Blog
Thoughts on AI, development teams, and the future of work.
- →
GLM vs Claude for Coding Agents: The Gap Got Small, But It Didn't Close
An open-weight model you can download for free now trades blows with Claude on coding benchmarks at a fraction of the price. So why do experienced developers still keep Claude around? Because a coding agent is not a benchmark.
August 4, 2026
- →
Residency Isn't Sovereignty: What 'EU Data Residency' Actually Buys Your LLM Traffic
Flipping on an 'EU data residency' toggle checks a compliance box, but residency (where the bytes sit) and sovereignty (whose law can compel them) are different guarantees. As of mid-2026, the gap between them is where the real risk lives.
August 2, 2026
- →
How to Benchmark Context Compression Without Fooling Yourself
A compression ratio on its own is a vanity metric. Here is a methodology for measuring whether context compression actually helps, instead of measuring whether it looks good.
August 2, 2026
- →
Why Message 50 Costs More Than Message 5
Your agent has no memory. It re-reads the whole conversation on every turn, which is why a long session gets more expensive with each message even when you barely type anything. Here is the honest math, and what caching and giant context windows actually change.
July 23, 2026
- →
Your LLM Fallback Plan Might Be Making Outages Worse
Multi-provider failover for LLM apps isn't the microservices playbook. The reflexes that work for fast, cheap, idempotent APIs can amplify outages and duplicate real side effects. Here's what actually holds up in 2026.
July 23, 2026
- →
Three Ways to Talk to a Model: Anthropic Messages, Chat Completions, and Responses
There is no single standard for sending a prompt to an LLM. There are three competing wire formats, they disagree on nearly every detail that matters, and the gap between them is why swapping providers is rarely a one-line change.
July 23, 2026
- →
How Much Does Claude Code Cost Per Developer? Anthropic Says $150 to $250 a Month. Uber's Power Users Paid Ten Times That.
There is finally a published benchmark for what Claude Code costs per developer. It is also nearly useless on its own, because the same seat that averages $200 a month produced $2,000 bills at Uber. The gap between the average and the tail is the whole story.
July 22, 2026
- →
Switching Models Mid-Session Is Easy. Keeping the Reasoning Is the Hard Part.
Translating a live agent session between Claude, GPT, and Gemini looks like a format-conversion problem. For text and tool calls it mostly is. Reasoning state is the wall nobody has climbed.
July 22, 2026
- →
Your AI 'Audit Logs' Are Probably Auditing the Wrong Thing
The feature your LLM gateway ships under the name 'Audit Logs' usually tracks who changed a setting, not what your models actually saw. On the AI request path, that gap is where compliance quietly breaks.
July 22, 2026
- →
OpenRouter Routes Your Requests. It Never Touches the Thing Driving Your Bill.
OpenRouter is the best router in the category, and routing decides which model serves a request. It doesn't change how many tokens you send. In agent sessions, that's the whole bill.
July 19, 2026
- →
You Can Finally Cap Every Developer's AI Spend. The Cap Isn't the Hard Part.
In 2026 every major coding-agent vendor shipped per-developer spend limits within a few weeks of each other. Setting the cap is now easy. Making spend predictable and cheap is the part that still isn't, and the limit barely touches it.
July 19, 2026
- →
Why 'We Have a Zero-Retention Agreement' Doesn't Solve Your PII Problem
Signing a zero-data-retention deal with your LLM provider feels like solving PII risk. It solves a much narrower problem than most teams think, especially once agents enter the picture.
July 18, 2026
- →
The Rule You Set Once Should Survive Every Session, Even the Long Ones
Every coding agent vendor solved rule distribution in 2026. Almost none of them solved rule persistence, the harder problem of whether a policy still carries weight two hundred messages into a long session.
July 18, 2026
- →
Self-Hosting vs. API: The Real Break-Even Math
Everyone repeats a rule of thumb about when self-hosting an open-weight model beats paying for API access. We built the actual calculation from primary benchmark data, and the number that decides it isn't the one people quote.
July 16, 2026
- →
Claude Code Isn't the Product. It's the Top of the Funnel.
Anthropic spent six months in 2026 breaking, half-fixing, and un-fixing the way your subscription pays for compute. That thrashing tells you exactly what Claude Code is actually for.
July 15, 2026
- →
Decoupling the Harness from the Horse: The Case for Model-Agnostic Coding Agents
If you can't swap the model behind your AI coding agent in five minutes, you don't own your workflow. You're renting it. Here's what mid-2026's pricing wars and vendor lock-in stories actually prove.
July 15, 2026
- →
Routing Claude Code to a Different Model: The Setup, and the Fine Print
The env var trick for pointing Claude Code at GLM, Kimi, or DeepSeek works. Here's what actually happens under the hood, and the gotchas most guides skip.
July 15, 2026
- →
Reclaiming the Context Window: The End of Context Clearing by Proxy
Sub-agents have become the duct tape of AI coding. The real motive isn't specialization — it's clearing context by proxy. There's a better way.
May 12, 2026
- →
Welcome to AJNT
Introducing AJNT — your AI team, managed for you.
March 10, 2026