward "cheaper or more efficient models where the frontier technology is not required." ### Three industry signals this week Three developments converge on the same theme. First, Anthropic's free Fable 5 window - available to Pro, Max, Team, and Enterprise subscribers - expires June 22; from June 23, all users pay usage credits billed by token. Second, Citadel's report frames the AI cost constraint as operative and binding now, not a future concern. Third, the Wall Street Journal reported on June 11 that OpenAI is weighing significant token price cuts ahead of anticipated competition with Anthropic's expected IPO - a signal that frontier inference pricing is actively contested. ### Model triage in practice The New Stack frames model triage as the discipline of routing each task to the cheapest model that still meets quality requirements - distinct from defaulting to the most capable available. The Citadel report notes that "narrower, more disciplined, and more token-efficient applications" - coding assistants, customer-support copilots, analytical drafting - represent the most durable productivity path, in contrast to "autonomous agents running everything everywhere all at once." For teams managing agentic pipelines, mixing model tiers by task type and cost profile is increasingly a financial necessity, not just an optimization. ## Scoring Rationale Solid practitioner-facing cost analysis anchored by a concrete 6x cost differential and corroborated by the Citadel Securities Tokenomics report - a credible primary macro source. Primarily an explainer/synthesis piece rather than breaking news; the individual signal stories (Fable pricing, Citadel report, OpenAI cuts) have been covered separately. Score reflects solid-tier practical relevance for AI deployment teams. Practice interview problems based on real data 1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)