7 min read
What xhigh Effort Bought Me on Short Tasks: Tokens, Not Answers
ultrathink now runs at xhigh effort in Claude Code. I ran six exact-answer tasks 96 times from low to max: every level got them right; higher just cost more.
Cost Optimization
77 articles
7 min read
Haiku 5.5 vs Opus 5.5 on Real Repo Chores in Claude Code
I gave Haiku 5.5 and Opus 5.5 the same five chores in this blog's repo. Both passed every run. Haiku cost 25 times less, and the diffs show what Opus adds.
7 min read
BigQuery Mirroring Is GA: Zero-ETL Multicloud Has a Bill Too
Fabric mirroring for BigQuery is GA, but it moves cost to Google: CDC compute, egress, keyless reseeds and schema ownership decide when to mirror or copy.
9 min read
Foundry Routines Are GA: Scheduled Agents Without Building a Scheduler
Foundry routines run an agent on a timer, a schedule or a GitHub or Teams event. When to use them, when to use a workflow, and guardrails for unattended runs.
8 min read
GPU-Accelerated Fabric Warehouse: How to Read the 7x Claim
Fabric Data Warehouse GPU query acceleration is in preview. Why the 7x benchmark won't predict your bill, and how to run a fair A/B test on your own queries.
9 min read
GPT-5.6 Sol, Terra or Luna: Pick the Tier per Job, Not per App
GPT-5.6 in Foundry has a 5x price spread across Sol, Terra and Luna. Choose a tier for each workflow step, back it with eval gates, then pick deployment types.
8 min read
Four LLM Caches and How to Tell Which One Will Pay Off
Prompt, response, semantic and retrieval caches save different money and add different risks. Here's a break-even check to run before you build one.
7 min read
Latency Budgets as a Cost Control: Pricing Each AI User Journey
Set a latency budget per user journey, then let it choose the Azure OpenAI deployment tier, model and reasoning effort you actually pay for.
8 min read
Where LLM Latency Goes, and Which Token Cuts Actually Help
Input and output tokens cost and delay answers differently, so split latency into first-token and generation time before trimming anything.
7 min read
A Token Waste Audit: Cut What the Answer Doesn't Use
Profile each LLM request by segment, rank the cuts by quality risk, and keep only the token savings that survive a paired evaluation against the baseline.
8 min read
Why Your Azure OpenAI Prompt Cache Keeps Missing
Azure OpenAI caches prompt prefixes automatically, yet many apps see few cache hits. How to measure the hit rate and fix the prompt shapes that break it.
8 min read
From F64 to F32: Three Fabric Sizing Mistakes and the Fixes
Why we moved Microsoft Fabric from F64 to F32: the three sizing mistakes, the monitoring and alerts we added, and when going smaller is the wrong call.
8 min read
Cost Per Query: Measure the Unit Cost Before You Cut It
Treat cost per query as the unit metric for an LLM app: decompose it from the usage object, then cut it in order with model choice, caching, context and batch.
9 min read
Microsoft Fabric Reality Check: What Holds Up and What Hurts
An honest look at running Microsoft Fabric on an F64: what OneLake and Direct Lake deliver, how throttling really works, and what the bill looks like.
8 min read
Azure OpenAI Hidden Costs: What Your Token Estimate Misses
Re-sent history, hidden reasoning tokens, billed failures, idle fine-tuned deployments and re-embedding: where Azure OpenAI bills drift from the estimate.
1 min read
The Ultimate Azure Cost Optimization Checklist for 2026
Start the new year with optimized cloud spend. These changes compound over time into significant savings.
1 min read
Kubernetes Cost Optimization on AKS: Lessons from 2025
Kubernetes cost optimization is continuous. Review these metrics weekly and adjust based on actual usage patterns.
1 min read
Semantic Caching for LLM Applications: Reducing Costs and Latency
Traditional caching requires exact matches. Users asking "What is Azure?" and "Can you explain Azure?" would generate two separate API calls. Semantic…
1 min read
Optimizing GPU Utilization for AI Workloads on Azure
Many organizations run GPU workloads at 30-40% utilization, paying for idle compute. Understanding workload patterns and implementing optimization…
1 min read
Cost Optimization Strategies for Azure OpenAI Deployments
Balance cost and quality by routing requests to appropriate models based on task requirements.
