Skip to content
Projects

QuesmaOrg/awesome-ai-tokenomics

View on GitHub

QuesmaOrg/awesome-ai-tokenomics has 219 GitHub stars and 44 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-16 00:08 UTC. Last push 2026-11-13.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

A curated list on AI token economics: what tokens cost, where they get wasted, and how to cut the bill. Tools, benchmarks, papers, and copy-paste configs for the token economy of LLMs and coding agents.

AI

README.md

Original

Awesome AI Tokenomics Awesome

Pricing, measurement, optimization, and governance of tokens used by AI models.

Tools, papers, benchmarks, and vendor pages on what AI tokens cost and where they go.

Contents

Legend

Each entry ends with a kind badge: tool: MIT (blue, with the license when known), or a gray badge for paper, bench, data, co for companies, and report. Plain entries are articles. GitHub-hosted tools also carry a live last-commit badge.

Monitor

Dashboards

  • antiburn - Desktop app for macOS, Windows and Linux that reads local session logs from coding agents, flags token-burn causes, and applies and verifies supported configuration fixes. tool: MIT last commit
  • ccusage - CLI that reads local agent logs to report token usage and cost across coding-agent sources, with caching-aware pricing. tool: MIT last commit
  • Claude Code Usage Monitor - Live terminal dashboard for Claude Code usage with burn-rate analytics, limit detection, and session-expiry forecasts. tool: MIT last commit
  • ClaudeBar - Menu-bar app for macOS that monitors AI coding quotas across multiple providers. tool: Apache-2.0 last commit
  • CodeBurn - Usage tracker for coding tools whose optimize command flags harness-waste patterns with dollar estimates. tool: MIT
  • Codex Usage Tracker - Local-first MCP tools and evidence console that index Codex CLI session logs into SQLite and report token, credit, allowance, and cost figures, with cached and uncached input split out. tool: MIT last commit
  • CodexBar - Menu-bar app for macOS that shows limits, reset timers, credit balances, and spending across AI providers. tool: MIT last commit
  • CodeZeno Usage Monitor - Windows taskbar widget that shows Claude Code quota and usage without opening a terminal. tool: MIT last commit
  • Datadog LLM Observability Cost - Datadog documentation on per-request LLM cost estimates computed from token counts and public pricing. co
  • gh-aw - GitHub Agentic Workflows runtime with per-run token and cost metering and budget caps that stop a workflow mid-run. tool: MIT last commit
  • Grafana Cloud GenAI Observability - Prebuilt Grafana Cloud dashboard for LLM cost, token usage, and latency, built on the OpenLIT SDK. co
  • OpenLIT - OpenTelemetry-native platform with a self-hosted dashboard for LLM cost, token, and latency observability. tool: Apache-2.0 last commit
  • OpenUsage - Native Swift macOS menu-bar meter for AI coding subscriptions, showing session and weekly limits, credits, and estimated spend. tool: MIT last commit
  • Opik Cost Intelligence - Comet enterprise feature that runs a local proxy on each developer machine and splits every coding-agent call into about 20 cost buckets from the provider's usage payload; Comet's 15% to 30% savings figure has no published method. tool: proprietary last commit
  • TokenTracker - Local-first token and cost dashboard for coding tools, with a desktop pet, native widgets, and achievements. tool: MIT last commit

eBPF Kernel Capture

  • AgentSight - Uses eBPF to watch an AI agent from the kernel boundary and correlate what it said it would do with what it did. tool: MIT last commit
  • OpenTelemetry eBPF Instrumentation - Zero-code eBPF instrumentation from OpenTelemetry that captures GenAI and MCP traces at the kernel layer without an SDK. tool: Apache-2.0

Observability

  • Braintrust - Closed-source eval and observability SaaS whose span metrics normalize cached prompt tokens across providers and expose per-user and per-model spend. co
  • Langfuse - Open-source platform for tracing, evaluating, and analyzing LLM and agent transcripts, with a prompt-management layer. tool: MIT

Observability Platforms

  • OpenAI Prompt Cache Diagnostics - OpenAI guide to a Responses API feature that returns a machine-readable reason why a prompt prefix was not reused from the cache. tool: proprietary

OTel for LLMs

  • OpenLLMetry - Open-source set of OpenTelemetry-based SDKs and instrumentations for LLM apps, built by Traceloop. tool: Apache-2.0 last commit
  • OpenTelemetry GenAI Semantic Conventions - OpenTelemetry conventions that define vendor-neutral token, cost, and cache attribute names for GenAI telemetry.

