30 Tools: Your AI Subscription Has a Meter. You Just Can't See It.

Andrej Karpathy told the No Priors podcast he feels "nervous when I have subscription left over." The former OpenAI researcher and Tesla AI lead treats unused tokens the way a PhD student treats idle GPUs: as waste. Jensen Huang, speaking on the All-In Podcast the same week, said any $500,000 engineer not consuming at least $250,000 in tokens should leave him "deeply alarmed."

Both men assume you know what you're consuming. You don't.

Neither ChatGPT nor Claude nor Gemini provides a real usage dashboard for paying subscribers. OpenAI shows you a message limit only after you've hit it. Anthropic's usage page displays rounded percentages buried in account settings. Google's Gemini offers subscribers minimal consumption visibility. You pay $20 to $200 every month for frontier AI models, and the companies selling you that access won't tell you how much of it you've used.

The result: more than 30 third-party tools now exist to do what a single progress bar could accomplish. That cottage industry is the tell.

Why developers started counting their own tokens

Three pressures hit at the same time in late 2025.

First, developers started stacking subscriptions. Claude Code Max alone runs $100 to $200 a month. Add Cursor Pro at $20 (or $60 for Pro+, $200 for Ultra). Toss in GitHub Copilot. Suddenly you're looking at $150 to $400 a month in AI tooling, and nobody's tracking where it all goes. That's $4,800 a year on tools that won't even show you a receipt.

Then the pricing models started moving underneath people's feet. Cursor switched from request-based to credit-based billing in June 2025. Windsurf moved from credits to daily quotas in March 2026. Nobody has settled on a stable model yet. Subscribers trying to budget against a moving target feel anxious, and rightly so.

Third, rate limits became the actual bottleneck. Claude Code Max plans tie limits to five-hour rolling windows, roughly 88,000 tokens for the 5x plan and 220,000 for the 20x. Hitting a rate limit mid-session kills your flow. Knowing whether you're spending tokens on real work or a runaway agent prompt is the difference between a productive afternoon and a frustrating one.

So the community built what the providers wouldn't. Here's everything they made.

Menu bar apps (macOS)

These sit in your toolbar and show live consumption. Glance up, see your burn rate, get back to work.

1. CodexBar

What it does: Shows current session spend and weekly limits directly in the macOS menu bar across 15+ providers including Claude Code, Codex CLI, Cursor, Gemini, GitHub Copilot, and OpenRouter. Dual-meter format with credits and countdown timers.

Install: brew install --cask codexbar

Pro: Broadest provider coverage in any menu bar tool. Free, open source, Homebrew install. 9,400 GitHub stars signal real community trust. Zero configuration needed.

Con: macOS only. No runaway-prompt detection. Reads local session files, so it can't track web-based usage like claude.ai conversations.

Price: Free (MIT license) | GitHub: steipete/CodexBar

2. TokenBar

What it does: Tracks usage across 20+ providers including OpenAI, Claude, Cursor, OpenRouter, Perplexity, Vertex AI, DeepSeek, and Mistral. Key differentiator: runaway-prompt detection that catches retries and background agent activity before costs spiral.

Pro: The only menu bar app with built-in anomaly detection. Community reports describe single Explore() operations consuming 90,000 tokens before you notice. TokenBar catches those. Pace indicators and incident alerts when spending snowballs.

Con: Not open source. macOS only. Smaller community than CodexBar, so edge-case bugs take longer to surface.

Price: $4.99 one-time | tokenbar.site

3. Tokemon

What it does: Focused exclusively on Claude Code. Polls usage every 30 seconds. Burn rate per hour, per-project token breakdowns, 24h/7d usage charts, and estimates for when you'll hit limits.

Pro: Deepest Claude Code analytics of any menu bar tool. Team analytics via Admin API and organization-wide dashboards. Raycast integration for keyboard-first workflows. Open source (MIT).

