Blog
Kimi K3.1 Pro vs Claude Sonnet 5.5: same price, 2.7x the bill
Kimi K3.1 Pro and Claude Sonnet 5.5 cost the same per token on FastMetal. On 36 identical prompts, Kimi billed 2.7x the output tokens and took a median 150 seconds to Sonnet's 14. The visible answers were about the same length.
GPT-6.1 Terra and Terra Pro are now in FastMetal chat
OpenAI's GPT-6.1 Terra and Terra Pro, released October 26, 2026, are now in FastMetal's chat at about half the price of GPT-6.1 Sol. We measured what Pro mode really costs per question.
Kimi K3.1 Pro API: pricing and how to use it
Kimi K3.1 Pro, Moonshot AI's successor to Kimi K3 released October 23, 2026, is now on FastMetal's API and chat: a 1M-token reasoning model at two-thirds of K3's price.
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash: same rate, 8x the cost
MiniMax M3.1 Flash and DeepSeek V4.1 Flash bill output at the same rate. On the same 14 prompts, M3.1 Flash thought its way to 8x the output tokens and 8x the cost.
MiniMax M3.1 Flash API: pricing and how to use it
MiniMax M3.1 Flash, released October 19, 2026, is now on FastMetal's API and chat: a fast agent and coding model with a 1M-token context at roughly half the price of MiniMax M3.
Mercury 3, Inception's diffusion LLM, now in FastMetal chat
Inception's Mercury 3, a diffusion language model that generates over 1,000 tokens per second with a 1M-token context, is now available in FastMetal's chat.
DeepSeek V4.1 Flash 1017 API: now on FastMetal
DeepSeek V4.1 Flash 1017, the October 17 snapshot of V4.1 Flash, is now on FastMetal's API and chat at the same price as V4.1 Flash, with higher coding and agent scores and shorter answers.
GPT-6.1 Sol vs Claude Sonnet 5.5: same price, 2.1x the output
GPT-6.1 Sol and Claude Sonnet 5.5 cost the same per token on FastMetal. On the same 41 prompts, Sonnet 5.5 wrote 2.1x the output tokens and still answered faster. The extra tokens are mostly annotations.
DeepSeek V4.1 Flash vs GLM 5.3 Flash: same answer, 12x the tokens
We ran DeepSeek V4.1 Flash and GLM 5.3 Flash on the same 43 prompts. The answers were about the same length, but DeepSeek used 1.7x the output tokens on average and 12x on one SQL task. All of it was thinking.
Which model do I trust with eight hundred websites?
I own eight hundred dropped domains and about four hours a week. This is the benchmark I built to decide which language model gets to write on all of them, and the two runs it took to stop measuring my own instrument.
Showing 1-10 of 35