1 min read
Semantic Caching for LLM Applications: Reducing Costs and Latency
Unlike exact-match caching, semantic caching uses embeddings to find similar queries even when worded differently.
1 min read
Cost Optimization Strategies for Azure OpenAI Workloads
Semantic caching avoids redundant API calls for similar queries. Reduce prompt length by removing unnecessary context. Use batch processing with the Batch…
1 min read
Azure Functions with AI: Serverless Inference Patterns
Use consumption-based billing to pay only for actual inference time. Implement caching for repeated queries and batch similar requests to amortize cold…
1 min read
Reserved Capacity for AI: Optimizing Long-Term Costs
Strategic use of reserved capacity can reduce AI costs by 30-50% for predictable workloads.
1 min read
FinOps for Microsoft Fabric
1. Start with visibility : You can't optimize what you can't see 2. Allocate costs : Enable accountability through showback/chargeback 3. Automate policies…
1 min read
Microsoft Fabric Capacity Optimization
1. Monitor utilization patterns : Understand your usage before optimizing 2. Stagger workloads : Avoid concurrent peaks 3. Right size jobs : Use appropriate…
1 min read
Cost-Based LLM Routing: Optimizing AI Spend
The difference is dramatic. A 1000-token task costs $0.02 with GPT-4o but $0.0002 with GPT-4o-mini.
1 min read
Cost Reduction Strategies with GPT-4o
GPT-4o is already 50% cheaper than GPT-4 Turbo, but there are more ways to optimize costs. Here are practical strategies I use in production.
1 min read
Serverless Model Serving: Pay-Per-Request ML Inference
Serverless model serving eliminates infrastructure management while providing cost-effective, scalable ML inference. This guide covers implementing…
1 min read
Model Selection Strategies: Right-Sizing Your AI Workloads
While the AI community anticipates Claude 3's tiered model approach, it's worth examining how to right-size your AI workloads today. Understanding when to…
1 min read
Semantic Caching: Intelligent Response Reuse for LLMs
1. Start conservative Higher threshold (0.95+) for critical applications 2. Tune with data Evaluate on your actual query patterns 3. Monitor quality Track…
1 min read
Cost Optimization for LLMs: Strategies That Work
1. Cache aggressively Identical queries don't need re processing 2. Tier your models Match model to task complexity 3. Optimize prompts Shorter prompts =…
2 min read
LLM Routing: Intelligent Model Selection for Cost and Quality
" in query or self. has code indicators(query lower): return ModelTier.STANDARD Long queries often need more capability if word count 200: return…
7 min read
Keeping Azure AI Search Vector Indexes Small Without Compression
Azure AI Search has no built-in vector compression yet. How to estimate vector index size, read the quota, and cut it with chunking and dimension choices.
8 min read
Fabric Throttling Explained: Reading Carryforward Before Users Do
How Fabric smoothing, carryforward and the four throttling stages work, how to read them in the Capacity Metrics app, and which levers actually help.
7 min read
Shipping on GPT-4 Turbo Preview in Azure OpenAI: A Production Checklist
GPT-4 Turbo is still a preview model on Azure OpenAI; here's how I'd deploy, budget, retry and test around 1106-preview before real users touch it.
1 min read
Cost Optimization in Microsoft Fabric
Cost optimisation isn't about cutting features — it's about aligning spend with business value. During holiday slowdowns I help teams identify idle…
1 min read
Prompt Compression: Reducing Token Usage
Tomorrow we'll explore semantic compression techniques in more depth. LLMLingua Prompt Compression Paper Token Optimization Guide
1 min read
Cost Optimization for AI Workloads
AI infrastructure costs compound quickly. A few patterns I've been applying with clients to bring Azure OpenAI and Azure ML spend under control: at the API…
1 min read
Spot Instances for ML: Cost-Effective Training at Scale
Tomorrow we'll explore cost optimization strategies for AI workloads. Azure Spot VMs Checkpointing Best Practices Azure ML Cost Management
1 min read
Caching Strategies for LLM Applications
Effective caching significantly reduces LLM costs and latency. Tomorrow, I will cover response streaming patterns.
