Verified
Spotify
Experimenting- Audio streaming
- 7,284 employees (Sep 2026)
- Google Cloud
- Large Taste Model
- Honk agent on Claude
- 99%+ of engineers use AI weekly
Spotify deploys AI mainly in its own product and engineering work. Spotify's Large Taste Model is built on open-source models trained on its listening data, its consumer features include DJ, Prompted Playlist and Talk to Spotify, and Honk, its coding agent, is built on Anthropic's Claude. More than 99% of Spotify's engineers used AI coding tools every week in Jun 2026, with co-CEO Gustav Söderström leading the push. Spotify has reported no business return it credits to AI, citing faster shipping with flat headcount instead.
engineering.atspotify.comTheir approach to AI
- The platform
- Spotify builds its own AI stack on Google Cloud. Its Large Taste Model learns from the 3.4 trillion events and taste signals Spotify's users generate each day (Investor Day, May 2026), Honk is a background coding agent, and Xirp routes engineering work between paid and self-hosted open-source models (Aug 2026).
- The models
- Anthropic's Claude runs Honk and Claude Code, and Spotify engineers also use OpenAI's Codex. Ads AI runs Gemini 2.5 Pro on Vertex AI, and Spotify fine-tunes open-weight models such as Qwen3 for recommendations. Spotify said in May 2026 that it will buy general reasoning on the open market rather than train its own frontier models.
- The reach
- More than 99% of Spotify's engineers used AI coding tools every week (Jun 2026), and AI-powered experiences such as DJ reached about a quarter of Spotify's active users (Aug 2026).
- The catch
- Spotify has published no business return it credits to AI, only throughput and engagement measures. Spotify's own count of AI tool users covers engineers, about half of its staff, not the whole company.
Snapshot
- Employees
- 7,284 full-time
- Sep 30, 2026; 7,287 on average in 2025
- Technology budget
- €1.39B R&D (2025)
- 8% of revenue; cloud and IT costs rose €30M in 2025
- AI/ML headcount
- Not disclosed
- 3,787 research and development staff on average in 2025
- Who owns AI
- Gustav Söderström, co-CEO
- Co-CEO since Jan 1, 2026; earlier Chief Product and Technology Officer
- Platform
- Large Taste Model, Honk, Xirp, Backstage
- Built in-house; Portal and Xirp also sold to other companies (Oct 2026)
- Models
- Claude, Codex, Gemini, open-weight models
- Claude behind Honk; Gemini 2.5 Pro in Ads AI; Qwen3 in recommendations
- Reach
- 99%+ of engineers weekly
- Jun 2026; data assistant used by 2,100+ staff
- Use cases
- Recommendations, DJ, Prompted Playlist, coding agents
- Plus Talk to Spotify, Personal Podcasts, ad tools and Wrapped
Key findings
Spotify's engineering blog said on Jun 3, 2026 that more than 99% of Spotify engineers use AI coding tools every week and that 94% say AI made them more productive. Spotify said pull request frequency was up 76%, and that adoption jumped after Anthropic released Claude Opus 4.5 in late 2025.
Spotify built Honk, a background coding agent that runs Claude through Anthropic's Agent SDK inside Spotify's own tooling. Gustav Söderström said in Feb 2026 that Spotify's best developers had not written a line of code since Dec 2025.
Spotify builds its Large Taste Model by further training open-source models on its own listening data, and Gustav Söderström said in Oct 2026 that early deployments show double-digit growth in discoveries.
Spotify said on Oct 22, 2026 that its AI investments accelerated feature shipping while keeping headcount flat, at 7,284 full-time employees on Sep 30, 2026. Spotify said in Aug 2026 that marketing and AI would add about €200 million of operating expense in 2026.
Spotify has reported no business return it credits to AI: its figures are throughput and engagement measures, and its count of AI tool users covers engineers, not the whole company.
Where Spotify sits
- Stage
- Experimenting
Spotify has reported no business return it credits to AI since May 2025, for any use case or for the company. Spotify cites throughput instead: pull request frequency up 76% (Jun 2026) and merged code changes up from about 8,100 in Aug 2025 to 17,000 in Aug 2026 (Sep 2026). For engagement, Spotify cites double-digit growth in discoveries from its Large Taste Model (Oct 2026). CFO Christian Luiga said on Oct 22, 2026 that AI investments accelerated feature shipping while keeping headcount flat, without giving a figure. Co-CEO Gustav Söderström's remark that revenue per employee is on track to double over 3 years (Aug 2026) is a forecast not credited to AI alone. On rollout, Spotify's engineering blog said on Jun 3, 2026 that more than 99% of its engineers use AI coding tools every week, and Spotify's internal data assistant had more than 2,100 users (Jun 2026). Spotify has not reported a count of AI tool users across the company, and its research and development staff were about half of its 7,287 average employees in 2025. Spotify is at Experimenting. One return it credits to AI in a use case, or its own company-wide count of AI tool users with a named owner, would move Spotify to Piloting.