Tracing

  • Arize Phoenix - Source-available (Elastic License 2.0) LLM tracing platform that records per-span token counts and USD cost via OpenTelemetry. tool: Elastic-2.0 last commit
  • claude-tap - Local trace viewer that intercepts API traffic from coding agents and shows per-request token breakdowns: input, output, cache read, cache creation. tool: MIT last commit
  • LangSmith Cost Tracking - LangChain documentation on cost tracking in LangSmith, its commercial LLM and agent observability service. co
  • Opik - Open-source (Apache-2.0) LLM observability platform from Comet, with per-span USD cost estimated from token usage. tool: Apache-2.0 last commit

Optimize

Caching

  • CacheLint (Silent and Costly) - Paper and static-analysis detector for prompt-caching mis-integration in agent code, built from 120 caching-related pull requests in 10 open-source agent systems and applied to 571 agent projects. paper
  • LMCache - Self-hosted KV-cache layer beneath vLLM that gives token-level cache-hit observability for teams that own their GPUs. tool: Apache-2.0 last commit
  • prompt-cache - Go LLM proxy with a three-tier semantic cache that uses a cheap verification model for borderline similarity matches. tool: MIT last commit
  • Redis LangCache - Managed semantic cache with a REST API that returns a stored response when a new query is similar to a past one. tool: proprietary

Cheap Local Models

  • llama.cpp - Open-source (MIT) local LLM inference engine with a built-in OpenAI-compatible server. tool: MIT last commit
  • Magnitude - Apache-2.0 local inference engine for agents, written in Rust, that tunes GPU kernels on the user's device and shares prefix caches across concurrent sessions; its maker reports 92% faster decode than llama.cpp on an M4 Pro. tool: Apache-2.0 last commit
  • Ollama - Runtime for running open-weight models locally on hardware you already own. tool: MIT last commit

Compression

  • Amazon Bedrock Prompt Shortening - AWS documentation for the prompt shortening setting in Bedrock Advanced Prompt Optimization, which returns a shorter prompt template only when it scores at least as high as the original on your own evaluation. tool: proprietary
  • Context Mode - MCP server that sandboxes tool calls and returns only the distilled result to the model. tool: Elastic-2.0 last commit
  • headroom - Apache-2.0 context-compression tool for LLM and agent pipelines. tool: Apache-2.0 last commit
  • lean-ctx - Rust MCP server that mediates what a coding agent reads, with a savings ledger and an accounting of its own context overhead. tool: Apache-2.0 last commit
  • LLMLingua - Microsoft prompt-compression library that uses a small model to drop low-information tokens before a prompt reaches the target LLM. tool: MIT last commit
  • llmtrim - Local proxy that compresses a coding agent's prompt, tool schemas, and history before forwarding, and can reroute Claude calls to Grok. tool: MPL-2.0 last commit
  • Minification of State-in-Context Agents - ICPC 2026 study of minifying code in a coding agent's context, measuring the input-token savings against the accuracy cost. paper
  • rtk - Single-binary Rust CLI proxy that compresses the output of common dev commands before it reaches a coding agent's context window. tool: Apache-2.0 last commit
  • snip - Single-binary Go CLI proxy that filters dev-command output before it reaches a coding agent, using declarative YAML filter pipelines. tool: MIT last commit
  • TOON - Compact, human-readable, lossless serialization of the JSON data model, designed for LLM input. tool: MIT last commit

Context Engineering

  • AgentDiet - Paper on an inference-time module that strips useless, redundant, and expired information from an agent's trajectory. paper
  • Anthropic Context Management - Anthropic documentation for context editing, the memory tool, and server-side compaction in the Claude API.
  • Claude Code Compaction Engine - Blog post describing the three-tier compaction engine Claude Code uses to trim a filling context window, and its cache and correctness failure modes.
  • CliffCompaction - CMU autocompaction proxy that shrinks an agent's context only by truncating or dropping content, never by LLM rewriting, and works in front of Claude Code or Codex. tool: MIT last commit
  • Codex CLI Compaction Issue - GitHub issue reporting that a Codex CLI upgrade made context compaction fire more often and raised token consumption for identical tasks.
  • Context Rot - Chroma study of how LLM output quality changes as input length grows, with task difficulty held fixed. report
  • ContextBudget - Paper on BACM, a method where an agent decides when and how much to compress its history based on the remaining context budget. paper
  • CTXAgent - Paper and open-source agent in which the model decides when to compact and curate its own context, compared on list-price cost and task success against Claude Code, Codex and OpenCode on sessions of 35 to 212 turns with the same model behind every harness. paper
  • Cursor Dynamic Context Discovery - Cursor blog post on loading tool schemas and large outputs on demand instead of eagerly, and on Composer self-summarization.
  • Repomix - Packs an entire repository into a single AI-friendly file, reports token counts, and uses Tree-sitter to compress code to signatures. tool: MIT last commit
  • RULER - NVIDIA long-context benchmark that measures how model quality holds up as the context window fills. bench last commit
  • Self-Compacting Language Model Agents - Paper introducing SELFCOMPACT, where the model itself decides when and how to compress a growing agent trace. paper
  • Serena - Open-source MCP toolkit that gives a coding agent IDE-grade semantic code retrieval and editing. tool: GPL-3.0-or-later last commit
  • trace-mcp - MCP server that pre-indexes a repository into a symbol and dependency graph so an agent can query call graphs, change impact, or file outlines instead of reading files. tool: MIT last commit
  • Unreal Agent - MIT async agent harness whose vendor-reported SWE-Marathon run shows the same 69/160 pass count as Codex with GPT-6.1 Sol at xhigh, for $445.75 against $850.52, after adding mid-session compaction. tool: MIT

