How I think about reducing AI costs
A practical framework for reducing AI costs - auditing token spend, killing legacy models, moving workloads to open weights providers, and fixing the agent and tool inefficiencies that quietly burn tokens.
Martin Alderson
Web development, AI tooling, and building better software
Latest posts
Watch out for cache read costs
For long agentic sessions cache reads are most of the bill, not input or output tokens. KV caches have shrunk dramatically while cache read pricing has barely moved.
I'm (mostly) picking models on speed now, not intelligence
For the first time I'm choosing my daily driver models on tokens per second rather than raw intelligence. Why ~100tok/s might be the new 100ms, where the speed gains stop mattering, and the price war that's coming.
The first known runaway AI agent - or a very bad marketing stunt?
Breaking down the Hugging Face security incident caused by OpenAI's own models during a benchmark run - the sandbox escape, the package proxy, and whether it's really a marketing stunt.
Winners and losers in the coming AI margin collapse (part 2)
Part two: as AI inference margins collapse, who actually captures the value? The hardware supply chain and consumers win, model inference commoditises, and the frontier labs' escape routes are managed agents and staying ahead.
GLM 5.2 and the coming AI margin collapse (part 1)
GLM 5.2 is the first open weights model I'd call a genuine competitor to Opus and GPT for agentic work - at ~15-20% of the price. Part one of why AI inference margins are about to collapse.
Expert-aware quantisation: near-Q4 quality at near-Q2 size?
Profiling a MoE model to find which experts matter for a specific task, then quantising the cold ones hard. The result: near-Q4 quality at near-Q2 size for local models.
A brief history of KV cache compression developments
How KV cache compression - from MQA and GQA to MLA and linear-attention hybrids - quietly unlocked the long context windows that make modern agentic LLMs possible.
xAI is looking more like a datacentre REIT than a frontier lab
xAI is renting huge amounts of GPU capacity to Anthropic and Google. Financial engineering ahead of the SpaceX IPO, a real compute shortage, or a genuine datacentre advantage? Probably all three.
Is datacentre sovereignty really that important?
The UK is obsessed with building AI datacentres at home. But the arguments for sovereignty - latency, tax, control - mostly don't hold up.
I went on the Built for Turbulence podcast
I joined Radical's Built for Turbulence podcast to talk about what AI agents are doing to the economics of software, the Figma Trap, and why running human-written code without AI audit is going to start looking reckless.
What's going on with Gemini?
Google's Gemini 3.5 Flash was the headline model at I/O - fast, but expensive and middling at coding. Why it makes more sense as a model built for Google itself, the TPU advantage, and Google's real weakness in coding agents.
Managed agents are the new Lambda
Managed agents (cloud-hosted agent harnesses) are powerful, but locking yourself into a frontier lab's platform now is risky - here's why and what to do instead.
Open weights are quietly closing up - and that's a problem
Open weights models keep frontier labs honest on price. If they disappear, we end up with a handful of oligopolists extracting consumer surplus.
29th August 2026: a scenario
A fictional scenario about what AI changes for cloud security, written because the technical version of the argument doesn't land with anyone except engineers.
Figma's woes compound with Claude Design
Figma's reliance on non-designer seats made it uniquely exposed to AI. Claude Design's launch deepens the problem.
A little tool to visualise MoE expert routing
I built a small tool to visualise how Mixture of Experts models route tokens through different experts. It's genuinely fascinating to watch.
Has Mythos just broken the deal that kept the internet safe?
What Anthropic's Mythos research preview tells us about the trajectory of frontier models, sandbox escapes, and the cybersecurity risk ahead.
What next for the compute crunch?
AI compute demand is growing exponentially while supply constraints bite hard. The next 18-24 months are going to be defined by shortages, rationing and price discovery.
Telnyx, LiteLLM and Axios: the supply chain crisis
A cascading wave of supply chain attacks has hit npm and PyPI in under two weeks. LLMs are making it worse, and current mitigations aren't enough.
Using agents and Wine to move off Windows
How I used Claude Code to fix Linux desktop issues, get 'garbage'-rated Windows apps working in Wine, and what it means for software ecosystems
Why Claude's new 1M context length is a big deal
Anthropic's 1M token context window on Opus 4.6 and Sonnet 4.6 is a genuine breakthrough - and they're not even charging more for it.