Con: Claude Code only. If you use Cursor or Copilot alongside Claude, you'll need a second tool. Polling every 30 seconds may feel aggressive to some.

Price: Free (MIT license) | tokemon.ai

4. TokenKite

What it does: Tracks Claude and Cursor usage from the menu bar. Two tracking modes: local estimation (parses ~/.claude/projects/ logs, fully offline) or OAuth-based precise tracking that pulls data directly from Anthropic's API.

Pro: The OAuth mode gives exact quota status, not estimates. Stored securely in Keychain. Per-model cost breakdowns (Opus vs. Sonnet vs. Haiku). Shows API-equivalent costs so you see subscription value. No Full Disk Access required.

Con: Only supports Claude and Cursor. No telemetry cuts both ways: great for privacy, but no community benchmarking.

Price: Free | tokenkite.com

5. OpenUsage

What it does: Plugin-based architecture where every provider is a module. Currently supports Codex, Claude, Cursor, Copilot, Antigravity, Amp, Factory, Gemini, JetBrains AI, Kiro, Kimi, MiniMax, OpenCode, Perplexity, Synthetic, and Windsurf. Exposes a local API at localhost:6736 that any script or editor status line can read.

Pro: Most extensible architecture. The local REST API (curl localhost:6736/v1/usage) means any tool in your stack can consume the data. Not just a viewer but a usage data platform. Plugin model means new providers ship fast.

Con: Newer project, so documentation and community support are still thin. The API-first design means the menu bar UI itself is minimal compared to CodexBar or TokenBar.

Price: Free | openusage.ai

6. SessionWatcher

What it does: Lightweight tool for Claude Code and Codex CLI. Rolling 5-hour window tracking and countdown. No API keys needed.

Pro: The simplest option. Install it, forget it exists until you need it. Tiny resource footprint.

Con: Only two providers. No historical analytics, no per-project breakdowns. You're paying for simplicity.

Price: $1.99 one-time | sessionwatcher.com

7. Cursor Usage Widget

What it does: Cursor-specific menu bar monitor. Tracks subscription status, model-specific usage breakdowns, monthly spending totals, and days until subscription end. Configurable refresh intervals and warning thresholds.

Pro: Purpose-built for Cursor. Shows per-model breakdown (Auto, Agent, Claude, GPT, etc.) with input/output token analysis and cache optimization metrics. Historical month navigation.

Con: Cursor only. If you use multiple AI coding tools, this solves one piece of the puzzle.

Price: Paid (Mac App Store) | cursorusage.com

CLI trackers (cross-platform)

These tools aggregate token usage data from multiple AI coding tools by reading local session files. No API keys required. All data stays local.

8. ccusage

What it does: The most mature and popular CLI tracker. Supports Claude Code and Codex CLI with daily, monthly, and session-based usage breakdowns including cost analysis.

Install: npx ccusage

Pro: 12,000 GitHub stars. The largest community of any tracking tool, which means bugs get caught fast and features ship regularly. The npx install means zero setup, no global packages.

Con: Only supports Claude Code and Codex CLI. If you use Cursor, Windsurf, or Copilot, ccusage is blind to that spending.

Price: Free (open source) | GitHub: ryoppippi/ccusage

9. Agentlytics

What it does: Broadest editor coverage available. Supports Cursor, Windsurf, Claude Code, VS Code Copilot, Zed, Antigravity, OpenCode, and Command Code, 16 editors total. Unified dashboard with KPIs, heatmaps, usage streaks, per-project analytics, and side-by-side editor comparisons.

Install: npx agentlytics

Pro: If you use multiple coding tools, nothing else comes close to Agentlytics for coverage. Zero configuration. Team Relay mode for sharing usage across organizations. All data local.

Con: 359 GitHub stars means a smaller community than ccusage. The breadth-over-depth tradeoff means individual editor integrations may be less polished than purpose-built alternatives.