Cost Controls

Gateways and Proxies

  • Bifrost - Go-based AI gateway from Maxim AI that fronts many models with low added latency. tool: Apache-2.0 last commit
  • Cloudflare AI Gateway Spend Limits - Cloudflare AI Gateway documentation on dollar-denominated spend limits that block or reroute requests once a budget is hit. co
  • Helicone - Open-source (Apache-2.0) LLM proxy that logs every request's cost, latency, and tokens. tool: Apache-2.0
  • Kong AI Gateway - AI layer of Kong's API-gateway platform that meters LLM, agent, and MCP traffic for billing, showback, and chargeback. co
  • LiteLLM - Open-source gateway fronting 100+ LLM APIs that computes per-request dollar cost from a pricing map, with spend limits. tool: MIT last commit
  • OpenRouter Provider Selection - OpenRouter documentation on routing requests across models and providers by price, with fallback on outages. co
  • Portkey AI Gateway Budget Limits - Portkey documentation on routing LLM traffic across providers and enforcing hard USD budget limits on virtual keys. co
  • Standard Compute Budgets and Limits - Standard Compute documentation on its hosted LLM router, sold as monthly plans with a fixed compute budget, that by default returns HTTP 402 once the budget is used instead of charging overage (unless Pay as you go is enabled). co

Harness Efficiency

  • SoL-Pi - NVIDIA Labs extension for the Pi coding harness that packages four token-efficiency mechanisms. tool: MIT
  • WOZCODE - Proprietary Claude Code plugin that publishes its own paired-task dataset comparing cost and tokens against stock Claude Code. tool: proprietary

Memory

  • claude-code-memory-setup - Practitioner recipe that pairs an Obsidian memory vault with a local AST code-graph tool. tool: MIT last commit
  • claude-mem - Coding-agent memory layer that captures every session, compresses it with AI, and re-injects relevant context in later sessions. tool: Apache-2.0 last commit
  • Cognee - Open-source (Apache-2.0) AI-memory platform that gives agents persistent memory through a self-hosted knowledge graph. tool: Apache-2.0 last commit
  • Karpathy's LLM Wiki - Gist by Andrej Karpathy describing a pattern where an agent builds and maintains a persistent markdown wiki from your sources.
  • LangMem - LangChain long-term memory library that extracts and consolidates facts from conversations and integrates with LangGraph's memory store. tool: MIT last commit
  • Letta - Platform for stateful agents, from the MemGPT project, that pages an LLM's context like an operating system. tool: Apache-2.0 last commit
  • Mem0 - Open-source memory layer that extracts salient facts from conversations and retrieves only the relevant ones per call. tool: Apache-2.0 last commit
  • Supermemory - Memory and context engine for agents. tool: MIT last commit
  • Zep / Graphiti - Memory platform for agents built on temporal knowledge graphs. tool: Apache-2.0

Multi-Agent Systems

  • CrewAI - MIT-licensed multi-agent orchestration framework whose crew object reports cache-aware token totals. tool: MIT last commit
  • CrewAI Hierarchical Process - CrewAI documentation on the hierarchical process, where a manager LLM plans and validates work delegated to other agents.
  • LangChain - MIT-licensed agent framework underneath LangGraph, LangMem, and LangSmith that defines cache-aware usage metadata and ships context-editing, summarization, and call-limit middleware. tool: MIT last commit
  • LangGraph - LangChain graph orchestration library for stateful agents, with per-super-step checkpointing and an opt-in node result cache. tool: MIT last commit
  • PerspectiveGap - Benchmark of 110 orchestration scenarios that tests how a model writes sub-agent prompts and counts the out-of-role information fragments it leaks into each sub-agent's prompt. bench
  • SupervisorAgent - Paper on a lightweight, modular framework for runtime, adaptive supervision of multi-agent systems. paper
  • The Coordination Tax and the Heterogeneity Dividend - Paper with a closed-form break-even model for multi-agent LLM pipelines and a 2,430-run measurement of planner-coder-reviewer pipelines against single calls and a router on code-generation tasks, reporting dollar cost, wall time and pass@1. paper

Prompt Agent Loop

  • benjamin-plus - JetBrains skill, a small instruction payload that changes how a coding agent looks things up and waits, with published A/B results. tool: MIT last commit
  • LOOP Skill Engine - Paper on recording an agent's first run of a repetitive task with full LLM reasoning, then replaying the extracted tool-call template without calling the LLM. paper
  • Orchestrator-Worker Model Tiering - Blog post on the pattern where a capable model plans and cheaper agents execute, using Claude Code subagents.
  • token-ninja - Intercepts deterministic commands like git status or npm test before they reach the model and runs them locally. tool: MIT last commit

Retrieval Memory

  • deja-vu - MIT single-binary local memory for 43 coding agents, built from session history on disk with no LLM; its maker's benchmark (11 runs per arm) cut median task tokens from 126,222 to 53,558. tool: MIT last commit
  • LlamaIndex - MIT-licensed data framework for LLM applications with a token-counting callback and an optional hard token budget. tool: MIT last commit
  • PageIndex - Vectorless RAG engine that builds a hierarchical tree index over long documents and has an LLM reason over the index instead of embedding chunks. tool: MIT last commit
  • Signet AI - Self-hostable memory and context layer that syncs memories, transcripts, and secrets across Claude Code, Codex, OpenCode, and other harnesses. tool: Apache-2.0 last commit
  • xmemory - Paper on schema-grounded agent memory that extracts structured facts instead of storing text. tool: proprietary

Retry and Reliability

  • Claude Code v2.1.199 - Claude Code release notes covering automatic retry of rate-limit errors with backoff and partial-output preservation.