How the stages are defined - Ownership
- Co-CEO Gustav Söderström, formerly Chief Product and Technology Officer
- Platform
- Hybrid: Large Taste Model, Honk and Xirp built in-house; frontier models bought
- Models
- Anthropic's Claude and OpenAI's Codex for coding; Gemini 2.5 Pro for Ads AI; open-weight models such as Qwen3 fine-tuned in-house
- Hosting
- Google Cloud Platform for the vast majority of computing and data; Gemini on Vertex AI; Claude and OpenAI hosting and GPU locations not disclosed
- Reach
- 99%+ of engineers use AI coding tools weekly (Jun 2026)
How they deploy
Buy general reasoning, train for taste
Spotify said at its May 21, 2026 Investor Day that it does not expect its advantage to come from owning a general reasoning frontier model. Spotify said buying that capability on the open market will remain the most cost-efficient strategy. Spotify instead trains what it calls its Large Taste Model, earlier its Large Personalization Model, on the taste signals only Spotify holds. Gustav Söderström said in Aug 2026 that Spotify builds the model by further training open-source models on its proprietary data.
Flat headcount, more compute per employee
Gustav Söderström said in Apr 2026 that Spotify had not increased its headcount, and had in fact slightly decreased it, while spending more compute per employee, choosing to ship more rather than cut staff. Christian Luiga said in Aug 2026 that Spotify's 2026 investment cycle is marketing and AI, not personnel. Spotify measures the result in pull requests, in definitions of done shipped in its 6-month planning cycles and in feature usage. Söderström said in Oct 2026 that the measures showed significant acceleration in shipped end-user features.
An agent other agents can call
Gustav Söderström said in Feb 2026 that Spotify was building an intelligent, agentic media platform that users can talk to. In Oct 2026, Söderström said Spotify is now an agent reachable in its own app and through other agents. Spotify integrated into ChatGPT in Oct 2025, into Anthropic's Claude in Apr 2026 and into Meta's Muse in Oct 2026. Söderström said playback and the reasoning over user data stay inside Spotify rather than passing to those partners.
AI music on licensed terms
Spotify said in Oct 2025 that it is building a generative AI research lab and product team to develop artist-first AI music products with Sony Music Group, Universal Music Group, Warner Music Group, Merlin and Believe. Spotify said the products would rest on upfront agreements that let artists choose whether to take part. Spotify signed covers and remix licenses with Universal Music Group (May 2026), Merlin (Aug 2026) and Believe (Oct 2026). The Financial Times reported in Oct 2026 that the covers and remix tool could launch in Q1 2027.
Selling the tools it built
Spotify launched Spotify Technology on Oct 8, 2026 as one brand for the developer products it sells to other companies: Portal by Spotify, Xirp, Confidence and Spotify for Backstage. Gustav Söderström said in Oct 2026 that selling the tools spreads the cost of Spotify's agent development work over more developers than its own few thousand. Söderström called the business option value and gave no financial guidance.
- Deployment model
- Spotify's AI work runs through the product and technology organization that Gustav Söderström led as Chief Product and Technology Officer before becoming co-CEO on Jan 1, 2026. Tyson Singer, SVP of Technology and Platform, oversees the developer platforms behind Spotify Technology, including Portal by Spotify, Xirp and Confidence. Chief Architect Niklas Gustavsson presented Spotify's AI coding rollout at Anthropic's Code with Claude conference in 2026. Spotify's product teams build consumer features on the Large Taste Model and on outside models, and the advertising organization runs Ads AI and AI ad creation tools. Spotify's internal data assistant serves employees across advertising, podcasts, music, audiobooks and finance.
- Model providers
- Spotify uses Anthropic's Claude for coding: Honk runs Claude through the Agent SDK (Jun 2026), and Gustav Söderström named Claude Code and OpenAI's Codex among the tools that accelerate Spotify's development (Apr 2026). Ads AI, Spotify's media planning system, runs on Google's Agent Development Kit with Gemini 2.5 Pro on Vertex AI (Feb 2026). For recommendations, Spotify fine-tunes open-weight models, including a Qwen3 variant with 1.7 billion parameters served for generative recommendation (Oct 2026). Spotify builds its Large Taste Model by further training open-source models on Spotify data (Aug 2026). Xirp lets Spotify engineers switch models mid-task and route each job to the model with the best performance for the price, including open-source models Spotify hosts itself (Aug 2026). DJ's commentary has been generated by OpenAI since its Feb 2023 launch, with a voice built by Sonantic, which Spotify acquired in 2022. Spotify's 2025 Wrapped used a frontier model to produce reference outputs and a smaller fine-tuned model to write 1.4 billion personal reports (Mar 2026).
