LLM Cost Optimization
An LLM cost and model optimization audit shows where the spend in a production system comes from. It then implements the changes that can lower it: prompt caching, model routing, prompt compression and retry rules. The report compares cost and quality test results before and after.
Fixed-price entry packageLLM Cost AuditEUR 4,2001 application · up to 3 weeks from scopingWhat is includedEvery euro attributed
Per feature, per segment, per model: the bill stops being one number nobody can explain.
Savings ranked by effort
Caching, routing, prompt compression, retry tuning, ordered by saving times effort.
Quality proven, not assumed
An eval run before and after shows the same or better quality at the lower cost.
23Minds runs LLM cost and model optimization audits for engineering teams whose LLM bill grows faster than traffic.
Show the full summary
The engagement instruments the system for per-call cost attribution, audits current usage, recommends and implements provider-side savings (prompt caching, smaller-model routing, prompt compression, retry tuning) and delivers a measured before/after with eval-protected quality. The size of the saving depends on the starting point; it is measured on your system, not promised up front. ISO 9001:2015 quality assurance backing.
What makes this service defensible.
- Top three quick wins implemented in the package
Not only recommended: applied in your code and measured.
- Before and after in writing
The audit report carries the measured saving with the method behind it.
- Fixed price per application
One LLM application, three weeks from scope approval.
- Eval-protected
No optimization ships without an evaluation run guarding the output quality.
The process, phase by phase.
The same workload can cost very different amounts.
Move the sliders to estimate monthly inference spend for a typical AI feature across the major providers. The point isn't the exact number - it's the gap between the cheapest and the premium option: that gap is your optimization headroom.
- Model selection per task, routed dynamically by call profile and stake level.
- Caching and prompt size shift the bill as much as model choice does.
Pick the way in that fits your budget and timing.
Every practice has a free first step, an entry package with a written deliverable and its price listed here, and a full project or retainer quoted after a first call. Once the scope is agreed, the price is fixed, with no hidden costs.
- 1Step 1 · Free
Intro call or self-assessment
60 minutes · 1 engineer · no obligation
FreeTalk to an engineerShow details
- Intro call with an engineer, or the online self-assessment
- A written next step within one business day
- No obligation and no sales deck
- Recommended2Step 2 · Fixed price
LLM Cost Audit
1 application · up to 3 weeks from scoping
EUR 4,200 netFixed price, no hidden costs; extras only from the Not included list.
Ask for this packageShow details
- Cost attribution for one LLM application: per feature, segment, model
- Recommendations ranked by saving times effort
- Top three quick wins implemented
- Written report with the measured effect
- Not included: Implementing the full recommendation list, architecture rework, negotiating contracts with model providers, any guaranteed saving level.
- 3Step 3 · Project or retainer
Full scope, quoted after a first call
Full scope · phased delivery · quoted per phase
from EUR 2,800 net per month, quoted after a first callStart the conversationShow details
- Continuous cost optimization across the platform
- Quality protected by evaluation
- Monthly report with the measured saving
The situations that bring teams to this service.
- LLM spend grows faster than user traffic and nobody knows why
- Prompt cache hit rate is low or not measured; every call pays full price
- All features route to the most expensive model regardless of stakes
Show the other problems (1)
- Retries on transient errors silently multiply the bill
- Engineering teams whose monthly LLM bill grew past EUR 10k and keeps climbing
- CFOs and finance partners asking why LLM spend exceeds plan
- Product engineering leaders shipping AI features whose unit economics are upside-down
- Mid-stage companies preparing for a funding round needing defensible AI cost numbers
Where this goes wrong, and what we do about it.
1. Saving tokens breaks the cache and raises the billWhat stays with you: cache hit rate before and after per prompt family, with the invoice line it maps to
What happensA plugin or a prompt rewrite trims the context to save tokens. The trimmed prefix no longer matches, prompt caching stops hitting, and the same traffic now costs more than before the optimisation. The change looks like a saving in every metric except the invoice.
What we do about itWe measure the actual cache hit rate before touching anything, treat the stable prefix as a structure to protect rather than text to compress, and re-measure after each change so a regression is visible in the same week.
2. The cost is attributed to the model and driven by the behaviourWhat stays with you: spend broken down by cause rather than by model, with the top drivers ranked and one of them fixed
What happensTeams negotiate price per token and compare providers, while the multiplier sits elsewhere: branching times retries times context growth. A documented case ran to tens of thousands of dollars on billions of missed input tokens in a single chain. Changing the model moves the unit price a little; the behaviour moves the total.
What we do about itWe break the spend down by what causes it - retries, branch depth, context growth per turn, per feature and per user segment - and fix the largest driver first, whether that is a routing rule, a retry policy or a product decision.
3. Caching is supported and nobody checked whether it hitsWhat stays with you: measured hit rate per layer, and budget limits with alerting at the level where spend is actually made
What happensThe provider documents prompt caching, so the team assumes it is working. Support is not the same as a hit. Without a measurement, a chain of middleware quietly rewrites the prefix on every call and the cache never engages, which shows up only as a bill that grows faster than usage.
What we do about itWe verify hits against the provider's own counters rather than the documentation, follow the request through every layer that can rewrite it, and set a hard spend ceiling so the next regression stops instead of accumulating.
Deliverables, not deliverable-shaped slides.
- Cost attribution dashboard - per feature, per user segment, per model
- Optimization recommendations ranked by saving × effort
- Implementation of the top wins (caching, routing, prompt rework, retry policy)
- Eval-protected quality measurement: same or better quality after optimization
- Written audit report with measured before/after savings
Questions procurement teams ask.
Where the dates and numbers come from.
- Anthropic, Prompt caching: pricing multipliers (checked 25 September 2026)
- OpenAI, Prompt caching (checked 25 September 2026)
- The Information, Uber CTO Shows How Claude Code Can Blow Up AI Budgets (14 April 2026)
- Bloomberg, Uber Caps Usage of AI Tools Like Claude Code to Manage Costs (2 June 2026)
- Fortune, Uber's CTO on AI spending (7 August 2026)
Talk to an engineer.
Describe where you are with LLM cost optimization. A reply within one business day.