Routing Model Selection

  • Antigravity CLI Changelog - Release notes for Antigravity CLI covering per-subagent model-tier routing and an effort control. tool last commit
  • Claude Code Router - Local gateway that puts Claude Code, Codex, and other coding CLIs behind one endpoint and routes each request by rules or a prompt tag. tool: MIT last commit
  • Claude Code via LiteLLM - LiteLLM tutorial for pointing Claude Code at a LiteLLM proxy so cheaper or non-Anthropic models can handle some work.
  • Cluster, Route, Escalate - Paper proposing a two-stage cost-aware cascade for LLM serving that combines routing and escalation. paper
  • Cursor Router - Cursor blog post on Auto mode, which classifies each request and routes it to a model under Intelligence, Balance, or Cost modes. co
  • Distilling Agent Behavior - Paper on distilling a large agent's behavior into a small task-specific model to cut per-token cost.
  • EC²-Bench (EC2-Bench) - Paper and benchmark comparing query routing, per-step routing, and edge-controlled delegation or advisory between a small edge model and a cloud model on long-horizon agent tasks, scoring task performance, cloud cost, and privacy exposure together. paper
  • GitHub Copilot Auto Model Selection - GitHub documentation on Copilot's Auto setting, which routes by model health and task complexity. co
  • GitHub Copilot Fusion - Copilot feature, previewed as HydraFusion, that picks one of three multi-model workflows (Single, Cascade, Critique) per request; in preview GitHub reported 36% to 67% lower estimated cost than Claude Opus 5 at -1.5 to +4.9 points of quality, tuned on the same benchmarks. tool: proprietary
  • MTRouter - Paper on picking a different model for each turn of a multi-turn conversation to meet a cost budget. paper
  • NadirClaw - OpenAI-compatible pre-router proxy for coding harnesses that picks the cheapest model predicted to answer and escalates on failure. tool: PolyForm Noncommercial 1.0.0 last commit
  • NVIDIA NeMo Switchyard - Apache-2.0, pre-1.0 routing library that picks a model per agent request from your own pool, also hosted on OpenRouter; its cost cuts (28% to 74% from partners, about one third internally) are self-reported. tool: Apache-2.0 last commit
  • OpenAI Decisions API - OpenAI guide to the Decisions API, which returns typed answers (probability, choice, score) about text, images, or JSON, with a section on input and prompt-cache billing. tool: proprietary
  • OpenCode - Open-source (MIT) coding-agent CLI with explicit cost and model routing configuration. tool: MIT
  • OrcaRouter - Paper on a production LLM router built on a LinUCB bandit. paper
  • pi-model-router - Heuristic per-turn model router for the pi coding agent, with an OpenCode counterpart that routes by keyword and word-count rules. tool: MIT last commit
  • Plano - Envoy-based proxy whose router matches queries to user-defined domains and actions using a small routing model. tool: Apache-2.0 last commit
  • ruflo - Open-source agent meta-harness for Claude Code and Codex with swarm orchestration, persistent memory, and cost-adjusted model routing. tool: MIT last commit
  • TypeSafe Jev - TypeSafe AI blog post on a hosted decision model that answers typed questions with a choice or score and a probability.
  • vLLM Semantic Router - Open-source, self-hostable router that sends routine queries to cheap or local models and hard ones to stronger backends. tool: Apache-2.0 last commit
  • Weave Router - Drop-in proxy that picks a model for every turn of a coding-agent session and weighs the cost of switching, relicensed from Elastic License 2.0 to Apache-2.0 on 2026-09-24; its benchmark results are vendor-reported. tool: Apache-2.0 last commit

Search and Retrieval Boundary

  • Exa - Search API for agents that bills content retrieval separately per type, such as query-scoped highlights or summaries. co
  • Firecrawl - Web scraping API that converts pages to markdown or structured JSON before they reach the model. tool: AGPL-3.0 last commit
  • Parallel Benchmarks - Parallel web search and extraction API benchmarks page with accuracy-versus-cost tables. co
  • Tavily - Search API for agents that returns short content snippets instead of full pages and prices by credit. co
  • Valyu - Search and deep-research API with an open benchmark harness for cost-versus-accuracy results on the DRACO benchmark. co

Serving Inference

  • RLM-Cascade - Paper on response-level speculative decoding at the gateway, where a cheap draft model answers first and a stronger verifier accepts or rewrites it. paper
  • SGLang - Serving framework for large language and multimodal models. tool: Apache-2.0 last commit
  • TensorRT-LLM - NVIDIA open-source LLM inference engine with KV-cache reuse, in-flight batching, and prefill-decode disaggregation. tool: Apache-2.0 last commit
  • vLLM - Open-source LLM serving engine that uses PagedAttention to manage KV-cache memory in blocks. tool: Apache-2.0 last commit

