The news of September 17thMozilla (15 SEPTEMBER) ISSUED THE SECOND EDITION OF THE OPEN SOURCE AI STATUS REPORT, WHICH REPORTED THATChinaBest open weight model of the company andUSAThe technology company's front-line closed-source modelThe capacity gap has narrowed to 4.4 months。
COMPARED WITH THE FIRST EDITION OF THE OPEN SOURCE AI STATUS REPORT, RELEASED IN JULY OF THIS YEAR, THE SECOND EDITION OF THE REPORT FURTHER LOWERS MACROJUDGMENT TO SPECIFIC MODELS, SPECIFIC MISSION DURATION AND SINGLE TASK COST。
| dimension (math.) | July first edition | September new edition | Main changes |
|---|---|---|---|
| Capacity gap | The performance gap between top open models and proprietary models like ChatGPT and Claude is 3% | The gap between the US front closed source model and the best open weight model in China is 4.4 months | Change from a single, integrated percentage to a closer “task time” indicator for actual work。 |
| Cost presentation | GPT-4 GRADE REASONING COSTS DROPPED TO ONE IN 50 IN 3 YEARS | Kimi K3 cost 30% of Fable 5; GM 5.2 single task cost is about one fifth lower | The new edition gives a comparison of the model with the direct cost of the mission。 |
| Use data | 79% DEVELOPER USES OPEN MODEL 51% DEPLOYED PRODUCTION; CLOSED SOURCE MODEL PRODUCTION DEPLOYMENT RATE 63% | August 2026 OpenRouter 8 of 10 models by token provide open weights | The focus has shifted from research by developers to the latest platform use ranking。 |
| Income data | OPEN MODEL COVERS ABOUT ONE THIRD OF THE ACTUAL AI USAGE, BUT ONLY 4% INCOME | REAFFIRMING THAT OPEN MODEL INCOME IS 4%, CLOSED SOURCE 96%, AND INDICATING THAT DATA COVER MAY TO SEPTEMBER 2025 | The new version complements the time frame of the income conclusion and the share of closed-source income。 |
The second edition focused on the Kimi K3 model on the dark side of the month, which scored only three points lower on the Artificial Analysis smart index than the Anthropic Fable 5, at a cost of about 301 TP3T。
IN TERMS OF SCOPE OF CAPACITY, METR MEASURES THE “TIME HORIZON” OF A MODEL OF A RELIABLE TIME FRAME FOR A MISSION, AND THE RELIABLE SUCCESS RATE IS 50%. CURRENTLY, THE OPTIMAL CLOSED-SOURCE MODEL IS 1.7 TIMES LONGER THAN THE BEST OPEN-SOURCE MODEL. THE OPEN-SOURCE MODEL HANDLES 7 HOURS OF TASK AND THE CLOSED-SOURCE MODEL HANDLES ABOUT 12 HOURS OF TASK。
Mozilla Chief Technical Officer Raffi Krikorian characterized the difference as depending on the mission scenario: the closed-source model premium was mainly on professional tasks that required between 8 and 12 hours, while open-source models were able to handle daily work at very low prices。
Krikorian states that the open model can handle 12 hours of mission after 4 months and the closed source model can handle approximately 20 hours of mission. At present, 8 to 12 hours of work are usually still areas where closed-source models can be completed and open models are difficult to handle。
This shift has begun to reshape the behaviour of enterprises. DoorDash is now using Kimi to handle day-to-day tasks while retaining Fable for more complex tasks. In August 2026, eight of the top 10 models on the routed platform OpenRouter ranked by token usage were open source models。