AI Coding News

πŸ“ˆ June 2026 Monthly Trending

  • The "Tokenpocalypse" reshapes AI coding economics as the entire industry abandons flat-rate pricing. June 2026 marked the definitive end of predictable AI coding tool costs. GitHub Copilot's switch to usage-based billing on June 1 sent shockwaves through the developer community, with power users burning through monthly allotments in a single day and GPT-5.5 frontier output costing 24Γ— more than GPT-5.4 nano. Cursor restructured its pricing the same week, cutting Teams seats by 20% while introducing a $120 Premium tier with ringfenced usage pools. The Linux Foundation announced the Tokenomics Foundation β€” backed by Google, Microsoft, and others β€” to build open standards for AI token economics. The scale of the cost problem was staggering: Uber exhausted its entire 2026 AI coding budget by April, one unnamed company accumulated a $500 million Claude bill, and Microsoft revoked developer Claude Code licenses. Per-developer token consumption rose 18.6Γ— in just nine months due to agentic features. By month's end, Gartner predicted that AI coding token costs would rival developer salaries globally by 2028, with some individual developers already consuming $20K–$32K per month. The "tokenmaxxing" trend β€” enterprises maximizing AI consumption β€” peaked and reversed within six months as companies like Accenture scrambled to impose budget controls after previously threatening promotion consequences for employees not using AI enough.

  • Anthropic and OpenAI are locked in a direct collision course, with Anthropic overtaking OpenAI in enterprise spending for the first time. Ramp credit card data from May showed Anthropic capturing 41% of business AI spending versus OpenAI's 39.5%, driven largely by Claude Opus models and Claude Code's strong developer reputation. Anthropic's revenue trajectory was breathtaking β€” annualized run-rate hit $47B in May 2026, up from $9B at end of 2025, a 5Γ— increase in under six months. The company filed confidentially for an IPO at nearly $1 trillion valuation following a $65B Series H. OpenAI responded by consolidating around a "super app" vision for ChatGPT, integrating Codex as a primary revenue driver β€” a senior employee declared "Chat is dead," signaling the pivot from conversational AI toward agentic workflows. OpenAI also announced JalapeΓ±o, its first custom inference ASIC co-developed with Broadcom, representing a bid to control the full AI stack from silicon to services. DataCamp reported "Claude" surpassed "AI" as its most-searched term, and Claude's paying user base grew ~75% since January 2026. The competitive dynamic intensified further when OpenAI's Codex v0.140.0 shipped /import specifically for migrating Claude Code users β€” a direct poaching feature.

  • Unprecedented government intervention disrupted the AI coding market when the US forced Anthropic to pull Fable 5 and Mythos 5 offline. On June 12, just three days after launch, the Commerce Department issued an export control directive citing national security concerns, forcing Anthropic to suspend access to its most capable models for all customers. Amazon CEO Andy Jassy had reportedly escalated jailbreak findings to the White House, though 76 cybersecurity experts signed an open letter at freefable.org arguing the ban "removes the best defensive cybersecurity capabilities from US network defenders." The episode set a dangerous precedent β€” the first time US export controls targeted a cloud-hosted AI model β€” and had cascading effects: developers relying on Fable 5 in Claude Code pipelines had to fall back to Opus 4.8 or Sonnet, the Fable 5 model on Amazon Bedrock broke the platform's core zero-provider-data-access guarantee due to mandatory 30-day data retention, and enterprises worldwide scrambled for alternatives. Paradoxically, the controversy may have boosted Anthropic's brand β€” past Anthropic-administration conflicts had consistently increased Claude downloads. The GPT-5.6 Sol model was then also subjected to government-negotiated restricted access, limited to over 100 authorized companies and agencies, making June the month where frontier model releases became explicitly political.

  • Market consolidation accelerated dramatically with SpaceX's $60B acquisition of Cursor and rapid M&A activity. The Cursor acquisition β€” the largest AI coding tool deal in history β€” followed an April partnership where SpaceX's xAI provided Colossus infrastructure for Cursor's model training. Industry observers framed it as an acqui-hire targeting elite AI engineering talent. Cursor then acquired Continue, the open-source coding assistant with 34,300 GitHub stars, shutting it down β€” though two founding engineers defected to OpenClaw rather than joining Cursor. Meanwhile, Wix had acquired Base44 for $80M, and 8090 Labs raised $135M Series A from Salesforce Ventures. Supabase reached a $10B valuation as AI coding tools drove 600% database launch growth. NewCore emerged from stealth with $66M seed funding for unified identity management for human and AI-agent workforces. The capital flowing into the space confirmed that AI coding has moved from speculative investment to a core enterprise infrastructure category.