- Data and infrastructure
- Spotify relies on Google Cloud Platform for the vast majority of its computing and primary data storage, including users' personal data and licensed audio. Spotify says moving off Google Cloud would take significant time and expense (2025 Form 20-F). Honk sessions run in Kubernetes pods across Spotify's cloud environment, with logs uploaded to Google Cloud (Nov 2025 and Jun 2026), and Ads AI runs on Vertex AI (Feb 2026). Spotify serves generative recommendation models on GPUs with vLLM, SGLang and NVIDIA's TensorRT LLM (Oct 2026). Spotify has not said where those GPUs sit or where the Claude and OpenAI models its engineers use are hosted. Spotify's data platform, a separate measure from the taste model's event count, processes 1.4 trillion data points a day across more than 70,000 datasets (Spotify Engineering, Jun 2026). Gustav Söderström said in Apr 2026 that Spotify uses third-party AI services but not with its core data.
Platforms and internal tools
Large Taste ModelDescribed May 2026
The Large Taste Model is Spotify's in-house personalization model, built by further training open-source models on Spotify's proprietary data and fed by the 3.4 trillion events and taste signals its users generate each day (May 2026). Spotify said in Aug 2026 that a new AutoPlay recommendation system powered by the model had raised active days, AutoPlay minutes and saves in its first 2 months. In Oct 2026, Spotify said early deployments show double-digit gains in content discovery.
- Models
- Open-source base models with continued pre-training
HonkInternal since 2025
Honk is Spotify's background coding agent. Honk runs Claude through Anthropic's Agent SDK inside Spotify's own tooling and works with the Fleet Management framework to make code changes across thousands of components. Engineers send Honk requests over Slack and get back a pull request or a testable app build. Spotify said in Nov 2025 that Honk's agents had produced more than 1,500 merged pull requests, and in Jun 2026 that a Java migration across its backend services took 3 days.
- Models
- Anthropic's Claude
Xirp2026
Xirp, which Spotify called Chirp in Jun 2026, is the engine under Honk. Spotify engineers use Xirp instead of calling Claude Code or Codex directly, so they can switch models mid-task and route each job to the model with the best performance for the price. Xirp also shares context across models, developers and teams and shows inference spend down to each developer (Aug 2026). Spotify also sells Xirp to other companies under the Spotify Technology brand (Oct 2026).
- Models
- Paid models and self-hosted open-source models
Backstage and Portal by Spotify
Backstage is Spotify's internal developer portal, open-sourced in 2020, and Portal by Spotify is its commercial version, sold since 2024. Spotify exposes Backstage's catalog through the Model Context Protocol (MCP) and command-line tools so that Claude can look up component owners and documentation. Spotify also uses Backstage standards and automated code checks to correct agents that pick the wrong technology (Jun 2026).
Data assistantAug 2025
Spotify's data assistant answers questions in plain English by writing and running SQL against Spotify's data warehouse, using clusters of datasets, vetted question and SQL pairs and documents curated by data experts. Spotify said in Jun 2026 that more than 2,100 employees had used it in more than 13,000 conversations, through Slack, an MCP server and a web interface.
- Users
- 2,100+ employees (Jun 2026)
Ads AIDescribed Feb 2026
Ads AI is Spotify's media planning system made of several AI agents, built on Google's Agent Development Kit and Gemini 2.5 Pro on Vertex AI. Ads AI turns an advertiser's request in plain language into a media plan drawn from historical campaign performance. Spotify said a plan that took 15 to 30 minutes by hand takes seconds.
- Models
- Gemini 2.5 Pro on Vertex AI
Vendors and partners
Google Cloud
Cloud for the vast majority of Spotify's computing and data, and Vertex AI with Gemini for Ads AI · Spotify's 2025 Form 20-F records a service agreement with Google for Google Cloud Platform with minimum spend commitments, and says a move to another cloud provider would take significant time and expense.
Anthropic2025
Claude models behind Honk and Claude Code; Spotify connector inside Claude · Spotify said in Jun 2026 that Claude Code adoption spiked after the Claude Opus 4.5 release in late 2025. Spotify launched an integration in Claude on Apr 23, 2026 and said it does not share its catalog with Anthropic for model training.
OpenAI2023
Commentary model behind DJ; Codex for engineers; Spotify app inside ChatGPT · Spotify launched DJ with OpenAI-generated commentary in Feb 2023 and announced its ChatGPT integration on Oct 6, 2025.
NVIDIA2026
Joint engineering on TensorRT LLM for Spotify's generative recommendation inference · Spotify and NVIDIA engineers added beam search, scoring and embedding features to TensorRT LLM, which Spotify says cut beam search inference cost by up to 54% on a Qwen3 1.7B model (Oct 2026).
Meta2026
Spotify integration in Meta's Muse agent and Meta Ray-Ban glasses · Spotify said on Oct 22, 2026 that it had launched an integration with Muse, and Gustav Söderström said Spotify was the only media service in Muse at launch.
Universal Music Group, Merlin and Believe2026
Licenses for AI covers and remixes on Spotify · Universal Music Group signed in May 2026, Merlin in Aug 2026 and Believe, with TuneCore, in Oct 2026. The Financial Times reported that Spotify was still negotiating with Sony Music Group and Warner Music Group (Oct 2026).