Price: Free (open source) | GitHub: f/agentlytics

10. tokscale

What it does: Supports Claude Code, OpenCode, OpenClaw, Codex CLI, Gemini CLI, Cursor, AmpCode, Factory Droid, Kimi, and Pi. Includes a global leaderboard, 2D/3D contribution graphs, and Kardashev-scale gamification.

Install: npx tokscale

Pro: Broad coverage with a playful design. Platform-specific filters (tokscale --claude). The gamification angle actually helps some developers stay conscious of their consumption patterns.

Con: The leaderboard and gamification features won't appeal to everyone. 1,400 stars, growing but still young.

Price: Free (open source) | GitHub: junhoyeo/tokscale | tokscale.ai

11. toktrack

What it does: Built in Rust with SIMD acceleration. Scans 3,500+ session files in approximately 40 milliseconds. Supports Claude Code, Codex CLI, Gemini CLI, and OpenCode. Daily/weekly/monthly views with model-level breakdowns.

Install: cargo install toktrack

Pro: The fastest scanner in the category, by a wide margin. If you have months of session history and want instant results, toktrack delivers. Preserves cost history even after CLI tools delete session files. Shareable SVG receipts.

Con: Requires Rust toolchain to install. 71 stars, which means a small contributor base. No Cursor or Copilot support.

Price: Free (open source) | GitHub: mag123c/toktrack

12. tokenusage (tu)

What it does: Rust-based CLI/TUI/GUI tool supporting Claude Code, Codex CLI, and Antigravity. Claims to be 214x faster than ccusage for Claude and 138x faster for Codex.

Install: cargo install tokenusage

Pro: Speed benchmarks are impressive if verified. Live monitoring mode and image generation for shareable usage cards give it a unique social angle.

Con: Only 8 GitHub stars. The speed claims need independent verification. Narrow provider support.

Price: Free (open source) | GitHub: hanbu97/tokenusage

13. tokentop

What it does: A real-time terminal TUI described as "htop for AI costs." Connects to Anthropic, OpenAI, and Gemini via OAuth. Tracks per-request cost, model, tokens, and duration with daily/weekly/monthly budgets.

Install: go install github.com/tokentopapp/tokentop@latest

Pro: The TUI design feels native for terminal-heavy developers. Spending limits with visual warnings. Cache analysis. Four different dashboard views.

Con: Requires Go toolchain. OAuth setup adds friction compared to tools that just read local files. 40 stars.

Price: Free (open source) | GitHub: tokentopapp/tokentop | tokentop.app

Browser extensions

These inject token counts directly into AI chat interfaces. No terminal required.

14. Claude Counter

What it does: A minimal browser extension that shows token count, cache timer, and usage bars on claude.ai. Intercepts Claude's SSE data stream to display exact, unrounded utilization fractions, more accurate than what Anthropic shows on its own /usage page.

Pro: Surfaces data Anthropic already collects but chooses not to display. The cache timer tells you whether continuing a conversation is cheaper than starting fresh. Works on Chrome, Edge, and Firefox.

Con: Only works on claude.ai web interface. No CLI or API tracking. Manual install (download zip, enable Developer mode) since it's not on the Chrome Web Store.

Price: Free (MIT license) | GitHub: she-llac/claude-counter

15. Claude Usage Tracker (Chrome extension)

What it does: Tracks token consumption from uploaded files, project knowledge, chat history, AI responses, Google Drive syncs, and MCP integrations. Sends notifications when usage limits replenish. 10,000+ Chrome Web Store users.

Pro: The most comprehensive Claude web tracking. Calculates tokens from sources most tools miss: project knowledge files, custom instructions, personal preferences. Tooltips on every element explain what you're seeing. 4.7 stars from 41 ratings.

Con: Claude.ai only. Some MCP integrations return "Knowledge" objects the extension can't access. Token calculation uses either Anthropic's API (needs key) or gpt-tokenizer (approximate).