How to use the Qwen 3.5 LLMs to OCR documents
Using Qwen 3.5 open weights models to OCR scanned PDFs - locally on consumer hardware or via OpenRouter for pennies
No, it doesn't cost Anthropic $5k per Claude Code user
The viral claim that Anthropic loses $5,000 per Claude Code subscriber doesn't survive basic scrutiny. Let's do the actual maths.
Is the AI Compute Crunch Here?
Claude Code has 2-3 million users. That's 1% of knowledge workers. The compute math gets scary from here.
Why on-device agentic AI can't keep up
On-device AI agents sound great in theory. The maths on KV cache scaling, RAM budgets, and inference speed says otherwise.
Using OpenCode in CI/CD for AI pull request reviews
Why I replaced SaaS code review tools with OpenCode running in CI/CD pipelines - cheaper, more secure, and works with any Git provider
Which web frameworks are most token-efficient for AI agents?
I benchmarked 19 web frameworks on how efficiently an AI coding agent can build and extend the same app. Minimal frameworks cost up to 2.9x fewer tokens than full-featured ones.
Who fixes the zero-days AI finds in abandoned software?
Anthropic's red team found 500+ critical vulnerabilities with Claude. But they focused on maintained software. The scarier problem is the long tail that nobody will ever patch.
Attack of the SaaS clones
I cloned Linear's UI and core functionality using Claude Code in about 20 prompts. Here's what that means for SaaS companies.
How to generate good looking reports with Claude Code, Cowork or Codex
A step-by-step guide to extracting your brand design system and generating on-brand PDF reports and slide decks using coding agents.
Self-improving CLAUDE.md files
A simple trick to keep your CLAUDE.md and AGENTS.md files updated using the agent's own chat logs - turning a tedious chore into a 30 second job.
Wall Street just lost $285 billion because of 13 markdown files
Anthropic's 'legal tool' that triggered a $285bn selloff is 156KB of markdown. The panic reveals a hard truth about the future of software.
Two kinds of AI users are emerging. The gap between them is astonishing.
A bifurcation is happening in AI adoption - power users shipping products in days versus everyone else generating meeting agendas. Enterprise tool choices are accelerating the divide.
Turns out I was wrong about TDD
I used to be a TDD sceptic - too much time writing tests for features that might get deleted. Then coding agents completely changed the economics of software testing.
Why sandboxing coding agents is harder than you think
Permission systems, Docker sandboxing, and log file secrets - why current approaches to securing coding agents fall short and what we might need instead.
The Coming AI Compute Crunch
Why DRAM shortages, not capital, will define AI infrastructure growth through 2027
Which programming languages are most token-efficient?
Comparing token efficiency across 19 popular programming languages using RosettaCode data - from Clojure to C, there's a 2.6x difference.
I ported Photoshop 1.0 to C# in 30 minutes
Using Claude Code to port 120k lines of Pascal and 68k assembly to modern C# - and what this means for cross-platform development
Why I'm building my own CLIs for agents
MCP tools eat thousands of tokens. A simple CLI with instructions in your CLAUDE.md file uses 71 tokens and works brilliantly.
Travel agents took 10 years to collapse. Developers are 3 years in.
Travel agents are the classic example of an industry killed by the internet. Software engineering is facing the same disruption, but the timeline is compressed.
Are we dismissing AI spend before the 6x lands?
Critics are judging models trained on last-gen hardware. There's a 6x wave of compute already allocated - and it's just starting to produce results.
Minification isn't obfuscation - Claude Code proves it
Using ASTs and AI agents to reverse engineer minified JavaScript in minutes instead of weeks
AI agents are starting to eat SaaS
Software ate the world. Agents are going to eat SaaS.
Has the cost of building software just dropped 90%?
Agentic coding tools are dramatically reducing software development costs. Here's why 2026 is going to catch a lot of people off guard.
Are we in a GPT-4-style leap that evals can't see?
Gemini 3 Pro's design capabilities and Opus 4.5's reduced babysitting needs represent a subtle but significant leap that traditional benchmarks completely miss.
I Finally Found a Use for IPv6
Using IPv6 with Cloudflare to run multiple services on a single server without a reverse proxy
How I use Claude Code to manage sysadmin tasks
A practical approach to managing production infrastructure using git-tracked markdown files and Claude Code for small teams
Could Excel agents unlock $1T in economic value?
Software engineers underestimate the scale of Excel usage. With agents now able to work directly in spreadsheets, we're looking at transforming how billions of dollars in business processes are managed.
Are we really repeating the telecoms crash with AI datacenters?
Looking at actual token demand growth, infrastructure utilization, and capacity constraints - the economics don't match the 2000s playbook like people assume