Leadership
AI, data and platform leaders
Executives
AI org map
Co-CEOs run one leadership group; a central technology and platform organization builds the AI tooling, and product teams build AI features on the Large Taste Model
Spotify has been run since Jan 1, 2026 by co-CEOs Gustav Söderström and Alex Norström, who report to Executive Chairman Daniel Ek and serve as joint chief operating decision makers. Söderström said in Feb 2026 that the co-CEOs run a single direct reporting group, which meets weekly for 3 hours with Spotify's VP and SVP layer to plan and unblock work across the company. Spotify has not named a successor to Söderström's former Chief Product and Technology Officer role, and its AI tooling sits in the technology and platform organization led by Tyson Singer.
Spotify does not disclose how many people work on AI or machine learning. Spotify averaged 3,787 research and development employees in 2025, out of 7,287 full-time staff, and calls its developers a few thousand (Oct 2026). Spotify's AI work sits in product and technology teams, including the personalization teams behind the Large Taste Model and the technology and platform organization under Tyson Singer, which builds Honk, Xirp and Backstage. Spotify began building a generative AI research lab and product team in Oct 2025. Spotify has not named the leaders of the personalization teams or the lab.
Reporting lines
Executive Chairman
Co-CEOs
Leadership group
Weekly 3-hour meeting of Spotify's VP and SVP layer, run by the co-CEOs (Feb 2026)
Technology and Platform
Tyson Singer, SVP
Honk, Xirp, Backstage, Portal and Confidence; reporting line inferred
Finance
Christian Luiga, CFO
Reporting line inferred
Architecture and engineering
Niklas Gustavsson, Chief Architect
Fleet Management and engineering standards; reporting line not published
Personalization and the Large Taste Model
Leader not named; placement under the co-CEOs' product organization inferred
Generative AI research lab
Announced Oct 2025 for artist-first AI music products; leader and reporting line not published
Spotify's 2025 Form 20-F lists the co-CEOs, the Executive Chairman, the CFO, the Chief Public Affairs Officer and the Chief Human Resources Officer as senior management, and names no chief technology, AI or data officer. Spotify has published no headcount for its AI or personalization teams, so the split of its AI workforce is not shown.
Named people
Use cases
Spotify runs AI in production in its listener product: recommendations and AutoPlay on its Large Taste Model, podcast recommendations from a large language model, DJ, Prompted Playlist, Taste Profile, Personal Podcasts and the 2025 Wrapped. Spotify also runs AI in its integrations with ChatGPT, Claude and Meta's Muse, in ad planning and ad creation tools, and in engineering through Honk, Claude Code, Codex and an internal data assistant. Talk to Spotify, Studio by Spotify Labs and licensed AI remixes are in beta or announced. Spotify has published no business-level return credited to AI for any of these use cases or for the company. Spotify reports throughput such as 76% higher pull request frequency (Jun 2026) and engagement such as double-digit growth in discoveries (Oct 2026).
| Use case | Area | Status | ROI |
|---|---|---|---|
| Recommendations and AutoPlay on the Large Taste Model Spotify deployed a new AutoPlay recommendation system powered by its Large Taste Model in 2026. Gustav Söderström said the system raised active days, AutoPlay minutes and track saves in its first 2 months (Aug 2026) and that the model is on track to add billions of listening hours a year (Oct 2026). | Listener experience | In production | |
| Podcast recommendations from a language model Spotify introduced generative recommendation for podcasts in Jan 2026. A large language model generates semantic IDs with beam search and a soft token built from each user's embedding. Spotify served the model first on vLLM and SGLang and is moving it to NVIDIA's TensorRT LLM. | Listener experience | In production | |
| DJ DJ combines personalized music picks with AI-generated commentary in a synthetic voice. Spotify said in Aug 2026 that DJ had served nearly 100 million Premium users since its 2023 launch and added French, German, Italian and Brazilian Portuguese. | Listener experience | In production | |
| Prompted Playlist and Taste Profile Prompted Playlist lets Premium users describe a playlist in their own words, drawing on their full listening history. Prompted Playlist reached about 14 million of the first 100 million users offered it (Aug 2026). Taste Profile lets users see and edit how Spotify models their taste, and launched to U.S. Premium subscribers in Q3 2026. | Listener experience | In production | |
| Talk to Spotify Talk to Spotify lets Premium users type or speak to the app to shape what plays, ask about songs and artists and explore their listening history. Talk to Spotify has been in beta since Jul 2026 for users 18 and older in the U.S., Ireland and Sweden. Gustav Söderström said in Oct 2026 that tens of millions of people speak to Spotify in English. | Listener experience | Pilot | |
| Personal Podcasts and Studio by Spotify Labs Personal Podcasts lets Premium users generate and schedule AI-made audio episodes, with a monthly allowance of credits and the option to buy more. Studio by Spotify Labs, a desktop agent that can use a browser, calendar and inbox to build private podcasts and briefings, began a research preview rollout on Jul 20, 2026. | Listener experience | In production | |
