Changelog

Dated changes to our numbers and pages, newest first. Each model page also shows its own checked date.

2026-11-04

New: Cheapest way to run each model — what $100 of rental buys you per model, context capped at each model's native max, no invented purchase prices. 20 pages indexed; the rest stay noindex until reviewed.
Correction (gpt-oss prose): Quick answers and FAQs no longer call the 65.4 GB / 13.8 GB file "BF16" weights; it is the "BF16" GGUF with MoE experts in native MXFP4.
Disclosure: B&H Photo declined our affiliate application; we have no relationship with B&H.
Correction (gpt-oss): gpt-oss-20b and gpt-oss-120b were labelled "16-bit (FP16/BF16)" / "BF16 (original)". They ship natively in MXFP4 (MoE experts MXFP4, attention/router BF16), so that row is now labelled MXFP4 (native) on model pages, the calculator, Ask, Scan and the data files. Sizes are unchanged: 13.8 GB file (14.0 GB at 8K context) for gpt-oss-20b, 65.4 GB file (65.7 GB at 8K) for gpt-oss-120b.
Prices: Runpod and Lambda pricing pages re-read; Vast.ai verified on-demand single-GPU offers re-read (05:25 PT). GPU prices.
New page: Cheapest GPU rental by VRAM (7 VRAM tiers).
New models: Qwen-Image, Wan 2.2, FLUX.1 schnell, SDXL, SD 3.5, LTX-Video (Hugging Face file listings checked 2026-11-04).
Updated: Mistral Large 4 and gpt-oss-120b pages got short direct answers using the existing sourced numbers.
Gaming: 1440p, 4K and 8 GB pages rechecked against their sources.
Fixes: removed plain Newegg links (no relationship); URLs with double slashes now redirect.

Prices: Novita AI per-token prices verified from Novita's model list (for example GPT OSS 120B $0.05 in / $0.25 out per 1M tokens).
New: "Run it in the cloud" table on model pages; local AI rig guides (16, 24, 32 GB); embeddable calculator at /embed/.
Disclosure: clarified that Vast.ai, Runpod and Novita AI links are referral links; pending programs are named as pending.

New: VRAM calculator tool page, Mistral Large 4, GPU buying guides (12–32 GB), DMCA and updated terms, contact page.
Redesign of the site.

New model: Qwen3.8-35B-A3B. Model data refreshed across the catalog; methodology published at Key facts.

Found something wrong? Report an error.