Blog
Mercury 2.5 vs Solar Pro 4: is a diffusion LLM really faster?
Mercury 2.5 is a diffusion LLM that generates tokens in parallel. On the same 43 prompts it answered 7.3x faster than Solar Pro 4 on average, while writing about half as much text and using 1.8x the output tokens.
GPT-6 Astra vs Claude Fable 5.1: same price, 2.4x the bill
GPT-6 Astra and Claude Fable 5.1 cost the same per token on FastMetal. On the same 43 prompts, Fable 5.1 used more output tokens on every one, about 2.4x in total, and took twice as long.
Gemini 3.7 Flash vs GPT-5.6 Luna: multimodal or cheapest text
Google's Gemini 3.7 Flash and OpenAI's GPT-5.6 Luna are both cheap, fast models for high-volume work. Luna costs about a third as much; Gemini takes audio and video. Here is how to pick.
Grok 4.6 vs GPT-5.6 Terra: STEM depth or long context
xAI's Grok 4.6 and OpenAI's GPT-5.6 Terra cost about the same to send a prompt to, but Grok's output is half the price and Terra's context is twice the size. Here is how to choose.
GPT-5.6 Terra vs Claude Sonnet 5: which mid-tier model to use
GPT-5.6 Terra and Claude Sonnet 5 are OpenAI's and Anthropic's mid-tier workhorses. Same input price, 1M-token context on both. Sonnet 5 for multi-step agents and coding, Terra for everyday OpenAI-style work.
RAG vs long context: is stuffing the whole corpus cheaper?
A million-token context window didn't kill RAG. Sending 1M tokens per request uses about 200 times the input of retrieving 5,000, caching narrows the gap but never closes it, and accuracy drops in the middle.
Muse Glimmer 30B vs Qwen3.6 27B: two ~30B open models compared
Meta's Muse Glimmer 30B and Alibaba's Qwen3.6 27B are both Apache 2.0, ~30B dense models. Qwen is stronger on coding; Glimmer's output costs half as much and its cache discount makes agent loops cheap.
Qwen3.8 Max vs Claude Opus 5: Opus for hard code, Qwen for volume
Qwen3.8 Max vs Claude Opus 5: what the published benchmarks actually say, why Qwen's agent scores need a discount, and what the same large coding task costs on each in USD.
Hermes vs OpenClaw: two open-source agents, one API key
Hermes and OpenClaw get compared a lot, but they do different jobs. What each one is for, how to connect both to FastMetal, and what to watch when an always-on agent runs on pay-as-you-go.
Hermes Agent with FastMetal: a coding agent without a Codex subscription
How to point the open-source Hermes Agent at FastMetal's API, and how to tell whether pay-as-you-go beats a Codex subscription for the way you actually work.