Test-Time Compute

Tool Protocol Overhead

  • Code Execution with MCP - Anthropic engineering post on agents calling MCP servers by writing and executing code so unused tool schemas stay out of the context window.
  • Coral - Self-hosted MCP gateway where agents write TypeScript programs that call provider tools in a sandbox, with Allow, Ask or Block rules per tool and working programs saved as reusable custom tools. tool: Apache-2.0 last commit
  • MCP Tool Descriptions Are Smelly! - Study of how poorly written MCP tool descriptions affect agent efficiency, using an LLM-jury scanner on MCP-Universe. paper
  • StackOne Falcon - Execution engine for agent tool calls that cuts tokens by narrowing the tools an agent sees and returning only mapped response fields. co
  • Tool Attention Is All You Need - Paper on the cost of MCP re-sending every tool's full schema on every turn, known as the MCP/Tools Tax. paper

Govern

Allocation Chargeback

  • AWS Bedrock Cost Attribution in Cost Explorer - Since 2026-10-08, AWS Cost Explorer and Budgets group and filter Amazon Bedrock cost by model, provider, inference type (input, output, prompt cache read), and feature, at no extra charge. tool: proprietary
  • CloudZero - Commercial cloud and AI cost-intelligence and FinOps platform. co
  • JetBrains AI Credits for Business - JetBrains blog post on moving business AI plans from monthly per-seat licenses to reallocatable credits with a governance dashboard.
  • Mavvrik - AI and hybrid-infrastructure cost governance and FinOps platform, formerly DigitalEx. co
  • Pay-i - SDK-based GenAI cost-observability platform, now Ascerta, that tracks token-level spend per call and rolls it up into cost-center allocation. co
  • Speakeasy - Hosted AI control plane with vendor-documented cost attribution by person and role from client-reported telemetry. co

Anomaly Detection

  • Denial-of-Wallet Attacks - Paper on attacks that exploit pay-per-token pricing to inflate a bill, through stolen credentials or agents steered into runaway token use. paper
  • Governance Decay - Paper showing that compacting an agent's context can silently erase governance and safety constraints. paper

Billing Audit FinOps

  • AgentMeasure - Open-source (MIT) measurement standard and local checker that audits vendor billing claims against a named-rule settlement spec. tool: MIT last commit
  • Cloud FinOps Skill - Markdown reference library for agents, packaged as a Claude skill and an MCP server, covering billing mechanics for AI, cloud, and SaaS spend. tool: CC BY-SA 4.0 last commit
  • FinOps for AI - FinOps Foundation practitioner framework for governing AI, GPU, and token spend. tool
  • Vaudit TokenAudit - Vaudit's LLM invoice-reconciliation product, from an independent spend-auditing and recovery platform. co

Budgets Caps

  • AWS Spend Limits - AWS project-level spend limits, announced 2026-09-16 for a limited set of customers, pause a project's resources when spend reaches the limit, with optional earlier controls on the top cost drivers. tool: proprietary
  • Claude Code Review Billing - Anthropic's Code Review has no flat per-review price, bills tokens through usage credits outside plan limits, defaults new repositories to a review after every push, and offers a monthly spend limit. tool: proprietary
  • Claude Max Plan Usage - Anthropic support article describing Max plan usage as multipliers of Pro, without absolute quota numbers.
  • TrueFoundry AI Gateway Budget Limiting - TrueFoundry documentation on budget limits in its AI gateway for GenAI deployments. co

Policy Enforcement

  • ActPlane - Policy-enforcement engine, built on eBPF at the OS level, for AI-agent harnesses like Claude Code and Codex. tool: MIT last commit
  • AEGIS - Open-source (MIT) pre-execution firewall and cryptographic audit layer for AI agents. tool: MIT last commit
  • MCPGuard-Dynamic - Research-grade kernel-level eBPF sandbox for MCP, published under Meta's GitHub organization. tool: MIT last commit

Spend Management

  • ChatGPT Enterprise Spend Controls - OpenAI post on usage analytics and spend controls for ChatGPT Enterprise and Business: credit caps, request workflows, and a Cost API. tool: proprietary
  • Claude Enterprise Admin Controls - Anthropic post on admin analytics and cost controls for Claude Enterprise and Team: spend caps, model defaults, and per-user cost analytics. tool: proprietary
  • Larridin Token Spend and Insights - Enterprise AI measurement platform that reports AI adoption, all-in AI spend split by tool and by human or agent actor, and engineering output, with budget alerts and model switching at a budget threshold. co
  • nable - MCP server that reports cloud and AI spend in one answer, running locally with a free local package and some paid features. tool: Apache-2.0 last commit
  • PointFive TokenShift - PointFive's AI Efficiency OS, which governs coding-agent token spend across Claude Code, Cursor, Codex, and other tools. co
  • Revenium - Tracks AI agent spend at runtime, attributing every model call and tool cost to its workflow, with automatic shutoff on runaway budgets. co
  • Vantage - FinOps platform that ingests token-level cost data from Anthropic and OpenAI usage APIs, plus Cursor and cloud spend. co
  • Vercel AI Gateway Budgets - Vercel changelog on per-API-key dollar spend caps with daily, weekly, or monthly refresh. tool: proprietary

