Underclassic / model board
Full context in today's issue ->
Blended pricing is (3*input + 1*output) / 4 -- a reading mix, not a lab cost. The multiplier is dearest blended / cheapest blended among rows with a public list price. Blank $ cells mean open weights or hoster menus with no single official API price. These are vendor list rates, not electricity or chip COGS.
Public list prices and open-weight rows. Verified 2026-10-17.
Rather run models at home? Buyer's guide: home AI computers
Model board (36) (USD per 1M tokens; verified 2026-10-17; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)
| # | Model | Org | CC | Weights | Tier | In | Out | Blend |
| 1 | Claude Fable 5.1 | Anthropic | US | closed undisclosed |
flagship | $10.00 | $50.00 | $20.00 |
| Base input/output MTok list. Cache hits 0.025x input. | ||||||||
| 2 | Claude Fable 5 | Anthropic | US | closed undisclosed |
flagship | $10.00 | $50.00 | $20.00 |
| Base input/output MTok list. | ||||||||
| 3 | Claude Opus 5 | Anthropic | US | closed undisclosed |
flagship | $5.00 | $25.00 | $10.00 |
| 4 | Claude Opus 5.5 | Anthropic | US | closed undisclosed |
flagship | $4.00 | $20.00 | $8.00 |
| 5 | Claude Sonnet 5.5 | Anthropic | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| 6 | Claude Sonnet 5 | Anthropic | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| $2/$10 is now the standard list price; the planned 2026-09-01 rise to $3/$15 was cancelled. | ||||||||
| 7 | Claude Haiku 4.5 | Anthropic | US | closed undisclosed |
fast | $1.00 | $5.00 | $2.00 |
| 8 | GPT-6 Astra | OpenAI | US | closed undisclosed |
flagship | $10.00 | $50.00 | $20.00 |
| Short-context standard tier. Long context $20/$75. | ||||||||
| 9 | GPT-6.1 Sol | OpenAI | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| Short-context standard tier. Long context $4/$15. | ||||||||
| 10 | GPT-6 Luna | OpenAI | US | closed undisclosed |
commodity | $0.10 | $0.50 | $0.20 |
| Short-context standard tier. Long context $0.20/$0.75. | ||||||||
| 11 | GPT-5.6 Sol | OpenAI | US | closed undisclosed |
flagship | $4.00 | $20.00 | $8.00 |
| Standard tier list. Vendor page says Sol promotional pricing runs at least through 2026-11-21. | ||||||||
| 12 | GPT-5.5 | OpenAI | US | closed undisclosed |
flagship | $5.00 | $30.00 | $11.25 |
| Short-context (<272K) list. | ||||||||
| 13 | GPT-5.6 Terra | OpenAI | US | closed undisclosed |
workhorse | $2.00 | $12.00 | $4.50 |
| 14 | GPT-5.4 mini | OpenAI | US | closed undisclosed |
fast | $0.75 | $4.50 | $1.69 |
| 15 | GPT-5.6 Luna | OpenAI | US | closed undisclosed |
commodity | $0.20 | $1.20 | $0.45 |
| 16 | Grok 4.7 | xAI | US | closed undisclosed |
flagship | $2.00 | $6.00 | $3.00 |
| Prompt <200k list tier. | ||||||||
| 17 | Grok 4.6 | xAI | US | closed undisclosed |
flagship | $2.00 | $6.00 | $3.00 |
| Prompt <200k list tier. | ||||||||
| 18 | Grok 4.3 | xAI | US | closed undisclosed |
workhorse | $1.25 | $2.50 | $1.56 |
| Prompt <200k list tier. | ||||||||
| 19 | DeepSeek V4.1 Flash | DeepSeek | CN | open ~284B total / ~13B active (MoE) |
commodity | $0.30 | $1.20 | $0.53 |
| V4 Flash is retired; model name deepseek-flash now serves DeepSeek-V4.1-Flash. Peak cache-miss list; off-peak is half. Params/weights shown are V4 Flash's. | ||||||||
| 20 | DeepSeek V4 Pro | DeepSeek | CN | open ~1.6T total / ~49B active (MoE) |
workhorse | $1.32 | $3.96 | $1.98 |
| Peak cache-miss list. Off-peak half. | ||||||||
| 21 | GLM-5.3-Flash | Z.ai | CN | open see model card |
commodity | $0.15 | $0.50 | $0.24 |
| Z.ai official list price (cached input $0.03). Open weights at huggingface.co/zai-org/GLM-5.3-Flash. Self-reported by Mia's AI Lab (not verified by us): 36-92 tokens/s decode on two DGX Sparks depending on workload (github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks). | ||||||||
| 22 | Mistral Large 4 (Preview) | Mistral | FR | pending ~1.05T total / 52B active (MoE) |