| Spotify Wrapped 2025 Archive Spotify used a language model to write 1.4 billion reports on up to 5 standout listening days per user for 2025 Wrapped. Spotify fine-tuned a smaller production model on a frontier model's outputs, and checked the reports with a second model acting as judge and with human review. | Listener experience | In production | |
| Spotify inside ChatGPT, Claude and Muse Spotify runs as an agent other assistants can call, integrated into ChatGPT (Oct 2025), Anthropic's Claude (Apr 2026) and Meta's Muse (Oct 2026). Users of those assistants can get recommendations, build playlists and start playback that runs through Spotify. | Listener experience | In production | |
| Licensed AI covers and remixes Spotify is building a tool for fans to make covers and remixes of participating artists' songs with consent, credit and compensation, licensed by Universal Music Group, Merlin and Believe. The Financial Times reported a possible Q1 2027 launch at $4 to $5 a month on top of a subscription. | Music and creators | Announced | |
| Ads AI media planning Ads AI splits media planning into specialized agents on Google's Agent Development Kit and Gemini 2.5 Pro. Ads AI turns an advertiser's goals, written in plain language, into a plan for budget, audiences and formats based on historical campaign performance. | Advertising | In production | |
| AI audio ad creation and agent access to the Ads API Alex Norström said in Aug 2026 that about 7,000 of Spotify's 33,000 active advertisers use its AI audio asset creation tool. Norström said Spotify had launched plugins and Model Context Protocol (MCP) servers so advertisers can create campaigns from Claude, ChatGPT and Gemini, and Spotify published its Ads API plugin for Claude Code in May 2026. | Advertising | In production | |
| Honk and Fleet Management code migrations Honk makes complex code changes across thousands of components inside Fleet Management, which has merged more than 2.5 million automated maintenance pull requests over the years. Engineers also call Honk from Slack to fix bugs or add features and get back a testable build. | Engineering and data | In production | |
| AI coding tools for engineers Spotify engineers use Claude Code, Codex and Honk through Xirp. Spotify reported weekly use by more than 99% of engineers and a 76% rise in pull request frequency, and said total merged changes more than doubled year over year to about 17,000 in Aug 2026. Spotify's incident reviews found no material direct contribution from AI-authored code. | Engineering and data | In production | |
| Internal data assistant Spotify's data assistant writes and runs SQL from plain English questions against expert-curated clusters of datasets. More than 2,100 employees have used the assistant since Aug 2025, and more than a quarter of them had never written SQL. | Engineering and data | In production |
Adoption
Spotify spread AI first through its product: The Echo Nest acquisition in 2014 brought machine learning personalization, DJ added generated commentary in 2023, and by Aug 2026 AI-powered experiences reached about a quarter of Spotify's active users. Inside the company, Spotify issued AI rules and training to all employees in Jun 2025. Spotify's engineering blog reported in Jun 2026 that more than 99% of engineers use AI coding tools every week, after Claude Code adoption spiked with Claude Opus 4.5 in late 2025. Gustav Söderström said in Aug 2026 that working through Honk from Slack had become ubiquitous among engineers. Outside engineering, the only usage count Spotify has published is its data assistant's more than 2,100 users (Jun 2026), and Spotify has not reported how many of its roughly 7,300 employees use AI tools.
By the numbers
Rollout, first pilot to today
2014
Spotify acquired The Echo Nest, a machine learning company whose work Daniel Ek said gave Spotify its personalization.
Feb 2023
Spotify launched DJ, with commentary generated by OpenAI and a voice built by Sonantic, the AI voice company Spotify acquired in 2022.
Feb 2025
Spotify began testing AI agents inside its Fleet Management system, which had automated about half of Spotify's pull requests since mid-2024.
Apr 2025
Spotify expanded AI Playlist, which builds playlists from written prompts, to more than 40 markets.
Jun 2025
Anna Lundström and Gustav Söderström issued AI rules to the whole company and introduced AI training for employees.
Oct 2025
Spotify announced its ChatGPT integration and partnerships with the major labels, Merlin and Believe on artist-first AI music products, and began building a generative AI research lab.
Nov 2025
Spotify said Honk's agents had produced more than 1,500 pull requests merged into production.
Dec 2025
Spotify introduced Prompted Playlist in beta, and Claude Code adoption among its engineers spiked after the Claude Opus 4.5 release.
Jan 2026
Gustav Söderström and Alex Norström became co-CEOs, and Spotify introduced podcast recommendations from a large language model.
Feb 2026
Gustav Söderström described Honk on the Q4 2025 call and said Spotify's best developers had not written a line of code since Dec 2025.
May 2026
Spotify presented its Large Taste Model at its Investor Day, signed an AI covers and remix license with Universal Music Group and announced Studio by Spotify Labs.
