ModelsOpenAI o3 Guide: The Tool-Using Reasoner Before Its Retirement
o3 put tool use inside the chain of thought, then took an 80% price cut. What it does well, where it hallucinates, and how to move off it before the API shuts it down.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
Models
ModelsPalmyra X5: Writer's 1M-Context Agent Model Before Its Sunset
Palmyra X5 pairs a 1M-token window with cheap input and an 8K output cap, and WRITER deprecates it on 14 December 2026. Where it fits and what replaces it.
Read
ModelsPareto 26.10 Preview: Reading a Composite Model's Debut
Pareto 26.10 Preview was Unbiased's opt-in build of a blended multi-model API, replaced by the stable 26.10 two weeks later. What it was and how to judge composites.
Read
ModelsPareto by Unbiased: A Composite Model Behind a Moving Alias
Pareto is an opaque multimodal composite whose model ID now redirects to the newest release. What is known, what is not, and how to test or pin it safely.
Read
ModelsPerceptron Mk1: A Video Model That Answers in Coordinates
Perceptron Mk1 is a closed vision-language model for video and embodied reasoning that returns points, boxes and timecodes. What it does, its limits and how to test it.
ReadModelsPerceptron Mk1.5: A Perception Model That Drives the Loop
Perceptron Mk1.5 adds audio, object tracking and tool calling to Mk1, and cuts latency. What changed, how to read its benchmarks and how to use it in a control loop.
ReadModelsPerplexity Sonar Guide: Cheap Search Answers and Hidden Fees
Sonar is Perplexity's entry search model: web-grounded answers with sources at $1 per million tokens. Why the request fee dominates the bill, and what its 2026 move means.
Read
ModelsPhi-4 Guide: A 14B Reasoner That Knows Few Facts
Microsoft's Phi-4 is a 14B dense, MIT-licensed model strong at maths and reasoning but weak at factual recall, with a 16K window. Where that profile fits and fails.
ReadModelsQwen-Plus Guide: A Price Tier That Changes Model Underneath
Qwen-Plus is an alias that has pointed at several models since 2024. How its snapshots differ, how its tiered pricing works, and why to pin a dated version.
ReadModelsQwen2.5 72B Instruct in 2026: When a 2024 Model Still Fits
Qwen2.5 72B Instruct is two years old and still on every price list. Its specs, its licence terms, its context trap, and the narrow jobs it still does well.
ReadModelsQwen2.5 7B Instruct: The Small Workhorse Two Generations On
A 7.6B dense, Apache-licensed model from September 2024. What Qwen2.5 7B Instruct is still good for, where it falls short, and when to move up to Qwen3.
ReadModelsQwen2.5 Coder 32B Instruct: The Dense Coder You Can Still Own
A 32.5B dense, Apache 2.0 coding model from late 2024 that scored 73.7 on Aider at launch. What it does well, its 32K and tool-calling limits, and its successors.
ReadShowing 241–252 of 731 articles