What happened
Loxi, a free AI product built by an Istanbul‑based team, achieved a 58.2% score on the JobBench knowledge‑work , placing it level with Claude Fable 5 and ahead of GPT‑5.6 SOL, Kimi K3, Qwen 3.8 Max and Claude Opus 4.8. The system completed all 65 benchmark tasks without time‑outs, refusals or empty outputs, and the team reports operating costs under one‑tenth of the largest models on the leaderboard. Loxi’s architecture relies on a small, task‑specific model combined with a purpose‑engineered “harness” that plans tasks, manages files, checks work and loads appropriate skills. The team spent eight months developing this harness layer and ran thousands of controlled experiments to tune the model‑harness fit. Independent research cited in the release shows similar gains when only the harness is improved, underscoring the importance of system design beyond model size.
The Istanbul team released Loxi, a free AI product aimed at professional knowledge‑work tasks such as spreadsheet manipulation, legal memo drafting, financial analysis and project planning. Unlike the frontier models it was compared against, Loxi does not rely on a large, general‑purpose model; instead it uses a small model trained specifically for these tasks.
On the JobBench —a 65‑task suite measuring real‑world professional deliverables—Loxi achieved a 58.2% score, matching Claude Fable 5 and surpassing several leading models, including GPT‑5.6 SOL and Kimi K3. The system completed every task without time‑outs, refusals or empty output folders.
The team attributes the performance to a “harness” layer that orchestrates task planning, file handling, self‑verification and skill selection. Over eight months, they conducted thousands of controlled experiments, adjusting one variable at a time to optimize the fit between the small model and the harness. Independent research cited in the release shows similar performance gains when only the harness is improved, reinforcing the importance of system design.
Loxi operates at under one‑tenth the cost of the largest models on the JobBench leaderboard, according to the team’s calculations. The product remains in a free beta, with the latest experimental build continuously updated based on user feedback.
Why it matters
The Loxi result demonstrates that specialized, cost‑efficient AI systems can rival the performance of much larger frontier models on professional knowledge‑work tasks. This challenges the prevailing assumption that only massive, general‑purpose models can deliver high‑quality deliverables in domains such as spreadsheets, legal memos, financial analysis and project planning. By showing that a small, purpose‑trained model plus a well‑designed harness can achieve comparable scores at a fraction of the cost, Loxi offers a viable alternative for enterprises facing rising AI expenses. The finding aligns with broader industry trends toward domain‑specific AI deployments, which Gartner predicts will comprise over half of enterprise generative‑AI use by 2027. It also highlights a shift in competitive advantage from raw model scale to system engineering, potentially reshaping how startups and established firms allocate R&D resources. For founders and business leaders, Loxi suggests a path to high‑performing professional AI without the heavy cloud‑ bills associated with the largest models, potentially lowering barriers to adoption in cost‑sensitive sectors. The announcement also ties into Turkey’s National AI Action Plan (2026‑2030), which emphasizes sector‑specific language models and substantial private investment. Loxi’s success may encourage further regional innovation and signal that AI competitiveness is expanding beyond the United States and China, offering a new dimension to the global AI race.
The result challenges the dominant narrative that only massive, general‑purpose models can handle complex professional tasks, suggesting that specialized, smaller models paired with sophisticated system engineering can achieve comparable outcomes at dramatically lower cost.
Cost efficiency is a critical concern for enterprises as AI expenses rise. Loxi’s sub‑10% cost claim offers a concrete example of how organizations might achieve high‑quality AI assistance without the financial burden of running the largest models.
The achievement aligns with Gartner’s projection that more than half of enterprise generative‑AI deployments will be domain‑specific by 2027, indicating a broader industry shift toward tailored AI solutions.
Loxi’s success may stimulate further investment in region‑specific AI initiatives, especially given Turkey’s National AI Action Plan, which aims to develop sector‑specific language models and attract $10 billion in private AI investment by 2030.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
Future updates to Loxi’s model and harness, adoption rates among enterprises, and any pricing changes if the product moves beyond its free beta. Watch for independent verification of cost claims and performance on subsequent JobBench releases. Also monitor whether other startups replicate Loxi’s approach, and how larger AI labs respond—potentially by releasing more domain‑specific offerings or improving their own system‑level engineering. Finally, observe policy developments in Turkey’s AI strategy that could support or accelerate similar specialized AI ventures.
Whether Loxi remains free or introduces a paid tier, and how any pricing changes affect its cost advantage.
Independent verification of the reported cost savings and scores on future JobBench releases.
Adoption metrics among businesses, particularly in finance and legal sectors that are early adopters of domain‑specific AI.
Responses from larger AI labs—whether they will accelerate domain‑specific model releases or enhance their own system‑level engineering.
Policy developments in Turkey’s AI strategy that could provide additional funding or regulatory support for similar specialized AI projects.