Price: Free | Chrome Web Store

16. Claude Usage Tracker (desktop app)

What it does: Different from #15. A native macOS app (with browser mode for Windows/Linux) that auto-detects 9+ tools including Claude Code CLI, Cursor, Windsurf, and Cline. Scans local session data and shows daily costs, model breakdowns, heatmaps, session logs, and monthly projections.

Pro: Works retroactively. Scans existing JSONL and log files the moment you open it, so your full usage history appears immediately. 100% local, no accounts, no telemetry. Open source (MIT). The heatmap visualization is genuinely useful for spotting usage patterns.

Con: Cost calculations use API token rates, not subscription pricing. For Pro/Max subscribers, the numbers show what usage would cost at API rates, not what you're actually paying. Per-project breakdown not yet available (on the roadmap).

Price: Free (MIT license) | Product Hunt

17. ChatGPT Token Counter (Chrome)

What it does: Monitors conversation tokens in ChatGPT and warns when you're nearing limits that OpenAI won't surface. Rated 4.8 stars on the Chrome Web Store.

Pro: Fills the exact gap OpenAI refuses to fill. Simple, focused, well-rated.

Con: ChatGPT only. Limited analytics compared to the Claude tracking ecosystem.

Price: Free | Chrome Web Store

18. Cursor Token Tracker (Chrome)

What it does: Reads data from cursor.com/dashboard and provides total cost tracking, daily averages, budget projections, cost trends, per-model breakdowns, most-active-time analysis, and CSV export.

Pro: Open source. Budget notifications at 80% and 90% usage thresholds. All processing happens locally.

Con: Requires you to visit cursor.com/dashboard. Not a passive background tracker.

Price: Free (open source) | Chrome Web Store

19. Chatterclock (Chrome)

What it does: Automatically counts ChatGPT messages and segments them by model. Tracks when you sent each message to help you understand rolling window resets.

Pro: Simple answer to the question "how many GPT-4o messages do I have left?" No configuration.

Con: Message counting only. No token-level granularity, no cost analysis.

Price: Free | Chrome Web Store

Claude Code-specific tools

Purpose-built monitoring for Claude Code power users who want deeper telemetry than general-purpose trackers provide.

20. Claude Code Usage Monitor

What it does: Real-time terminal monitoring with ML-based predictions, P90 percentile calculations, and Rich UI. Auto-detects your plan type and adapts limits based on your actual usage patterns from the past eight days.

Install: uv tool install claude-monitor or pip install claude-monitor

Pro: The ML prediction engine is unique. It learns from your last 192 hours of sessions to calculate personalized limits. Multi-level warning system with cost and time predictions. Configurable refresh rates from 0.1 to 20 Hz.

Con: Python dependency. The "smart" features add complexity. If you just want a number on screen, this is overkill.

Price: Free (open source) | GitHub: Maciek-roboblog/Claude-Code-Usage-Monitor

21. claude-code-otel

What it does: Full observability stack combining OpenTelemetry, Prometheus, Loki, and Grafana. Tracks cost, tokens, sessions, tool usage, and user activity metrics (DAU/WAU/MAU).

Pro: The most comprehensive monitoring solution if you're willing to run the infrastructure. Pre-built Grafana dashboards. If you already run a Prometheus stack, this drops right in.

Con: Requires running Prometheus, Loki, and Grafana. Massive overkill for individual developers. Built for teams.

Price: Free (open source)

22. claude_telemetry

What it does: Drop-in replacement for the claude command (aliased as claudia). Exports telemetry data to Logfire, Sentry, Honeycomb, or Datadog.

Pro: If your team already pays for Datadog or Honeycomb, this sends Claude Code data to the dashboards you already check. Zero new infrastructure.

Con: Requires an existing observability platform. Not useful as a standalone tool.

Price: Free (open source)

23. claude-code-metrics-stack

What it does: Local Grafana dashboard tracking cost, tokens, sessions, and productivity metrics.

