Morph

Pure usage-based inference host, re-read 2026-09-11 through a headless browser over CDP (the site sits behind a Vercel bot challenge: WebFetch, curl and r.jina.ai all get 429). Free = 200 requests/month, then pay-per-token ("practically no rate limits"); PAYG top-up starts at $10 with +$5 free on the first top-up, and carries a 5.5% credit purchase fee ($0.80 minimum) that dedicated capacity does not. Hosted-model per-M tokens (input / cache-read / output), all 1M context: Kimi K3 $2.50/$0.29/$14.00; Kimi K3 fast $6.00/$0.60/$22.50; GLM-5.3 744B $1.19/$0.20/$3.74; GLM-5.3-Flash $0.20/$0.04/$0.70; DeepSeek V4 Flash 0731 $0.14/$0.036/$0.40; DeepSeek V4.1 Flash $0.15/$0.01/$0.60 (text and images, reasoning on by default). Specialized: Fast Apply $0.80/$1.20 per M in/out (morph-v3-fast, 262K ctx) or $0.90/$1.90 (morph-v3-large); WarpGrep, now morph-warp-grep-v2, $0.80/100K, 100K ctx (1M on Pro); Compact $0.20/$0.50 per M in/out; Model Router $0.005/request; Reflex $0.001/event realtime or $0.0005/event batch (both halve past 1M events/mo). Dedicated reserves B200 OR B300 by the GPU-hour, from $9.06 per reserved B200-hour behind a 99.9% monthly availability SLA, with a 90-day initial commitment; $8.64 / $8.43 / $8.16 are the 2x / 4x / 8x card rates. The dedicated catalog is seven models and wider than the chat one: it still carries Qwen 3.6 27B, Gemma 4 31B and MiniMax M3 428B, which left chat but not the product. Careful with one figure the vendor contradicts itself on: the dedicated calculator quotes DeepSeek V4 Flash 0731 at $0.14/$0.0028/$0.28, against the $0.14/$0.04/$0.40 of its own price table (2026-09-18). The table is the more credible of the two. 2026-10-02: producto nuevo, System One (morph-systemone-v1) a $0,042 de entrada y $0,01 cacheada por millon, 64K de contexto: 'decisiones tipadas, un estado entra y sale una distribucion calibrada por pregunta'.