Observed Signal · Nov 6, 2026 · Technical Release · Source: Linas Newsletter · Impact: 4/5 · Sentiment: Positive
Revolut slashes AI coding costs by 90% with open-source models
Revolut, Europe's most valuable private company, cut its AI coding bill by 90% by replacing metered Anthropic and OpenAI APIs with self-hosted, open-weight GLM models from Z.ai, running on 56 Nvidia B200 GPUs through Together AI. The move, disclosed by David Abitbol (Head of AI Operations) at AI Engineer New York, resulted in a blended cost of $0.30 per million tokens and a 70% reduction in lead time from feature pick-up to merge-ready code for its 2,000 engineers. The company tuned its GPU fleet instead of expanding it, reserving frontier models for tasks requiring peak capability. Revolut, which reported over $6B in 2025 revenue and is known for its own AI initiatives like the PRAGMA foundation model, has nearly two years of data across its engineering org, making this a notable case study for AI cost optimization in enterprise settings.
This case study from a major fintech demonstrates a significant cost-saving strategy in AI infrastructure, potentially influencing how enterprises and AI tool providers approach model deployment and pricing. It highlights the shift towards open-weight models and self-hosting for cost efficiency, which is highly relevant to the broader AI and AdTech ecosystem.
Track Revolut Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your Polaris7 workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Revolut reduced AI coding costs by 90% by switching from metered Anthropic/OpenAI APIs to self-hosted open-weight GLM models.
- The open-source AI stack runs on 56 Nvidia B200 GPUs through Together AI, at a blended cost of $0.30 per million tokens.
- The switch led to a 70% reduction in lead time for code delivery for Revolut's 2,000 engineers.
- Revolut reported over $6B in revenue for 2025, up 46%, from over 80 million customers.
- Revolut built its own foundation model for banking, PRAGMA, with NVIDIA, and runs an in-house AI lab, Revolut Research.
Connected Companies & Entities
6 Entities mapped“Revolut cut its AI coding bill by 90% by moving 2,000 engineers to self-hosted open-weight models.”
“Revolut moved off metered Anthropic and OpenAI APIs.”
“Revolut moved off metered Anthropic and OpenAI APIs.”
“Revolut uses Z.ai's open-weight GLM model, self-hosted on Nvidia GPUs.”
“The models run on 56 Nvidia B200 GPUs, and Revolut built PRAGMA with Nvidia.”
“The GPUs are provided through Together AI.”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Revolut Launches PRAGMA; Fintechs Converge on AI
Revolut launched Revolut Research to develop PRAGMA, a proprietary encoder-style foundation-model family (≈10 million–1 billion parameters) pre-trained on multi-source behavioral data from roughly 80 million customers across 100+ countries — tens of millions of user records and tens of billions of events. Revolut reports material production uplifts: ~2.3× better credit-default detection and 65% more fraud cases caught. The announcement is presented as part of an industry-wide convergence (Nubank, Visa, Mastercard, Stripe, Plaid and NVIDIA have comparable efforts; NVIDIA published a blueprint and completed a reported $7B “Poolside” deal). The newsletter highlights engineering, latency, compute-economics, proprietary-data moats, and regulatory constraints (notably the EU AI Act and explainability) that limit deployment in high-risk credit use cases today.
Revolut Publishes PRAGMA Financial Foundation Model
Revolut published PRAGMA (arXiv:2604.08649), a family of encoder-style Transformer foundation models pre-trained on banking event sequences from roughly 25 million users across 111 countries (≈40 billion events, 207 billion tokens). PRAGMA uses a single shared backbone to produce user-level embeddings transferred across tasks including credit scoring, fraud detection, lifetime value prediction and communication engagement. Revolut reports large production gains on several tasks (e.g., +130.2% Credit Scoring PR-AUC, +64.7% Fraud Recall, +79.4% Communication Engagement PR-AUC) while noting a 47% degradation in AML performance. The announcement is positioned as part of a recent wave of financial-data foundation models from Stripe, Mastercard and Visa and raises questions about data-scale moats, regulatory constraints, and commercial opportunities for fintechs and startups.
Freelancer Cuts AI Costs 62% Using Context Windows
A developer describes how they reduced monthly AI API spending by 62% through careful choice of models based on context window needs, token pricing, caching, streaming, and fallbacks. The author shares per‑million‑token pricing observed via a multi‑model aggregator called Global API (pricing for DeepSeek V4 Flash/Pro, Qwen3‑32B, GLM‑4 Plus, GPT‑4o), a reusable Python client that routes calls through Global API, and practical habits (aggressive caching, streaming, model-task matching, quality monitoring, graceful fallbacks). The post includes example billing math, informal benchmark metrics, and a reported monthly token distribution that keeps total AI infrastructure spend under ~$80/month versus $400+ if using an expensive flagship model for all tasks.
Receive all future signals for Revolut & competitors in your briefing
Never miss material M&A deals, leadership changes, account wins, or product rollouts. Polaris7 monitors 20,000+ primary sources 24/7, delivering verified telemetry directly into your executive briefing.