Unit Economics

  • Paid - Monetization platform for AI agents that meters agent events, configures pricing and billing, and links each billed action to its AI cost. co

Understand

Buyer Incentives

  • Claude Subscription Arbitrage - Zed blog post on routing agentic workloads through Pro and Max plans and Anthropic's changes to subscription terms.
  • Cursor Spend Controls Changelog - Cursor changelog entry for admin spend controls: budget caps, credit metering, and usage dashboards.
  • GitHub Copilot Billing Backlash - Discussion thread on reaction to GitHub's move from flat-rate Copilot plans to metered AI Credits.
  • Lanai - Lanai's AI @ Work platform, which discovers sanctioned and shadow AI workflows across an organization and maps token spend to the KPIs they drive. co
  • OpenAI Researcher Coding-Agent Spend - OpenAI reports its median researcher used over $600 per day of coding-agent inference at API prices by mid-August 2026; Epoch AI's fit puts the recent doubling time at about 34 days. report
  • State of FinOps - FinOps Foundation annual survey of how organizations manage cloud and AI spend. report
  • The $47k Claude Code Bill - Teardown of a viral Claude Code bill story that traces the cost to quadratic context re-ingestion.
  • Uber AI Coding Spend Cap - TechCrunch report on Uber capping per-employee AI-coding spend per tool.

Compression Efficacy

Consolidation

  • Cisco Acquires Galileo - Cisco announcement of its intent to acquire Galileo, an LLM and agent evaluation and observability platform. co
  • Tokenomics Foundation - Linux Foundation project building open standards for AI token spend. co

Market Competitors

  • Aider - Open-source (Apache-2.0) terminal coding agent with per-message dollar-cost tracking and a public polyglot leaderboard. tool: Apache-2.0 last commit
  • Amp - Coding agent spun out of Sourcegraph, with monthly tiers and credits billed at provider API prices. tool
  • Cline - Open-source (Apache-2.0) AI coding agent built on a bring-your-own-API-key cost model. tool: Apache-2.0 last commit
  • Factory - Pricing page for Factory's Droids coding agents, sold as subscription tiers with usage-based rate limits. co
  • Gemini CLI to Antigravity CLI Transition - Google Developers Blog post on transitioning from the open-source Gemini CLI to Antigravity CLI. tool
  • OpenHands - MIT-licensed open-source coding-agent platform with a free local mode, a free cloud tier, and an at-cost LLM option. tool: MIT last commit

Market Sizing

Model Economics

Pricing Models

  • AICostBudget AI API Pricing Dataset - Source-linked AI API pricing records with normalized JSON and CSV exports. data
  • Anthropic API Pricing - Anthropic pricing documentation covering how tokens are metered and priced on the Claude API.
  • Anthropic Batch Processing - Anthropic documentation on asynchronous batch processing, which trades latency for a lower price.
  • Anthropic Prompt Caching - Anthropic documentation on prompt caching, including cache-read and cache-write pricing.
  • Bessemer AI Pricing Playbook - Bessemer playbook on the shift in AI pricing from per-seat to usage and outcome-based models.
  • ChatGPT Credit Rate Card - OpenAI help article with the credit-based rate card for ChatGPT Business, Enterprise, and Edu workspaces.
  • ChatGPT Pricing - OpenAI pricing page for ChatGPT subscription tiers and Codex bundling.
  • Claude Fable 5.1 and Mythos 5.1 - Anthropic announcement with the rate card for Claude Fable 5.1 and Mythos 5.1.
  • Claude Fable 5.2 and Mythos 5.2 - Anthropic announcement of Claude Fable 5.2 and Mythos 5.2, with per-token and fast-mode rates and the data retention requirement.
  • Claude Haiku 5.5 - Anthropic model page for Claude Haiku 5.5, with release date, per-token rates split at a 100K-token prompt threshold, cache rates, and a note on token counts under the newer tokenizer.
  • Claude Opus 5.5 Overview - Anthropic documentation for Claude Opus 5.5, including pricing, cache rates, and default effort.
  • Claude Sonnet 5.5 Overview - Anthropic documentation for Claude Sonnet 5.5, including pricing and effort levels.
  • Claude Tokenizer Vocabulary - Substack analysis of how a smaller Claude tokenizer vocabulary changed token counts for plain English text.
  • Cursor Pricing - Cursor documentation on token-based metering for legacy plans, split into first-party and third-party pools.
  • Cursor Subscription Plans - Cursor documentation on its subscription plans, the weekly included-usage pool, and how Claude, GPT, and Gemini models are billed at API list price from prepaid credit.
  • DeepSeek API Pricing - DeepSeek's official pricing page listing per-token rates for its models, including cache hits and peak-window rates.
  • Devin Agent Compute Units - Devin documentation on Enterprise billing in Agent Compute Units.
  • Devin Desktop Quota - Devin documentation on token-based quota for Devin Desktop, formerly Windsurf.
  • Doubleword - Provider of asynchronous and batch inference on open models that publishes a cost-per-token table. co
  • Fable 5 Subscription Changes - Anthropic announcement on how Fable 5 is offered across subscription plans and usage credits.
  • Gemini 4 Argon Pricing - Gemini API pricing section for Gemini 4 Argon, listing rates for each service tier, scheduled rate changes, and implicit-only context caching.
  • Gemini 4 Carbon Pricing - Gemini API pricing section for Gemini 4 Carbon, listing rates for each service tier, the scheduled rate change, and implicit-only context caching.
  • Gemini API Pricing - Google's Gemini API pricing page with per-token rates and dated rate changes for Flash models.
  • Google AI Subscriptions - Google page listing Google AI subscription tiers and their Gemini and Antigravity rate limits.
  • GPT-6 Astra API Pricing - OpenAI model page for GPT-6 Astra, with per-token rates and the long-context surcharge.
  • GPT-6 Sol and Luna API Pricing - OpenAI model page for GPT-6 Sol and GPT-6 Luna, with per-token rates and the long-context surcharge.
  • GPT-6.1 Sol Pricing - OpenAI model page for GPT-6.1 Sol, with per-token input, cached-input, cache-write, and output rates and the long-context surcharge.
  • LiteLLM Custom Pricing - LiteLLM documentation on custom pricing, including flex and priority service-tier cost keys.
  • LLMflation - Andreessen Horowitz analysis of falling LLM inference prices alongside rising total AI spend.
  • Mistral Large 4 Pricing - Mistral model page for Mistral Large 4, listing release status, license, per-token input, cached-input, and output rates, and a link to the open weights.
  • OpenAI Agents API Observability - OpenAI documentation on usage and billing for hosted agent runs in the Agents API. tool: proprietary
  • OpenAI API Pricing - Pricing page listing per-token input, cached-input, cache-write, and output rates for each OpenAI API model.
  • OpenAI Deprecations - Page listing deprecated OpenAI APIs and models, including the wind-down of self-serve fine-tuning.
  • Tokenization Multiplicity and Overcharging - Paper on how the same output can be billed a different token count depending on tokenization. paper
  • Tokenomics Foundation Prompt Cache Explainer - Explainer on prompt caching that defines Cache Hit Rate and Cache Cost Efficiency and maps cache economics across access modes. tool: CC BY 4.0