1 min read
Token Budgeting: Cost-Effective LLM Applications
Token costs for GPT-4 in mid-2023 are real enough to design around: roughly $0.03 per 1K input tokens and $0.06 per 1K output tokens for the 8K context…
1 min read
LLM Caching Strategies: Reducing Costs and Latency
LLM caching strategies are essential for production systems. By combining exact matching, semantic similarity, and intelligent invalidation, you can…
1 min read
Azure OpenAI Cost Optimization Strategies
One token is roughly 4 characters in English. A 1000-word document is about 1300 tokens. Start with these strategies and refine based on your usage…
8 min read
Counting and Capping Tokens in the Azure OpenAI Preview
How to count tokens with tiktoken, size max_tokens, log real usage and cap spend for Azure OpenAI completion models in the January 2023 preview.
8 min read
GPT-3.5 on Azure OpenAI: What You Get Before text-davinci-003
What GPT-3.5 means on Azure OpenAI in January 2023: which models are in the catalogue, what they cost, and how to design for text-davinci-003's arrival.
1 min read
Right-Sizing Azure Resources: A Data-Driven Approach
1. Collect metrics CPU, memory, network, IOPS 2. Analyze patterns Peak, average, trends 3. Identify candidates Under utilized resources 4. Recommend changes…
1 min read
Auto-Scaling Best Practices in Azure
Auto scaling is powerful but requires careful tuning. Start with conservative settings, monitor behavior, and adjust based on real data. Combine metric…
1 min read
Azure Spot Instance Strategies for Cost Savings
Spot VMs use Azure's excess capacity at significant discounts. Key characteristics: Up to 90% discount Can be evicted anytime 30 second eviction notice Best…
1 min read
Reserved Capacity Planning for Azure
Azure offers reservations for: Virtual Machines Azure SQL Database Cosmos DB Storage Azure Synapse Analytics App Service Azure Databricks Azure Cache for…
1 min read
Azure Cost Optimization Tips for 2023
Cost optimization is an ongoing practice, not a one time project. The combination of right sizing, reservations, storage optimization, and operational…
1 min read
Black Friday Azure Tips: Cost Optimization for Peak Traffic
Don't wait for auto scaling to catch up during traffic spikes: Reduce database load with intelligent caching: Use read replicas and query optimization:…
7 min read
Measure Your Azure Carbon Footprint Before You Try to Shrink It
How to use the Emissions Impact Dashboard for Azure to baseline cloud emissions, find the workloads that matter, and cut waste with tools you already have.
7 min read
Azure Advisor Score (Preview): Read It as a Trend, Not a Grade
How the preview Azure Advisor Score is weighted, why dismissals inflate it, and how to pair it with Resource Graph queries before you report it upward.
7 min read
Azure Hybrid Benefit as Code: Only Claim Licences You Can Prove
Set Azure Hybrid Benefit in Bicep, audit it with Resource Graph and Azure Policy, and size core licences so the discount never becomes a compliance finding.
8 min read
Sizing an Azure VM Reservation: Buy the Floor, Not the Average
How to work out how many Azure Reserved VM Instances to buy: find your steady usage floor, use instance size flexibility, and read the recommendations API.
8 min read
Designing for Eviction: Architecture Patterns for Azure Spot VMs
Spot VMs only save money if the design expects eviction: a regular base tier, diversified Spot pools, queue-driven workers and a Scheduled Events handler.
7 min read
Azure Spot VMs: Eviction Policy, Max Price and the 30-Second Notice
Azure Spot VM eviction explained: capacity vs price eviction, Deallocate vs Delete, max price, and handling the Scheduled Events Preempt notice in Python.