Pro: Self-contained. Runs locally without cloud dependencies. Good middle ground between raw CLI output and a full OTEL stack.

Con: Still requires running Grafana locally. Less polished than commercial alternatives.

Price: Free (open source)

24. claude-code-usage-analyzer

What it does: Detailed breakdowns by model and token type with statistical insights including mean, median, and P95 percentiles.

Pro: Statistical depth. If you want to know your P95 token consumption per session, this is the tool.

Con: Analysis-only. No real-time monitoring, no alerts.

Price: Free (open source)

25. cccost

What it does: Instruments Claude Code for actual token tracking. Outputs .usage.json files that statusline scripts can read.

Pro: The lightest option in this category. Designed to feed data into your existing terminal setup (tmux statusline, Starship prompt, etc.) rather than imposing its own UI.

Con: Requires scripting knowledge to do anything useful with the JSON output. No built-in visualization.

Price: Free (open source)

26. MyTokenTracker

What it does: Web dashboard with per-project cost tracking, model breakdowns, session history, and subscription savings analysis. Uses a lightweight Claude Code hook that runs after each session.

Pro: The only tool with per-project cost breakdowns in a web UI. Shows week-over-week trends and subscription savings vs. API costs. Works with CLI, VS Code, and JetBrains.

Con: Data leaves your machine (sent to their dashboard). Free tier limited to 100 session logs per day and 7-day retention. Pro plan is a recurring $9.99/month.

Price: Free tier (100 sessions/day) | Pro $9.99/mo | Team $29.99/mo | mytokentracker.io

Enterprise and observability platforms

For teams managing AI spend across an organization. These track API usage at the infrastructure level.

27. Langfuse

What it does: Open-source observability platform. Real-time analytics with support for cached, audio, image, and reasoning tokens. Integrates with 50+ libraries and frameworks. OTEL-native.

Pro: 23,900 GitHub stars. Self-hostable. The most mature open-source option for production LLM monitoring. Free tier includes 50,000 units/month.

Con: Enterprise-grade complexity. Not designed for individual developers tracking personal subscriptions. Setup requires infrastructure.

Price: Free tier (50K units/mo) | Pro $59/mo | langfuse.com

28. LangSmith

What it does: LangChain's observability platform. Detailed cost breakdowns for LLM calls, tools, and retrieval steps. Three views: Trace Tree (per-run breakdown), Project Stats (aggregated), and Dashboards (trends).

Pro: Deep LangChain integration. Automatic cost calculations for major providers. Customizable spending limits and trace retention.

Con: Enterprise pricing starts at $75,000. The LangChain ecosystem tie-in is a strength or a lock-in depending on your stack.

Price: Free tier | Enterprise from $75,000 | smith.langchain.com

29. Helicone

What it does: Open-source LLM observability. Supports 100+ providers. Proxy-based architecture with caching, rate limiting, and cost tracking.

Pro: One-line integration (just change your base URL). The proxy approach means you don't need to modify application code. Caching can reduce costs substantially.

Con: Proxy architecture adds a network hop. Self-hosting requires infrastructure management.

Price: Free tier | Cloud pricing scales with usage | helicone.ai

30. Portkey

What it does: AI gateway handling 50 billion tokens daily across 200+ providers. Semantic caching, budget controls, and an Analytics API for custom billing dashboards.

Pro: Scale. If your organization processes billions of tokens, Portkey is built for that volume. Semantic caching cuts token usage by 30-90% on repetitive tasks. Hard spending limits prevent bill shock.

Con: Enterprise pricing. Not for individual developers. The gateway architecture means all your AI traffic routes through Portkey's infrastructure (or your self-hosted instance).

Price: Free tier | Enterprise pricing | portkey.ai

31. Arize

What it does: Enterprise monitoring using OpenInference standards. Tracks prompt, completion, and total token counts. Includes a Prompt Playground for side-by-side model testing. Cache token tracking.