Jun 2026
Spotify reported that more than 99% of its engineers use AI coding tools every week and that its data assistant had more than 2,100 internal users.
Aug 2026
Spotify described Xirp, which routes engineering work across paid and self-hosted models, and said AI-powered experiences reached about a quarter of its active users.
Oct 2026
Spotify launched Spotify Technology to sell its developer tools, integrated with Meta's Muse and said its AI investments had accelerated feature shipping with flat headcount.
Economics & controls
Spend and ROI
Technology budget
Spotify spent €1.393 billion on research and development in 2025, 8% of revenue, down 6% mainly because of lower social charges on share-based pay. Spotify's information technology costs rose €30 million in 2025 on greater use of cloud computing (2025 Form 20-F). Spotify's purchase obligations with minimum spend commitments, including its Google Cloud Platform agreement and some content and marketing commitments, totaled €1.575 billion at Dec 31, 2025. Spotify does not break out AI spending.
What AI itself costs
Christian Luiga said in Aug 2026 that marketing and AI-related investments would add about €200 million of operating expense in 2026, without splitting marketing from AI. Spotify's Q3 2026 deck attributed the 19% rise in operating expenses, excluding currency and social charges, mainly to marketing alongside cloud and AI spend. Gustav Söderström said in Apr 2026 that the compute behind Spotify's AI work covers coding tools, training its own large models and the inference behind new features. Spotify said that expense growth would moderate from Q4 2026 into 2027. Spotify and NVIDIA said in Oct 2026 that TensorRT LLM work cut beam search inference cost by up to 54% on a Qwen3 1.7B recommendation model.
Reported ROI
Spotify has reported no business-level return credited to AI. Spotify cites throughput instead: pull request frequency up 76% (Jun 2026), merged code changes up from about 8,100 in Aug 2025 to 17,000 in Aug 2026 (Sep 2026) and definitions of done roughly doubling (Apr 2026). For engagement, Spotify cites double-digit growth in discoveries from the Large Taste Model (Oct 2026). Gustav Söderström said in Oct 2026 that AI-built features raise engagement, retention and, for audiobooks, average revenue per user, without a figure for AI's share.
Headcount and output
Spotify had 7,323 full-time employees at Dec 31, 2025 and 7,284 at Sep 30, 2026, and Christian Luiga said in Aug 2026 that Spotify would stay flat on employees in 2026. Gustav Söderström said in Aug 2026 that revenue per employee is on track to double over 3 years with no increase in headcount, a forecast that Spotify does not credit to AI alone. Spotify's average headcount fell from 9,123 in 2023 to 7,287 in 2025 after reorganizations in 2023.
Pricing AI features
Gustav Söderström said in Aug 2026 that inference, the cost of running AI models, adds a variable cost per user, so Spotify meters its AI features. Personal Podcasts come with a monthly credit allowance in Premium plus paid top-ups, and every credit purchase carries a defined amount of inference. Spotify has not disclosed revenue from AI features or from Spotify Technology's commercial tools.
Governance and risk
Rules and training for employees
Spotify's Chief Human Resources Officer Anna Lundström and Gustav Söderström issued a set of AI rules to the whole organization and introduced training from prompt engineering to advanced courses (Jun 2025). Xirp shows Spotify its inference spend down to each developer and lets it route work to models it chooses, which Söderström described in Aug 2026 as a cost control.
Quality controls on AI-written code
Spotify reviews every major incident monthly and, as its AI use grew, began asking whether AI-authored code contributed. Tyson Singer wrote in Sep 2026 that the reviews had found no material direct contribution, but that the volume of change outran some verification controls. Spotify is strengthening code review, testing, rollout, monitoring and rollback in response. Backstage standards and automated code checks give Claude immediate feedback when it uses the wrong technology in Spotify's codebase (Jun 2026).
AI-generated music
Spotify removed more than 75 million spammy tracks in the 12 months to Sep 2025. Spotify also announced a policy against unauthorized AI voice clones and impersonation, a music spam filter and AI disclosures in song credits through the DDEX standard. Spotify's artist-first principles commit it to upfront agreements with rightsholders, artist choice over participation, compensation and credit (Oct 2025).
Data boundaries
Gustav Söderström said in Apr 2026 that Spotify uses third-party AI services but not with its core data. Söderström said in Oct 2026 that outside agents such as Muse ask Spotify for results while Spotify does the reasoning over its user data. Spotify said in Apr 2026 that it does not share its music, podcast or video catalog with Anthropic for training Claude.
Board oversight and disclosed risks
Spotify's audit committee reviews enterprise risk management, including data protection and cybersecurity. Spotify's 2025 Form 20-F warns that its use of AI, including generative AI in features such as DJ and AI Playlist, may not benefit the business. The filing also warns that laws such as the EU Artificial Intelligence Act may raise compliance costs and that competitors may build cheaper or better AI products.
Recent signals
Oct 30, 2026
The Financial Times reported that Spotify plans to launch its AI covers and remixes tool in Q1 2027 and is weighing a price of $4 to $5 a month.