7 min read
AKS Cost Review: Fix Requests First, Then Nodes, Then Billing
A practical order for cutting AKS spend in early 2022: right-size pod requests, tune the cluster autoscaler, then use spot, stop/start and reservations.
1 min read
Cost Optimization Strategies for Azure Workloads
Cost optimization in 2021 became a core cloud competency. The tools are powerful; success requires discipline and cultural change.
2 min read
Job Clusters vs All-Purpose Clusters: Choosing the Right Approach
The job cluster versus all-purpose cluster decision in Databricks is primarily a cost decision: all-purpose clusters stay running between tasks (you pay for…
2 min read
Cost Optimization with AKS Spot Node Pools
AKS Spot node pools are the cost reduction lever that can cut compute spend by up to 90%—at the cost of accepting that Azure may evict spot nodes with a…
1 min read
Azure Blob Lifecycle Management: Automating Data Tiering and Deletion
Lifecycle management policies automate the tedious work of data tiering and retention, ensuring compliance while optimizing storage costs without manual…
1 min read
Azure Storage Tiers: Optimizing Costs with Hot, Cool, and Archive
Azure Storage tiers provide a powerful mechanism for balancing performance and cost, enabling organizations to store massive amounts of data economically…
1 min read
Azure SQL Database Serverless: Auto-Scaling and Auto-Pause for Variable Workloads
Serverless is perfect for dev/test environments, infrequently used applications, and workloads with predictable quiet periods. For consistent…
1 min read
Consumption vs Premium Azure Functions - Making the Right Choice
Feature Consumption Premium Scaling 0 to 200 instances 1 to 100 instances Cold starts Yes No (pre warmed) Max timeout 10 min (default 5) 60 min (default 30)…
2 min read
Maximizing Value with Azure Hybrid Benefit
Azure Hybrid Benefit is the discount that nobody in finance remembers to claim and every consultant I know flags in the first cost review. If you have…
2 min read
Maximizing Savings with Azure Reservations
Reservations are the Azure cost conversation nobody has early enough. Pay-as-you-go is convenient; reserved instances are a commitment that rewards the…
7 min read
Triaging Azure Advisor Cost Recommendations Before You Act
Which Azure Advisor cost recommendations are safe to act on, which need judgement, and how to pull, tune and suppress them with the CLI and Python SDK.
7 min read
Cosmos DB Serverless vs Provisioned: Working Out the Break-Even
How to work out whether Azure Cosmos DB serverless (preview) beats 400 RU/s of provisioned throughput, with the break-even maths and a way to measure it.
8 min read
Synapse Serverless SQL After GA: Designing a Lake Query Layer
How I'd structure Synapse serverless SQL now it's GA: a user database, credentials, views over the lake, CETAS, and habits that keep the per-TB bill small.
1 min read
Azure Blob Storage Lifecycle Management
Savings: $1,700/year for 1TB Set up lifecycle policies on day one. Future you will thank past you for the cost savings.
1 min read
Azure SQL Elastic Pools: Cost-Effective Multi-Tenancy
Elastic pools are ideal for SaaS multi-tenancy where tenant databases have varied, unpredictable loads.
9 min read
Windows Virtual Desktop for Remote Work: Building It Properly
How to stand up Windows Virtual Desktop on the new Azure Resource Manager model: host pools, FSLogix profiles on Azure Files, scaling, and when not to use it.
7 min read
Application Insights for ASP.NET Core: Telemetry Worth Paying For
Instrumenting an ASP.NET Core 3.1 API with Application Insights: connection strings, sampling, custom metrics and alerts, without paying for noise.
10 min read
Azure SQL Elastic Pools: When Pooling Saves Money and When It Doesn't
How to size an Azure SQL Database elastic pool for database-per-tenant SaaS, measure tenant load first, and spot the workloads where pooling costs more.
9 min read
Azure Blob Lifecycle Policies: Tiering and Deletion Without Surprises
How I design Azure Blob Storage lifecycle policies to move data to Cool and Archive and delete it, plus the early-deletion and rehydration traps to avoid.