Pro: Standards-based approach (OpenInference) means less vendor lock-in. The Prompt Playground lets you A/B test models and see cost differences before committing.

Con: Enterprise pricing and complexity.

Price: Open-source tools free | Cloud from $50/mo | arize.com

32. Maxim AI (Bifrost)

What it does: Semantic caching gateway that delivers cached responses in under 50 milliseconds. Reduces API spending by 20-40% on predictable queries. Smart routing directs simple tasks to cheaper models automatically.

Pro: The caching and routing features can pay for themselves in a single billing cycle. Integrates with 12+ providers.

Con: Gateway architecture. Another piece of infrastructure to manage and depend on.

Price: Contact for pricing | getmaxim.ai

33. Vantage

What it does: Native Cursor cost tracking with breakdowns by model, token type, developer, and usage category. Distinguishes between included-plan usage and overage charges.

Pro: The only tool that separates subscription-included usage from overages. Shows whether your plan tier is right-sized. Monitors max-mode usage separately. Five-minute setup.

Con: Cursor-focused. Enterprise pricing.

Price: Free tier | Enterprise pricing | vantage.sh

Routing and multi-provider gateways

34. 9Router

What it does: Open-source smart routing tool supporting Claude Code, Cursor, Antigravity, Copilot, Codex, Gemini, OpenCode, Cline, and OpenClaw. Three-tier smart fallback routing with quota tracking and auto token refresh.

Pro: If one provider hits a rate limit, 9Router automatically routes to the next. Quota tracking for Claude Code, Codex, and Gemini plus spending limits per provider.

Con: Adds routing complexity. Newer project with less community testing.

Price: Free (open source)

35. OpenRouter

What it does: Commercial routing service providing access to hundreds of models. Activity dashboard with CSV/PDF export, per-key credit limits with auto-reset.

Pro: Enterprise usage monitoring with filtering by model, API key, and time period. If you route through OpenRouter, you get cost tracking for free as a side effect.

Con: Commercial service, not a standalone tracking tool. You're adopting a provider, not just a monitor.

Price: Pay-per-token | openrouter.ai

Team and financial tracking

36. Faros AI

What it does: Connects AI tool usage from Claude Code, GitHub Copilot, and Cursor to engineering outcomes. Tracks token usage, costs by model, and output metrics (commits, PRs). Correlates to DORA metrics: lead time, deployment frequency, and change failure rate.

Pro: The only tool that answers "are we getting engineering value from our AI spend?" Connects tokens to code output, not just cost. Single pane of glass across all AI coding tools.

Con: Enterprise pricing and sales process. Requires connecting multiple data sources.

Price: Enterprise | faros.ai

37. ToolSpend

What it does: Connects AI service subscriptions with banking data to show actual spend across ChatGPT, Claude, Cursor, Perplexity, ElevenLabs, and more.

Pro: Identifies underutilized seats and duplicate tools. Pre-renewal alerts with team-level attribution. Bridges the gap between engineering metrics and finance.

Con: Requires connecting banking/payment data. Privacy implications for some organizations.

Price: Contact for pricing

38. Prompts.ai

What it does: Tracks tokens in real-time within a prompt editor. Supports 35+ models. Pay-as-you-go TOKN credit system. Cost estimates update instantly when switching models.

Pro: Token tracking built into the editing workflow, not bolted on after. Audit trails and compliance tools for regulated industries.

Con: Locks you into the Prompts.ai editor. Not a monitoring overlay for your existing tools.

Price: Pay-as-you-go (TOKN credits) | prompts.ai

Web-based calculators

39. findskill.ai Token Calculator

What it does: Free web tool that estimates token counts using the characters/4 heuristic. Supports all major models with context window and pricing tables.

Pro: No install, no account. Paste text, get a count. Good reference for context window limits and per-token pricing across models.

Con: Estimates only (accurate to within 10% for English text). Less accurate for code, JSON, non-English text, or text with special characters. Not a monitoring tool.

