News & guides
The DeepSeek V4 Family Explained: Pro, Flash and When
Two models, one release date, a 3x gap in active parameters and a 3x gap in price. How to split traffic between DeepSeek V4 Pro and Flash sensibly.
· 9 min read
The Electricity Cost of Inference: Estimate It Yourself
Energy per token is a property of your deployment, not of the model. Here is the equation, which term dominates, and how to measure your own figure.
· 8 min read
The Hidden Cost of Failed Runs: Tokens Spent on Nothing
A failed agent run costs full price and produces no output. Why cost per successful run is the only figure worth tracking, and how to compute it.
· 8 min readThe Hidden Cost of Retries: Paying Twice for the Same Answer
Retry logic makes pipelines reliable and quietly doubles spend on the items that need it. How to measure what retries actually cost you.
· 8 min read
The Lost-in-the-Middle Problem: Position Beats Relevance
Models retrieve facts from the start and end of a long prompt far more reliably than from the middle. Why that happens and how to arrange prompts around it.
· 8 min read
The MCP Security Model: Trust Boundaries You Actually Have
MCP standardises connection, not trust. Here are the real boundaries in an MCP deployment and the attacks that cross them, with practical mitigations.
· 9 min read
The MiniMax Family Explained: M3 and Native Multimodality
MiniMax M3 pairs a 1M window with native multimodality, and its published pricing genuinely disagrees between sources. What to verify before you commit.
· 9 min read
The Moonshot Kimi Family Explained: K2.6 and K3
Two very different models share the Kimi name. What separates K2.6 from K3 on scale, context and licence, and which one your workload actually wants.
· 9 min read
The Qwen Family Explained: The Models You Can Own
Qwen is the tier that runs on hardware you have. What the 3.6 generation offers, why dense matters, and how to read a family with many sizes.
· 9 min read
The Real Cost of Migrating Between Model Providers
Changing the base URL takes an afternoon. Re-tuning prompts, rerunning evals and dual-running takes weeks. A worked estimate structure you fill in with your own numbers.
· 10 min read
The Real Cost of Open vs Closed Models
Open weights cost nothing to download and plenty to run. Here is how licence, hosting and switching costs actually compare, with current per-token rates.
· 8 min read
The ROI of Flat-Rate Billing for AI Development Work
The break-even maths for flat-rate AI plans, why usage variance decides the answer more than the average does, and the workloads where metered billing still wins.
· 9 min readReady to try it?
Every model behind one key. Start free and upgrade when it sticks.