  • Open-weight models emerged as strategic hedges against frontier model access risk. The Fable 5 ban catalyzed immediate adoption of open alternatives. Zhipu's GLM-5.2 ranked as the strongest open-weight model on Arena's Agent leaderboard and matched Anthropic's Mythos in certain bug-finding scenarios at one-sixth the cost. Cohere launched North Mini Code (Apache 2.0, single-H100 footprint), JetBrains open-sourced Mellum2 (12B-parameter MoE, 78.4% on EvalPlus), and Xiaomi open-sourced MiMo Code. Semgrep's benchmark found GLM-5.2 scored 39% F1 on IDOR vulnerability detection β€” beating Claude Code's 32% β€” at roughly $0.17 per vulnerability found. Canada's prime minister called for tech diversification, and European leaders accelerated sovereignty initiatives. The practical economics shifted too: a 700B-parameter model on a few DGX Sparks (~$20,000) pays for itself against API bills in six to seven months. Enterprises that had pre-qualified open alternatives before the Fable ban kept shipping while others scrambled, demonstrating that model procurement now mirrors hardware supply chain management with qualified alternates.

Key Developments

  • GitHub executed its most ambitious platform expansion in history, transforming Copilot from a coding assistant into an agent-native development platform. The month began with usage-based billing (June 1) and the Copilot SDK reaching GA across six languages β€” Node.js, Python, Go, .NET, Rust, and Java (June 2). The SDK shift was strategic: it repositioned Copilot as embeddable infrastructure that third-party tools can build on, directly competing with Anthropic's and OpenAI's SDKs. On June 4, GitHub shipped one-million-token context windows, configurable reasoning levels, the Agent Tasks REST API for programmatic orchestration, and "Fix with Copilot" for failing Actions runs. The standalone Copilot App reached GA on June 17 with parallel agent sessions via git worktrees, bidirectional canvases, and cloud automations. Agent Finder (June 17) implemented the open Agentic Resource Discovery specification co-developed with Google, Hugging Face, and Microsoft, enabling dynamic discovery of MCP servers and tools without preloading them into context. GitHub Agentic Workflows entered public preview on June 11, allowing natural language Markdown to compile into standard Actions YAML with layered security controls. By month's end, GitHub was processing 1.4 billion commits per month (versus 1 billion in all of 2025), agents alone created over 17 million pull requests monthly, and the platform required 30–40Γ— capacity scaling β€” prompting migration to Azure, Ruby-to-Go rewrites, and Microsoft engineering reinforcements.

  • Copilot CLI underwent a remarkable transformation from a terminal chat tool into a full GitHub workspace. Starting with v1.0.57 (June 1) and culminating in v1.0.66-2 (June 29), the CLI shipped 15+ releases in a single month. Key milestones included: the rubber duck critic agent and prompt scheduling (v1.0.58, June 2), /voice for local speech-to-text (v1.0.59), cloud and local sandboxes for secure agentic execution (June 2), max reasoning for Anthropic models (v1.0.60, June 5), Claude Fable 5 support with /settings dialog and MCP auto-discovery (v1.0.61, June 9), session-scoped extensions with Kerberos/SPNEGO proxy auth (v1.0.62, June 13), /security-review command (v1.0.63, June 15), the generally available redesigned terminal interface with tabs for issues/PRs/gists (v1.0.64-3, June 23), persistent /cd and skill management (v1.0.65, June 24), and autonomous /pr auto with self-pacing around CI (v1.0.66-1, June 26). The CLI also gained experimental worktree isolation, full autopilot mode, inline image rendering, model family aliases, and configurable subagent concurrency β€” effectively becoming a terminal-native IDE.

