# PublicAI launches index combining 15 LLM leaderboards > The company combines 3,621 scores covering 808 models into one Overall ranking. It also published a JSON API and MCP server that require no API key. - Publication: Go Big News - By Go Big News Staff (Staff News) - Category: AI - Published: 2026-09-09T04:03:52.069+00:00 - Canonical: https://www.gobignews.com/ai/publicai-launches-index-combining-15-llm-leaderboards --- On September 8, 2026, [PublicAI](https://publicai.io) announced an index that combines 15 public LLM leaderboards covering 808 models into one Overall score. The company said in a post on X that it does not run any of the benchmarks itself. The Index places 3,621 scores from 14 publishers on a 0-100 scale across 51 domains. A model needs scores from two independent publishers to enter the Overall ranking, while missing scores are not filled in and thin evidence is pulled toward 50. Sources include LMArena, Artificial Analysis, Terminal-Bench, ARC-AGI-2, LiveBench, LiveCodeBench, Aider, BFCL, OSWorld, MMMU-Pro, Kagi, SimpleBench and Vals AI. The [product page](https://publicai.io/model-index) also lists GDPval-AA and AA-Briefcase as Artificial Analysis evaluations. Vals AI contributes 22 evaluations covering areas including legal, finance, tax, medical and cybersecurity. PublicAI z-scores each figure across the models listed by its source board before mapping it to the common scale. Published error bars reduce a figure’s weight. Launch posts and blogs contribute to domain columns at one-tenth of a board’s weight but do not enter the Overall index. The printed Overall weights are 25% for LMArena Text, 25% for Artificial Analysis, 20% for LiveBench, 15% for Terminal-Bench and 15% for ARC-AGI-2. On September 8, the page listed 193 ranked models and 615 provisional models. Models without an Overall score can appear as estimates, such as ~55 with an asterisk, but are not ranked. Claude Fable 5.1 led the table at 70.9, followed by GPT-6 Astra at 70.2 and Claude Fable 5 at 68.8. The Index is available through a JSON API and an MCP server at https://publicai.io/model-index/mcp. Its four tools are `rank_models`, `get_model`, `whats_new` and `describe_index`, with no API key required. PublicAI said the ranking refreshes daily and is also distributed through a Monday Index Weekly email, an RSS feed and a change query for agents. OpenRouter and Hugging Face provide information about model access and open-weight sizes, rather than scores. PublicAI also excluded SWE-bench Verified, saying the newest of its 181 entries was dated February 17, 2026, and none of the models in the current table appeared on it. The ranking uses the highest published reasoning-effort tier and is a snapshot of each publisher’s table on the date read. PublicAI’s core business collects and annotates text, audio, video and mapping data and provides model evaluation. Its LinkedIn profile lists the company as founded in 2023, headquartered in San Francisco and employing 11–50 people. PublicAI announced $10 million in funding in June 2025, comprising an $8 million Series A after an initial $2 million. In its [launch blog](https://publicai.io/blog/introducing-the-publicai-index), dated September 7, PublicAI said it plans to expand toward 20 recognized boards, including τ²-bench, SEAL, Epoch AI and HealthBench as they publish tables. --- ## Companies in this story - Publicai: https://www.gobignews.com/company/publicaiio --- ## X profiles in this story - @publicaidata: https://www.gobignews.com/profile/publicaidata