Loxi runs on a small model its team trained for knowledge work, costs a fraction of what frontier systems charge, and scored 58.2% on JobBench, level with Claude Fable 5. The company behind it is not a Silicon Valley lab. It is a small team in Istanbul, and their result says more about where AI competition is heading than any funding round this year.
While the largest AI labs compete on parameter counts and billion-dollar compute budgets, a small team in Istanbul has been running a different experiment. They took a model they trained specifically for knowledge work, built a harness around it, and entered it into JobBench, an evaluation built from tasks that professionals across 35 occupations said they would actually hand to an agent.
Loxi scored 58.2%. That puts it level with Claude Fable 5 and ahead of Kimi K3, Qwen 3.8 Max, Claude Opus 4.8, and GPT-5.6 SOL. The model underneath is not a frontier system. The product is free. And it runs at under a tenth of the cost of the largest models on the leaderboard.
For founders watching AI costs climb, that last number deserves attention.
Table of Contents
The leaderboard is not telling the whole story
Most benchmark tables compare models. Loxi’s result suggests that comparison misses something important. Between the model and the user sits a layer called the harness: how a task is planned, how files are read and written, how the agent checks its own work, and which skills it loads for a spreadsheet versus a legal memo. For eight months, that layer has been the entire focus of the Loxi team.
The effect is not unique to them. LangChain changed only its agent runtime and moved from thirtieth to fifth place on Terminal Bench 2.0, raising its score from 52.8% to 66.5% without touching the underlying model. A separate arXiv paper found that holding the base model fixed and redesigning the harness lifted pass@1 from 69.7% to 77.0%.
The Loxi team describes the outcome plainly: same class of model, same tasks, same judge, different harness, different score. Benchmark tables, they argue, credit the model for work the harness is doing.
What JobBench actually tests
JobBench was built from a survey of 1,500 professionals across 35 occupations who were asked which of their own tasks they would delegate to an agent. The tasks that came back are the work people actually want to hand off: messy files, contradicting sources, and deliverables that demand real judgment. Every output is scored against dozens of binary, expert-written criteria, and a correct number reached through faulty reasoning earns nothing.
Loxi completed all 65 tasks on the main split. No time-outs, no refusals, no empty folders. You can see the latest results on their page.
Domain-specific models are quietly winning
The Loxi result fits a pattern that enterprise buyers have started to notice. Gartner projects that more than half of enterprise generative AI deployments will be domain-specific by 2027, up from about 1% in 2024, and that organizations will use small, task-specific models three times as often as general-purpose large language models.
Finance and law offer early proof. Bridgewater Associates worked with Thinking Machines Lab to fine-tune a model on how its own investment managers operate, and the result beat frontier models at one-fourteenth of the cost. Harvey AI fine-tuned the Kimi 2.6 model on a legal benchmark and reported roughly 40% improvement, reaching frontier-level performance at one-eleventh of the cost.
For founders building AI products, the lesson is straightforward. You do not need the largest model. You need the right model, trained on the right work, wrapped in the right system.
A Turkish team, and a country moving in the same direction
Loxi comes from an Istanbul-based team, and its timing overlaps with a formal push in Turkey. The country’s National AI Action Plan for 2026 to 2030, issued by presidential decree in August 2026, sets targets of 1 GW of data center capacity, at least $10 billion in private investment, AI literacy training for 5 million people, and 10,000 advanced specialists by the end of the decade.
The plan explicitly calls for developing and exporting sector-specific language models in health, energy, and smart manufacturing. Industry Minister Mehmet Fatih Kacır has said Turkey intends to be among the leaders of the global technology transition rather than a spectator.
The startup side has moved as well. Turkey reached eight unicorns in 2026, and the number of candidates in the Turcorn 100 program nearly doubled from 23 in March 2025 to 43 in the first half of 2026. Ankara-based Herdr raised $6 million in seed funding for an open-source AI agent runtime for terminals.
The map may be wider than it looks
A July 2026 Bank of America report describes an AI race consolidating around the United States and China. The US retains its lead in private investment, advanced chips, and compute infrastructure, while China leverages manufacturing scale and low energy costs. Outside those two, BofA names South Korea as the strongest candidate, followed by the UAE, with Canada, Germany, Israel, the Netherlands, Singapore, Switzerland, and the UK forming a second tier.
That map leaves little space for anyone else, which is what makes Loxi’s result worth noting. A country’s position in AI is not decided only by capital and compute. Software engineering depth and systems design can put a small team in the same performance band as the largest labs.
What this means for the labor market
PwC’s Global AI Jobs Barometer, published in June 2026, found the labor market splitting along two tracks. Specialized roles, where AI automates routine work and elevates human judgment, are growing twice as fast as democratized roles and carrying 42% faster wage growth. Companies using AI most effectively report 52% employment growth against 36% for the least exposed, and 24% wage growth against 17%.
JobBench measures the work at the center of that shift: routine but difficult, repetitive but demanding close attention, and drawn from what professionals themselves said they would hand off. An agent that performs them reliably does not replace a person. It gives that person more room for the part of the job that requires judgment.
The takeaway for founders
Loxi’s numbers point to something larger than a single benchmark result. Domain-specific models are gaining ground. Harness engineering is becoming a discipline in its own right. Players outside the US and China are beginning to appear on the board.
The next phase of competition may depend less on who trains the largest model and more on who builds the best system around it. For founders and business leaders, that is a shift worth paying attention to, especially if your AI bill is growing faster than your margin.