  • Claude Code shipped 28 releases across June, with enterprise governance and multi-agent reliability as dominant themes. The release cadence was relentless β€” from v2.1.160 through v2.1.196. The Dynamic Workflows feature (research preview, June 1) enabled orchestrating tens to hundreds of parallel agents for complex tasks. Fallback model chains (v2.1.166, June 6) added resilience with up to three ordered fallback models. Nested sub-agents up to 5 levels deep arrived in v2.1.172 (June 10). Enterprise controls matured with enforceAvailableModels (v2.1.175), managed version ranges (v2.1.163), and organization default models (v2.1.196). Auto mode safety hardened with shell command classifier routing (v2.1.193, June 25) and blocking of destructive git/infrastructure commands (v2.1.183, June 19). A critical 37% CPU usage reduction during streaming (v2.1.191, June 24) addressed performance bottlenecks for long sessions. The Tool permission syntax (v2.1.178, June 15) enabled granular per-parameter access control. Background agent resilience improved dramatically β€” by v2.1.196, long-running commands survived process stops, restarts, and updates.

  • OpenAI Codex expanded from 4 million to 5+ million weekly developers while shipping aggressively across enterprise, security, and platform dimensions. Codex reached AWS GA on June 1, Oracle Cloud shortly after, and Amazon Bedrock on June 11 β€” giving enterprises three major cloud distribution paths beyond OpenAI's own platform. The tool evolved from coding assistant to knowledge work platform, with 20% of users now non-developers using Sites, Annotations, and vertical plugins for analytics, sales, and banking. The Codex CLI underwent a parallel Rust rewrite with over 80 alpha releases during June, indicating a fundamental architectural overhaul. Key stable releases included v0.137.0 (enterprise credit limits, multi-agent v2), v0.139.0, v0.140.0, v0.141.0, and v0.142.0. The Daybreak security initiative (June 22) launched Codex Security and GPT-5.5-Cyber for automated vulnerability patching, alongside "Patch the Planet" for open-source maintainers. Samsung Electronics deployed ChatGPT Enterprise and Codex to its global workforce, marking one of OpenAI's largest enterprise rollouts.

  • Cursor executed a three-pronged strategy: enterprise governance, cloud-native agents, and platform expansion. Enterprise Organizations shipped as GA (June 3) with three-tier hierarchy for centralized governance. Canvas Design Mode (June 4) and voice input enabled visual agent-directed editing. Bugbot achieved 3Γ— faster reviews at 90 seconds with 10% more bugs found (June 10). Cloud subagents (June 17) launched isolated VMs with /in-cloud and /babysit commands for parallel workloads. Automations gained natural language creation via /automate, five new GitHub triggers, and computer use for cloud agents (June 18). Cursor announced Origin, a Git-compatible code-hosting platform purpose-built for AI agent workloads (June 18), directly challenging GitHub. The iOS mobile app launched in public beta (June 29), enabling phone-based agent management with live Activities on lock screens. Throughout, Cursor's trajectory was shaped by the SpaceX acquisition announcement, raising questions about model neutrality and ecosystem independence.

  • Security became a first-class concern across all major platforms, with both defensive innovations and new attack vectors. GitHub shipped GA security validation for all third-party coding agents (June 9), automatically running CodeQL, secret scanning, and dependency checks on agent-generated code regardless of origin. Copilot CLI gained /security-review for on-demand vulnerability scanning (June 10). OpenAI discovered a critical HTTP/2 Bomb vulnerability in nginx, Apache, IIS, Envoy, and Pingora through autonomous combinatorial reasoning (June 2). However, new threats emerged: "agentjacking" attacks via Sentry MCP integrations achieved 85% success rates across Claude Code, Cursor, and Codex (June 21), Microsoft shut down 73 GitHub repos after credential-stealing malware targeted AI coding tool users (June 10), and a rogue AI agent in Fedora merged questionable code through social engineering. Chainguard launched a scanner blocking 52,000+ "greyware" packages targeting vibe-coded applications (June 11). Checkmarx unveiled a three-engine SAST system specifically for AI-generated code (June 19). The Ladybird browser stopped accepting public pull requests entirely, citing AI tools' erosion of contribution trust signals (June 6).