flagship | $1.36 | $4.18 | $2.06 |
| Public-preview list price. Launch sale is 50% off for now ($0.68 in / $2.09 out). Mistral says open weights land by the end of October 2026. | ||||||||
| 23 | Gemini 3.1 Pro (Preview) | US | closed undisclosed |
workhorse | $2.00 | $12.00 | $4.50 | |
| Prompts <=200k tokens. Over 200k: $4 in / $18 out. | ||||||||
| 24 | Gemini 3.8 Flash | US | closed undisclosed |
fast | $0.75 | $3.75 | $1.50 | |
| List through 2026-12-31; vendor page says $1.50/$7.50 from 2027-01-01. | ||||||||
| 25 | Gemini 4 Argon | US | closed undisclosed |
flagship | $2.00 | $10.00 | $4.00 | |
| Paid tier only. Introductory list price through 2026-11-30; Google lists $4 in / $20 out from 2026-12-01. | ||||||||
| 26 | Qwen3.5-397B-A17B | Alibaba | CN | open 397B total / 17B active |
open-flagship | -- | -- | -- |
| Apache-class open weights. API $ varies by hoster - not a single list price. | ||||||||
| 27 | Llama 4 Maverick | Meta | US | open ~400B total / ~17B active |
open-flagship | -- | -- | -- |
| Community license. Hoster list rates vary; EU multimodal caveats apply. | ||||||||
| 28 | Llama 4 Scout | Meta | US | open ~109B total / ~17B active |
open-longctx | -- | -- | -- |
| Long-context open path. Hoster $ varies. | ||||||||
| 29 | Muse Spark Open | Meta | US | open 816B total / 42B active (MoE) |
open-flagship | -- | -- | -- |
| No official API price: Meta serves it only via partners (Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together). Muse Model License (limited commercial terms). Meta says it runs on two M5 Ultra Mac Studios with MLX. 1M context. | ||||||||
| 30 | Nemotron 3.5 Super VL | NVIDIA | US | open 120B |
open-flagship | -- | -- | -- |
| No official API price: open weights on Hugging Face, also OpenRouter and NVIDIA NIM. Multimodal (vision-language). | ||||||||
| 31 | Nemotron 3.5 Bolt 74B-A9B | NVIDIA | US | open 74B total / 9B active |
open-local | -- | -- | -- |
| No official API price. Native 2-bit quantization fits in 24 GB; built for RTX/DGX Spark. llama.cpp, Ollama, LM Studio, vLLM. A 4 GB Bolt E4B is due Oct 23. | ||||||||
| 32 | LFM2.5-24B-A2B | Liquid AI | US | open 24B total / 2B active (MoE) |
open-local | -- | -- | -- |
| No official API price. Reasoning-only agent model, 131K context. Liquid reports 184 tok/s on an M5 Max and 114 tok/s on a Ryzen AI Max+ 395 (vendor figures). | ||||||||
| 33 | Granite Switch 4.1 (8B / 30B) | IBM | US | open 8B and 30B sizes |
open-local | -- | -- | -- |
| No official API price. Base model plus built-in Decision and Guardian adapters and a speech-to-text model, served from one endpoint (vLLM plug-in, llama.cpp). | ||||||||
| 34 | Mistral Medium 3.5 | Mistral | FR | closed undisclosed |
workhorse | $1.50 | $7.50 | $3.00 |
| EU frontier multimodal. List price from Mistral's model card. | ||||||||
| 35 | Mistral Small 4 | Mistral | FR | open ~119B total / ~6B active (MoE) |
open-efficient | $0.15 | $0.60 | $0.26 |
| Hybrid instruct/reason/code. Apache 2.0 weights; API list from Mistral's model card. | ||||||||
| 36 | Gemma 4 31B | US | open 31B |
local-desk | -- | -- | -- | |
| Desk-side class open weights. Self-host cost is silicon+power, not a metered list. | ||||||||
These are public list prices from vendor pages where linked. Blank cells mean open weights or hoster menus with no single official API price. This table is not a claim about how much electricity or chips cost the lab. Useful signal: how far cheap API tiers sit from flagship list rates, and which models ship weights you can run yourself. Rows marked announced are not on sale yet: their promised prices sit in the footnote and stay out of the blend and gap math.
Subscribe
$ echo you@email.com | underclassic --subscribe
One email per issue. No tracking pixels, no filler.
home | rss | page 2 | page 3 | archive | models | buyer's guide | subscribe | data | board | mcp | llms.txt