跪拜 Guibai
← All articles
AI Programming · AIGC · Frontend

AI Models 1.1.0 Folds LLM Pricing, API Addresses, and SOTA Timelines into One Directory

By 王若风 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Choosing and wiring up an LLM still costs developers hours of cross-referencing vendor docs and price pages; this directory turns that into a set of filters and a Base URL copy, which directly shortens the model-selection loop for any team that ships against multiple providers.

Summary

Swapping an LLM into an application typically requires checking context lengths and per-token costs, then digging through vendor docs for model IDs, Base URLs, and which API key works with which plan. AI Models 1.1.0 collapses that two-step research into a single directory. The November 15 release holds 1,806 model records and 47 provider entries, built on Pi data plus official vendor sources. Combined search narrows by model name, ID, provider, and Chinese/English tags, then stacks filters for context range, flagship status, open source, and domestic origin. A Model PK view pairs candidates side-by-side and estimates the cost of a real request: in the example, DeepSeek V4 Pro costs $0.17 for a 100k input / 10k output request, while V4.1 Flash costs $0.042 for the same volume. The API Address window organizes Base URLs by protocol (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Google Generative AI) and by plan, so a key for Volcano Engine Ark's pay-as-you-go API maps to /api/v3, while Coding Plan and Agent Plan use separate paths. The new SOTA timeline lists original flagships in reverse release order, with vendor, model ID, capability tags, and a link to the official source. Preferences for filters, currency, favorites, language, and theme persist in the browser, and shareable links preserve the selected provider set.

Takeaways
— 1,806 model records and 47 provider entries are in the November 15 release, based on Pi data plus official vendor sources.
— Search covers model name, model ID, provider, and Chinese/English tags, and combining terms (e.g., 'DeepSeek 开源') narrows to records matching all words.
— A Model PK comparison estimates per-request cost: DeepSeek V4 Pro at $1.30 input / $4.00 output per million tokens vs. V4.1 Flash at $0.30 / $1.20 for the same 100k input / 10k output request, giving $0.17 vs. $0.042.
— The API Address window lists Base URLs per protocol and plan; for Volcano Engine Ark's Chat Completions, the pay-as-you-go API uses /api/v3, Coding Plan /api/coding/v3, and Agent Plan /api/plan/v3, all on the same domain.
— Four protocols are supported for Base URL lookup: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google Generative AI.
— The SOTA timeline shows only original flagship models, sorted by release date, with vendor, model ID, capability tags, and an official source link.
— Free models require input and output unit prices, plus cache and image prices, to be explicitly zero.
— Browser preferences persist for filters, sorting, currency, favorites, language, and theme, and a shared link with provider parameters reproduces the same candidate set.
Conclusions

The directory's decision to group a vendor's regions and plans together (e.g., Ark's three API paths) surfaces a common operational trap: the same model name can hide different Base URLs and billing terms depending on the key and plan.

The cost PK demonstration with SKU-like unit prices shows that choosing between a flagship and a lighter model is not a single static comparison; a real request volume often flips the economic case.

A manually verified SOTA list backed by source links is a different kind of trust signal from the automatic benchmark leaderboards that dominate Western coverage; it trades freshness for credibility.

Concepts & terms
Base URL
The root endpoint a client calls to reach an LLM API. Different plans or access methods under the same vendor often share a domain but use distinct path prefixes, so copying the right one is what makes a request succeed.
SOTA timeline
A chronological list of original flagship models from each vendor, sorted by release date. Here it carries a manually verified 'SOTA' marker and a link to the official source, rather than relying on benchmark scores.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