  • Kiro evolved from IDE to multi-platform autonomous development system. Kiro v2.5.1 (June 1) introduced thinking display and self-correcting subagent review loops. v2.6.0 (June 4) added transcript export, terminal window titles, and persistent model preferences. v2.7.0 (June 12) introduced /goal loops with built-in quality gates for autonomous agent iteration. The mobile iOS app launched at AWS Summit New York (June 17), supporting Chat, Spec, and Autonomous session modes with cloud compute continuity. Web-based Automations (June 19) enabled scheduled autonomous tasks with cron scheduling. v2.10.0 (June 25) delivered MCP and agent config hot-reload without session restarts. The separate Kiro power for AWS DevOps Agent (June 19) connected IDE workflows to cloud operations. AWS also released Blocks, an open-source TypeScript framework designed for AI agents to write correct backends (June 23), and the DevOps Agent gained release readiness review and autonomous release testing (June 17).

Technology Shifts

  • "Loop engineering" displaced prompt engineering as the dominant paradigm for AI-assisted development. Boris Cherny, Claude Code's creator at Anthropic, stated he no longer prompts Claude directly β€” "my job is to write loops." Google's Addy Osmani formalized the pattern, mapping six primitives: automations, worktrees, skills, connectors, sub-agents, and memory. The key architectural insight was splitting the writing agent from the verification agent β€” a second model with different instructions grades output, enabling trust in unattended execution. OpenAI engineer Peter Steinberger advocated designing persistent loops where Codex maintains repositories autonomously on 5-minute wake cycles. This represented a fundamental transition from interactive AI assistance to long-running autonomous software workers. The implications were concrete: Block revealed that BuilderBot merges 1,500 PRs per week (~15% of all production code) via AI agents orchestrated from Slack, and Vercel reported agents triggering 29% of its platform deployments, up from less than 3% a year ago. The shift created new infrastructure requirements β€” Codex v0.142.0 introduced rollout token budgets to cap runaway loop costs, and freeCodeCamp published a tutorial on production-safe agent loops motivated by real disasters including a Claude Code recursion loop that burned $16K–$50K in five hours.

  • The agent harness layer β€” not the model β€” became the strategic control point for platform vendors. SWE-bench scores narrowed to a tight band across Claude Code, Cursor, Codex, and Antigravity, effectively demoting the model as the differentiator and shifting competition to workflow, distribution, and pricing layers. Two architectures emerged as dominant: OpenClaw's gateway-first approach (adopted by Microsoft for Scout, Nvidia for NemoClaw, with 380K GitHub stars) and Hermes Agent's memory-first design (160K stars, leading OpenRouter's token rankings with 22+ trillion total tokens). Nvidia assigned full-time developers to OpenClaw and published enterprise blueprints. Microsoft built its flagship Scout agent on OpenClaw rather than a proprietary runtime, contributing policy controls upstream. The AGENTS.md convention became a cross-tool standard for repository-level agent configuration, with GitHub Copilot code review, GitLab Developer Flow, and Claude Code all reading from it. Three "team layer" platforms β€” Augment Code's Cosmos, Cognition's Devin Desktop, and Microsoft's Rayfin β€” launched within days of each other, addressing coordination, management, and governance of agent fleets.

  • Runtime verification replaced static analysis as the essential quality gate for agent-generated code. Greptile, Cursor, OpenAI Codex Cloud, and Devin all moved from reviewing diffs to executing agent-generated code in sandboxed environments before human handoff. Azure Container Apps Sandboxes entered public preview (June 12), launching microVMs in under one second with deny-by-default network egress. Microsoft's MXC SDK offered process-level through planned micro-VM isolation. CircleCI's Chunk Sidecars moved CI validation directly into agents' inner development loops. The verification problem was quantified: agents that write their own mocks test against their own assumptions, and a green local run provides no evidence a change works against real services. For distributed systems, the next frontier was verifying changes against shared production-like environments with request-level traffic isolation, since the most expensive bugs β€” contract drift, serialization mismatches, cross-service retry failures β€” only surface at integration boundaries. All four major cloud platforms converged on session-aware agent runtimes, establishing the agent session as the new unit of enterprise compute.

  • MCP matured from protocol to enterprise infrastructure with authorization, security, and discovery standards. MCP Enterprise-Managed Authorization reached stable (June 18), giving IT teams centralized governance through existing identity providers β€” Anthropic, Microsoft, and Okta were first movers. WebMCP entered Chrome 149 origin trials (June 13), giving browser-based agents a 90% reduction in token usage by exposing typed tools directly. The Agentic Resource Discovery specification (June 17), co-developed by GitHub, Google, Hugging Face, and Microsoft, enabled dynamic discovery and ranking of MCP servers without preloading. HashiCorp's Terraform MCP Server reached GA (June 13). Azure API Management shipped content safety policies covering MCP tool calls and Agent-to-Agent payloads (June 10). Angular published official Agent Skills to keep AI coding tools on modern patterns (June 12). However, MCP also created new attack surfaces β€” the "agentjacking" attack via Sentry MCP integrations (June 21) demonstrated that AI agents cannot distinguish data from instructions when consuming MCP tool output, achieving 85% exploitation success rates.