Oct 22, 2026
Christian Luiga said Spotify's AI investments accelerated feature shipping while keeping headcount flat, and Gustav Söderström said early deployments of the Large Taste Model show double-digit growth in discoveries.
Oct 22, 2026
Spotify launched an integration with Meta's Muse agent and rolled Taste Profile out to U.S. Premium subscribers.
Oct 20, 2026
Believe and its TuneCore platform signed an agreement covering AI covers and remixes on Spotify.
Oct 16, 2026
Spotify and NVIDIA described TensorRT LLM work that cut beam search inference cost by up to 54% for Spotify's generative recommendation models.
Oct 8, 2026
Spotify launched Spotify Technology, one brand for the developer tools it sells, including Portal by Spotify, Xirp and Confidence.
Sep 16, 2026
Tyson Singer wrote that merged code changes more than doubled year over year and that incident reviews found no material direct contribution from AI-authored code.
Aug 4, 2026
Gustav Söderström described Xirp and the new AutoPlay system on the Large Taste Model, and Spotify announced an AI covers and remix license with Merlin.
Jul 2026
Spotify began a beta of Talk to Spotify, a conversational assistant for Premium users in the U.S., Ireland and Sweden.
Jun 10, 2026
Spotify described its internal data assistant, used by more than 2,100 employees since Aug 2025.
Jun 3, 2026
Spotify said more than 99% of its engineers use AI coding tools every week and that pull request frequency was up 76%.
May 21, 2026
Spotify's Investor Day set out its Large Taste Model strategy, a covers and remix license with Universal Music Group and Studio by Spotify Labs.
Apr 28, 2026
Gustav Söderström said Spotify had not increased its headcount, and had in fact slightly decreased it, while spending more compute per employee on AI.
Apr 23, 2026
Spotify launched an integration inside Anthropic's Claude for recommendations, playlists and playback.
Feb 10, 2026
Gustav Söderström described Honk and said Spotify's best developers had not written a line of code since Dec 2025.
Quick answers
- How does Spotify deploy AI?
- Spotify deploys AI mainly in its own product and engineering work. Spotify's Large Taste Model is built on open-source models trained on its listening data, its consumer features include DJ, Prompted Playlist and Talk to Spotify, and Honk, its coding agent, is built on Anthropic's Claude. More than 99% of Spotify's engineers used AI coding tools every week in Jun 2026, with co-CEO Gustav Söderström leading the push. Spotify has reported no business return it credits to AI, citing faster shipping with flat headcount instead.
- What does Spotify use AI for?
- Spotify runs AI in production in its listener product: recommendations and AutoPlay on its Large Taste Model, podcast recommendations from a large language model, DJ, Prompted Playlist, Taste Profile, Personal Podcasts and the 2025 Wrapped. Spotify also runs AI in its integrations with ChatGPT, Claude and Meta's Muse, in ad planning and ad creation tools, and in engineering through Honk, Claude Code, Codex and an internal data assistant. Talk to Spotify, Studio by Spotify Labs and licensed AI remixes are in beta or announced. Spotify has published no business-level return credited to AI for any of these use cases or for the company. Spotify reports throughput such as 76% higher pull request frequency (Jun 2026) and engagement such as double-digit growth in discoveries (Oct 2026).
- Who leads AI at Spotify?
- AI at Spotify is led by Tyson Singer (SVP, Technology and Platform); Niklas Gustavsson (Chief Architect and VP of Engineering). Ownership of AI on the org chart: Co-CEO Gustav Söderström, formerly Chief Product and Technology Officer.
- What AI platform does Spotify use?
- Spotify builds its own AI stack on Google Cloud. Its Large Taste Model learns from the 3.4 trillion events and taste signals Spotify's users generate each day (Investor Day, May 2026), Honk is a background coding agent, and Xirp routes engineering work between paid and self-hosted open-source models (Aug 2026).
- Which AI models does Spotify use?
- Anthropic's Claude runs Honk and Claude Code, and Spotify engineers also use OpenAI's Codex. Ads AI runs Gemini 2.5 Pro on Vertex AI, and Spotify fine-tunes open-weight models such as Qwen3 for recommendations. Spotify said in May 2026 that it will buy general reasoning on the open market rather than train its own frontier models.
- How many Spotify employees use AI?
- More than 99% of Spotify's engineers used AI coding tools every week (Jun 2026), and AI-powered experiences such as DJ reached about a quarter of Spotify's active users (Aug 2026).
- What return has Spotify reported from AI?