Reliability SLAs

Unit Economics

  • Big-T Notation - Tokenomics Foundation vocabulary that classifies workloads by how token consumption grows with requests, calls per request, and agent depth. tool: CC BY 4.0
  • Cloud Capital Gross Margin in the Age of AI - Cloud Capital article on how inference and compute costs affect gross margin for AI-native software.
  • Cost-of-Pass - Paper defining the expected dollar cost of one correct answer as inference cost divided by success rate. paper
  • DORA AI Insights - DORA article on how AI adoption affects delivery throughput and stability. report
  • Faros AI Engineering Report - Faros report on developer telemetry and the effect of AI on engineering velocity. report
  • getDX AI Coding Assistant Pricing - Guide from getDX to AI coding assistant pricing and return on investment.
  • METR AI Productivity Study - METR randomized controlled trial of AI tool use by experienced open-source developers. paper

Measure

Benchmarks Evals

  • Claw-SWE-Bench - Benchmark showing how adapter and harness design change an agent's score on the same model backbone. bench last commit
  • Coding Benchmarks Are Misaligned with Agentic SE - Position paper from Tessl arguing that current coding benchmarks do not measure what people assume. paper
  • Deterministic Anchoring - ISSTA 2026 paper on injecting static-analysis facts as plain-text comments for code agents. paper
  • GitHub Copilot Agentic Harness Evaluation - GitHub blog post comparing Copilot's agentic-harness resolution rate with cost per task across benchmarks and models. report
  • Harness-Bench - Benchmark that holds task, model, and budget fixed while varying only the agent harness. paper
  • livenerf - Pre-registered, append-only daily benchmark that tests whether a Claude model gets worse after launch, with a locked 78-question panel, a control model, and a stated detection limit of about 7.5 accuracy points per 10-day window. bench last commit
  • LongMemEval - Benchmark for long-term memory in chat assistants, published at ICLR 2025. bench last commit
  • MemoryBench - Pluggable harness that runs memory systems head-to-head across datasets like LoCoMo. bench last commit
  • Prompt Compression in the Wild - ECIR 2026 study of end-to-end speed-ups from LLMLingua prompt compression. paper
  • promptfoo - Open-source CLI and CI harness for testing LLM prompts and agents that records per-eval token usage and cost. tool: MIT last commit
  • RedundancyBench - Benchmark for step-level redundancy detection in agent trajectories. paper
  • RouterArena - Open evaluation platform and leaderboard for LLM routers that select a model per query. bench last commit
  • SWE-bench - Software-engineering benchmark of real GitHub issue tasks from Python repositories, published at ICLR 2024. bench last commit
  • SWE-Effi - Paper that re-ranks SWE-agents on a SWE-bench subset by cost under resource constraints. paper
  • Terminal-Bench - Benchmark for AI agents in real terminal and CLI environments, with a public leaderboard. bench