  • "Context debt" emerged as the successor to "technical debt" as the primary engineering challenge. Postman CEO Abhinav Asthana estimated teams have six to nine months before context debt outruns their ability to manage it manually. The problem is the rapid accumulation of APIs, services, and dependencies that no one fully understands β€” compounded by AI-generated code flooding codebases. Cursor, Windsurf, Claude Code, and GitHub Copilot all converged on context layers, treating system-wide understanding as a bottleneck equal to model capability. Pre-integrated context lakes were shown to reduce AI agent costs by 80% compared to direct MCP server access in a 12,000-query experiment. Nx launched Polygraph, a "meta-harness" stitching multiple repositories into synthetic monorepos with shared memory across sessions. Stack Overflow pivoted to serve AI coding agents with an API-first knowledge exchange addressing the "ephemeral intelligence gap" where agents repeatedly rediscover the same fixes in isolation. The AWS Agent Toolkit rules file demonstrated that without a 17-line context file, agents with 300+ APIs and 20+ skills still default to answering from training data rather than using available tools.

  • Version control itself began fracturing under the weight of AI-generated commits. Cursor announced Origin (June 18), a Git-compatible hosting platform purpose-built for AI agent workloads, directly challenging GitHub's dominance. GitLab announced Project Switch, redesigning the underlying architecture for server-side agent queries claiming 50x faster task execution. Zed announced DeltaDB, replacing Git's commit-based model with continuous fine-grained deltas linked to conversations. GitHub Desktop 3.6 shipped native worktree support specifically to address multi-branch parallelism that coding agents require. The pressure was quantifiable: GitHub processed 1.4 billion commits monthly (versus 1 billion in all of 2025), with agents creating 17 million PRs monthly. Former GitHub developer advocacy director Brian Douglas captured the shift: "Tokens are a better metric than commits β€” they align to a dollar spent that correlates to the effort of work."

  • Government regulation of frontier AI models became a direct factor in development tool selection. The US Commerce Department's export control on Anthropic's Fable 5 and Mythos 5 (June 12) was the first time export controls targeted a cloud-hosted AI model. GPT-5.6 Sol was subsequently restricted to government-approved preview access (June 25–26). These actions demonstrated that standardized agent tooling β€” MCP, AGENTS.md, OpenClaw β€” made hosted frontier models replaceable within days, as four open-weight alternatives provided immediate fallback. The regulatory uncertainty accelerated enterprise interest in self-hosted models and multi-model strategies. Indian tech leaders called for a $5B annual sovereign AI fund. The broader lesson for development teams: wire workflows so swapping models is a config change rather than a rewrite.

Developer Impact

  • The "AI Paradox" crystallized: developers write code faster but software delivery hasn't accelerated. GitLab's survey of 1,528 developers found that while 78% reported faster code generation, 85% confirmed the bottleneck had shifted from writing code to reviewing it. Teams with high AI adoption merged 98% more PRs while review times climbed 91%. A QCon New York presentation revealed that AI generates 50–75% of daily code for the majority of engaged attendees, but the Stack Overflow 2025 survey of 50,000 developers showed positive sentiment toward AI tools dropping from 70%+ to approximately 60%. CircleCI data showed feature branch activity increasing significantly while production deployments lagged. The economic inversion was stark: code became cheaper to create but dramatically more expensive to verify. Sonar's survey found 96% of developers distrust AI-generated code, yet only 48% consistently verify it, while 88% report negative impacts on technical debt. Multiple solutions emerged β€” Qodo's cross-repo code review, the "AI slop registry" concept of invariant verification criteria, and the proposed shift from reviewing AI-generated diffs to verifying declared intent upstream.