- Spotify has reported no business return it credits to AI since May 2025, for any use case or for the company. Spotify cites throughput instead: pull request frequency up 76% (Jun 2026) and merged code changes up from about 8,100 in Aug 2025 to 17,000 in Aug 2026 (Sep 2026). For engagement, Spotify cites double-digit growth in discoveries from its Large Taste Model (Oct 2026). CFO Christian Luiga said on Oct 22, 2026 that AI investments accelerated feature shipping while keeping headcount flat, without giving a figure. Co-CEO Gustav Söderström's remark that revenue per employee is on track to double over 3 years (Aug 2026) is a forecast not credited to AI alone. On rollout, Spotify's engineering blog said on Jun 3, 2026 that more than 99% of its engineers use AI coding tools every week, and Spotify's internal data assistant had more than 2,100 users (Jun 2026). Spotify has not reported a count of AI tool users across the company, and its research and development staff were about half of its 7,287 average employees in 2025. Spotify is at Experimenting. One return it credits to AI in a use case, or its own company-wide count of AI tool users with a named owner, would move Spotify to Piloting.
Open questions
- How many Spotify employees outside engineering use AI tools, and will Spotify report a company-wide count with a named owner?
- What return in cost, headcount or revenue does Spotify credit to Honk, Claude Code and its other AI coding tools, beyond pull request and shipping counts?
- How much of the roughly €200 million of extra 2026 operating expense goes to AI compute and models rather than marketing?
- Where do the Claude and OpenAI models Spotify's engineers use run, and where are the GPUs that train the Large Taste Model and serve its recommendations?
- Which open-source base models does the Large Taste Model start from, and how large is it?
- Will Spotify put a figure on the retention and revenue effect of the Large Taste Model, DJ and Prompted Playlist?
- How much revenue do Personal Podcasts credits, Spotify Technology's tools and the planned AI remix add-on bring in?
- Who leads Spotify's personalization teams and its generative AI research lab, and who holds the product and technology duties Gustav Söderström held before Jan 2026?
- Will Sony Music Group and Warner Music Group license Spotify's AI covers and remixes tool?
Sources
Spotify
- Co-CEOs' remarks at Spotify Investor Day (May 21, 2026)
- Studio by Spotify Labs launch (May 21, 2026)
- Spotify integration with Claude (Apr 23, 2026)
- Spotify Engineering: Introducing Spotify Technology (Oct 8, 2026)
- Spotify Engineering: Generative recommendation inference with NVIDIA TensorRT LLM (Oct 16, 2026)
- Spotify Engineering: AI changed how Spotify builds (Sep 16, 2026)
- Spotify Engineering: The context layer behind Spotify's data assistant (Jun 10, 2026)
- Spotify Engineering: Coding is no longer the constraint (Jun 3, 2026)
- Spotify Engineering: Natural language interface to the Ads API with Claude Code plugins (May 1, 2026)
- Spotify Engineering: Background coding agents for dataset migrations, Honk part 4 (Apr 22, 2026)
- Spotify Engineering: Inside the Archive, the tech behind 2025 Wrapped (Mar 12, 2026)
- Spotify Engineering: Our multi-agent architecture for smarter advertising (Feb 19, 2026)
- Spotify Engineering: 1,500+ PRs later, Honk part 1 (Nov 6, 2025)
Filings, decks and earnings calls
- Q3 2026 shareholder deck (Oct 22, 2026)
- Q2 2026 shareholder deck (Aug 4, 2026)
- Q1 2026 shareholder deck (Apr 28, 2026)
- Q4 2025 shareholder deck (Feb 10, 2026)
- Spotify 2025 Form 20-F (filed Feb 10, 2026)
- Spotify announces co-CEOs and Executive Chairman (Sep 30, 2025)
- Q3 2026 earnings call transcript, Equibles (Oct 22, 2026)
- Q2 2026 earnings call transcript, Equibles (Aug 4, 2026)
- Q1 2026 earnings call transcript, Equibles (Apr 28, 2026)
- Q4 2025 earnings call transcript, Equibles (Feb 10, 2026)
- Q3 2025 earnings call transcript, Equibles (Nov 4, 2025)
Partners and conferences
News and analysis
- Music Business Worldwide: Spotify plans Q1 2027 launch for AI covers and remixes tool (Oct 30, 2026)
- Pulse 2.0: Spotify introduces Talk to Spotify (Jul 2026)
- RouteNote: What is Talk to Spotify (2026)
- RouteNote: Claude becomes Spotify's latest AI partner for music discovery (Apr 2026)
- MBI Deep Dives: Notes from Spotify's 2026 Investor Day (May 2026)
- TechCrunch: Spotify says its best developers haven't written a line of code since December (Feb 12, 2026)
- Music Business Worldwide: Spotify to develop artist-first AI music products (Oct 16, 2025)
- Music Business Worldwide: Spotify has deleted 75m+ spammy tracks (Sep 25, 2025)
- Fortune: Spotify's new CHRO on AI training (Jun 12, 2025)
- Engadget: Spotify's new AI DJ (Feb 22, 2023)
Data confidence
- Confirmed
- Stated in a primary or reputable public source.
- Estimated
- Derived or approximated from public reporting.
- Unconfirmed
- Reported, but not yet corroborated.
Spotify profile · Enterprise AI diffusion report by Imagine · facts checked against their sources on Nov 9, 2026
