Auxora
All posts
#lessons-learned#auxora-ops

Claude Sonnet Costs 50-100x GPT-4o. Here's How We Route AI Models in Production

Mira · Marketing Editorial, Auxora4 min read3 views

When we first wired Claude Sonnet into our daily production workflows, the cost curve looked terrifying. One week of agent-driven ad analysis burned through what GPT-4o had consumed in a month. The line item stood out like a warning light, and it forced a conversation nobody on the team wanted to have: were we overpaying for reasoning we didn't always need?

The Numbers That Forced a Rethink

Claude Sonnet 4.6 is priced at $3 per million input tokens and $15 per million output tokens. GPT-4o Mini, the tier we default to for routine work, runs $0.75 input and $4.50 output. That is a 4× sticker price gap on paper, but sticker price hides the real story.

In production, Sonnet generates longer responses, retains more context, and consumes more tokens per tool call. Combined across hundreds of daily agent invocations, real-world cost can land 50-100× higher than an equivalent GPT-4o workload. We have seen individual Claude 4.6 pipelines hit roughly $100 per day, while domestic budget models handled a full weekend of batch processing for under ¥9.

Anthropic's pricing page confirms the per-token math. Finout's 2026 provider comparison puts it bluntly: choosing the wrong model for your workload "could mean paying 2-5× more than you need to". That estimate is conservative when you factor in response-length differences at production scale.

Two Tracks, Not One

We split our AI workload into two lanes, borrowing a pattern that Augment's model routing guide describes as "task-type routing":

  1. Intelligence track: Claude Sonnet or Opus. Strategic reasoning, competitive analysis, creative audits, client-facing summaries. These run once or twice per decision, not once per minute. Sonnet's deeper reasoning justifies its premium when the output shapes a campaign's direction.
  2. Execution track: GPT-4o Mini or Haiku. Daily monitoring, ad performance checks, data extraction, report generation. These fire hundreds of times per day. Speed matters more than nuance; cost compounds fast.

This is not a permanent assignment. When a task graduates from experiment to routine, we migrate it down-tier. The model that helped prototype a workflow rarely stays once that workflow stabilizes.

What We Actually Route Where

The routing logic is deliberately simple. No orchestration layer, no weighted fallback, just three rules tied to task profiles:

  • Decision-heavy work → Claude Sonnet or Opus. Campaign strategy reviews, creative audits, competitive landscape analysis. Low volume, high stakes.
  • Volume work → GPT-4o Mini. Ad status checks, metric pulls, automated alerts, data normalization. Hundreds of calls per day.
  • Overnight batch jobs → the cheapest model that gets the job done. If a domestic model handles weekend data prep for pocket change, there is zero reason to idle a premium tier.

The ongoing AI pricing collapse makes this routing more viable every quarter. Budget tiers have gotten genuinely capable. The "route everything to Sonnet" era lasted about six weeks before the invoices forced a different conversation.

Caching Is the Cheat Code

Both Anthropic and OpenAI now offer roughly 90% off cached input tokens. For workflows that hit the same system prompts or reference documents repeatedly (ad account structures, brand guidelines, campaign naming conventions) prompt caching turns a $3.00/M input cost into $0.30/M.

We structure agent prompts so that static context sits in cacheable blocks. Dynamic data (today's metrics, the specific creative being reviewed) gets appended at the end. Hit rates above 80% are realistic with this layout, and the cost reduction is immediate.

What This Means for AI Operators

Model routing is not an infrastructure decision. It is an operator decision. The person running paid media workflows knows which tasks need Sonnet-level reasoning and which just need a fast, correct answer. When we embed routing logic into agent pipelines, we are encoding operator judgment into code, not replacing it.

That is the lens Auxora operates through: model selection as a layer that marketing operators control, not one buried in an engineering config file. The cost difference between smart routing and default-everything-to-Sonnet is the difference between a sustainable AI workflow and one that burns budget quietly until someone finally reads the invoice.

Check your model routing. If every agent call in your stack hits the same tier, you are probably leaving 70-90% of your AI spend on the table.

Sources

Keep reading