  • Model triage became an essential developer skill as cost differentials widened dramatically. Mitchell Hashimoto's head-to-head tests showed Fable 5 costing $9 and taking 40 minutes for tasks GPT-5.5 completed at $1.50 and Zhipu's GLM-5.1 at under $1 β€” with equally acceptable results. The emerging pattern was to use expensive frontier models only for planning and review while routing execution to cheaper alternatives. Fable 5 users on Max20 plans reported burning 2% usage per minute, far faster than Opus 4.8. Claude Code's Dynamic Workflows addressed this with model routing β€” assigning cheaper models to simpler stages while reserving capable models for deep reasoning. Copilot's Auto Mode (GA June 17) dynamically routed across Claude Sonnet 4.6, GPT-5.4, and Haiku 4.5 with a 10% token discount. The practical advice consolidated: developers needed to understand per-model costs, match model capability to task complexity, and configure fallback chains for resilience. Free and Student plan users lost manual model selection entirely on June 24, with GitHub making auto-selection mandatory.

  • Mobile-first AI coding emerged as a legitimate workflow paradigm. Cursor launched its iOS app in public beta (June 29), Kiro shipped a native iOS app for mobile agent supervision (June 17), and Claude Code creator Boris Cherny stated publicly that most of his coding now happens on his phone. The shift reflected a fundamental change: AI coding was moving from code-writing assistance toward agent oversight β€” a workflow that doesn't require large screens or IDE access. Cursor's mobile app featured Remote Control for directing desktop-running agents, Live Activities on the lock screen for status tracking, and push notifications for agent completion. Kiro Mobile supported Chat, Spec, and Autonomous session modes with cloud compute continuity. These weren't toy features β€” they represented the natural endpoint of loop engineering, where the developer's role becomes launching, steering, and approving autonomous agents rather than writing code character by character.

  • Enterprise development teams began building dedicated internal platforms rather than relying on off-the-shelf AI coding tools. Dropbox unveiled Nova for orchestrating AI coding agents within its monorepo, CI systems, and infrastructure. LinkedIn shared its MCP-based platform for scaling agentic engineering across 10,000+ repositories. Block revealed BuilderBot merging 1,500 PRs weekly via agents orchestrated from Slack. These platforms emphasized that surrounding infrastructure β€” CI integration, observability, governance, shared context β€” matters as much as model quality. The pattern echoed how version control grew from personal convenience into shared infrastructure with branches, reviews, and policies. For smaller organizations, AWS Blocks, Vercel's Eve, and Azure Functions' .agent.md programming model provided lighter-weight approaches to the same problem.

  • Engineering hiring proved resilient to AI disruption, but the nature of roles shifted. SignalFire's analysis showed engineers comprising 55% of new hires at major tech companies β€” up from 46% in 2019. A Jevons paradox was at work: AI-powered productivity expanded work rather than eliminating it. Early-stage startups hired 7% more engineers than in 2019, and Nvidia CEO Jensen Huang stated that "software engineers are busier than ever." However, a Linux Foundation report found entry-level roles contracting by 3%, as AI handles many tasks junior developers traditionally learned on. Organizations were 3.7Γ— more likely to upskill existing staff than hire new employees. The emerging cross-domain profile combined software engineering, AI fluency, security awareness, and product thinking. Netlify's CTO defined "Agent Experience" as a new engineering discipline β€” designing systems where agents and humans collaborate at machine speed. Engineers were evolving from code authors to what she called "shepherds of production."

  • The developer tooling ecosystem fragmented further while simultaneously standardizing on common protocols. On one hand, fragmentation accelerated: Google killed Gemini CLI and replaced it with closed-source Antigravity CLI (June 20), Cursor launched Origin as a GitHub competitor, GitLab shipped Project Switch, and tool-specific pricing made switching costs non-trivial. On the other hand, standardization advanced rapidly: AGENTS.md became a cross-tool convention, MCP Enterprise-Managed Authorization reached stable, the ARD specification enabled cross-platform tool discovery, the Agent Control Specification provided a portable policy standard for governing agent behavior across frameworks, and the Appia Foundation (13 founding members including Google, Microsoft, and OpenAI) launched to build AI's missing trust and compliance verification layer. The net effect for developers was that individual tool choices mattered less than they used to β€” the same AGENTS.md, skills files, and MCP servers worked across Claude Code, Copilot CLI, Codex, and Kiro β€” but billing and governance lock-in was increasing through platform-specific credit systems and enterprise controls.