Price: Free | findskill.ai/blog/ai-token-counter

40. OpenAI Tokenizer

What it does: Official OpenAI tool for exact token counts using the tiktoken tokenizer.

Pro: Exact counts for OpenAI models. The ground truth.

Con: OpenAI models only. Web interface, not an API you can integrate.

Price: Free | platform.openai.com/tokenizer

Which stack should you install?

If you're reading this list and feeling overwhelmed, here are three starting points.

Individual developer, free: Install CodexBar for always-on menu bar monitoring (brew install --cask codexbar). Run Agentlytics weekly for per-project breakdowns (npx agentlytics). That two-tool stack covers real-time awareness and historical analysis across every major provider.

Claude Code power user: Add ccusage (npx ccusage) for the deepest Claude-specific analysis. If you want predictions and burn-rate alerts, layer on Claude Code Usage Monitor (pip install claude-monitor). For per-project cost breakdowns in a web dashboard, try MyTokenTracker.

Team or enterprise: Start with Langfuse (self-hosted, open source, 23.9k stars) or Helicone for observability. Add Faros AI if you need to connect token spend to engineering outcomes. For Cursor specifically, Vantage gives you the clearest view of included vs. overage costs.

The math behind the curtain

All of this exists because the economics demand opacity. A $20 Claude subscriber costs Anthropic roughly $60 in compute. A $200 Max subscriber runs up approximately $570 in inference costs. Every active subscriber is a loss leader, subsidized by the ones who pay but barely use the service.

One power user documented 10 billion tokens over eight months. At API rates, that would have cost north of $15,000. What did this person actually pay? Eight hundred dollars in subscription fees. The quiet majority who subscribe, send a dozen messages a week, and never think about context windows picked up the tab.

If every subscriber could see their usage in real time, two things would happen. Light users would realize they're overpaying and downgrade. Heavy users would optimize their consumption, extracting even more value from plans that already lose money. Both outcomes hurt the provider.

This is the gym membership model applied to AI. The economics depend on people paying for capacity they don't touch. Transparency is the enemy of that model. And the providers know it.

What the numbers actually look like

When developers measure their consumption with these tools, the results are stark. Community analyses estimate average Claude Code usage at about $6 per developer per day in API-equivalent spend, with the 90th percentile under $12. That translates to $180 a month in compute, which is why the Max 5x plan at $100 always looked like a bargain. It was. For the subscriber.

But the extremes from community forums are staggering. One Cursor user burned 170 million tokens in two days. Someone on Claude Code hit $150 in API-equivalent costs in 48 hours on a mid-size repo. A CTO reported $4,000 in two weeks. And one developer generated 28 million tokens to produce 149 lines of code, a debugging spiral that cost more in compute than most developers spend in a month.

Every single data point came from the third-party tools listed above. The providers contributed nothing.

The test for every AI subscription provider is simple. If the tools you sell promise productivity but the costs stay hidden, the current arrangement benefits exactly one party. And it's not the one with the credit card on file.

The meter exists. They just won't show it to you. So 40 open-source projects did it instead.

OpenAI Ships a Codex Plugin for Claude Code, Putting Its Agent Inside a Rival's Tool

OpenAI on Monday open-sourced a plugin that lets developers run its Codex coding agent directly inside Anthropic's Claude Code, its most prominent competitor. The plugin, built by OpenAI's Dominik Kun

The Implicator

GitHub Reverses Policy, Will Train AI on Copilot User Data Starting April 24

GitHub announced on March 25 that it will begin using interaction data from Copilot Free, Pro, and Pro+ subscribers to train its AI models, reversing a policy that previously excluded all plan tiers f

The Implicator

OpenAI Launches GPT-5.4 With Built-In Computer Use and 1 Million Token Context Window

OpenAI released GPT-5.4 on Thursday, its latest frontier model that combines coding improvements from GPT-5.3-Codex with two capabilities new to its mainline lineup, native computer use and a one-mill

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]