Cache Accounting

  • Gemini Context Caching Pricing - Google's Gemini API pricing page, including the per-hour storage charge for explicit context caching.
  • Open-SWE-Traces Cache Simulation - Replay simulation over NVIDIA's Open-SWE-Traces that splits an agentic bill into output tokens and cache traffic.
  • OpenAI Conversation State - OpenAI guide to conversation state in the Responses API, including billing semantics and server-side compaction.
  • OpenAI Prompt Caching - OpenAI guide to prompt caching and the cached_tokens usage field.
  • The Hidden Costs of Forgetting: Trajectory Deletion in SWE Agents - Paper that runs 42 trajectory-deletion configurations on two SWE-bench agents and prices each one under flat and caching-aware pricing, showing token cuts that look cheaper under flat pricing cost more once cached prefixes are priced. paper
  • The Physical Cost of Context - Paper that measures coding-agent context policies by model calls, uncached prefill and GPU energy on self-hosted serving, and shows observation masking that cuts prompt tokens can raise uncached prefill by breaking prefix-cache reuse. paper

Cost Anatomy

  • Anthropic Pricing Line Items - Anthropic pricing documentation that enumerates the meters on a Claude API bill.
  • Claude Vision Token Pricing - Anthropic vision documentation on how images convert to billed tokens, with OpenAI and Google equivalents.
  • Cursor Token Mix - Cursor telemetry puts cache reads at 89.9% of all tokens and output at 0.6% (May 2026), and a September read finds the top 10% of users consume nearly two-thirds of tokens. report
  • TensorZero Token-Count Divergence - TensorZero blog post on how the same input produces different token counts across vendor tokenizers.

Datasets

  • FreeInference Agentic Trace - CC-BY-4.0 request-level trace of 12,002 real agent sessions from 14 harnesses over 16 weeks, with per-request token counts, cache-read tokens, and prefix hashes for cache replay, and no message text. data

Energy Carbon

  • Alumet - Rust measurement framework (EUPL-1.2 or later) that reads hardware energy counters and attributes energy to processes, cgroups, and Kubernetes pods. tool: EUPL-1.2-or-later last commit
  • claude-carbon - MIT-licensed Claude Code plugin that estimates each session's CO2 locally from token counts and shows it in the status line. tool: MIT last commit
  • CodeCarbon - Open-source (MIT) library for estimating a workload's energy use and CO2e emissions. tool: MIT last commit
  • EcoLogits - Library that estimates the energy and carbon footprint of calling generative-AI APIs, as a hosted counterpart to CodeCarbon. tool: MPL-2.0 last commit
  • Epoch AI Energy per Query - Epoch AI analysis estimating from first principles how much energy one LLM query uses.
  • Google AI Inference Environmental Impact - Google disclosure of the energy, carbon, and water footprint of a median Gemini Apps text prompt. report
  • ML.ENERGY Leaderboard - Leaderboard that measures real GPU inference energy across models and tasks. bench

Harness Overhead

  • Claude Code System Prompts - Piebald AI extraction of Claude Code's compiled prompt strings and tool descriptions for each release. data last commit
  • Claude Code vs OpenCode Token Overhead - Systima study of harness scaffolding token overhead in Claude Code and OpenCode. report
  • FrontierHarness Eval - Runta benchmark of coding harnesses on one model through one gateway, comparing pass rates and cost per completed task. bench
  • HarnessTax - Arena and Berkeley study of seven models across Claude Code, Codex CLI, and Pi on 30 SWE-bench Lite and 30 Terminal-Bench 2.0 tasks: success rates moved within a few points while cost per rollout varied up to 5.07x for the same model (3.49x in native model-harness pairs). report
  • HarnessTax Replication and Config Tax - Independent rerun of the HarnessTax protocol on 30 Terminal-Bench 2.0 tasks (540 trials, one row per trial published) that compares Codex CLI and Pi on the same model and measures what a personal CLAUDE.md, skills and rules setup adds to cost per task. report
  • System Prompts and Models of AI Tools - Collection of extracted agent system prompts and tool-schema JSON files across many vendors, published as raw text. data last commit
  • System Prompts Leaks - CC0 corpus of extracted system prompts for AI harnesses and models, with dated Anthropic prompt series. data last commit
  • TestDino Harness Overhead Benchmark - TestDino sent the same one-word prompt through 11 coding agents in one clean repo on 2026-10-08 and recorded first-call input tokens: 1,894 (Pi) to 31,480 (Claude Code), a 16.8x spread against a 31,798 Kiro CLI estimate, median of 3 runs.
  • The Harness Tax - Paper that holds model, tasks, and gateway fixed across coding-harness configurations and compares tokens per solved task. paper
  • When Instruction Files Go Stale - Paper measuring tool calls, tokens and task success when a coding agent's AGENTS.md or CLAUDE.md file holds a stale, corrected or removed line, across 732 Claude Code runs. paper

Metering

  • AgentsView and caut - Open-source tools that read local session logs to aggregate token usage and cost across coding-agent vendors. tool: MIT last commit
  • How Do AI Agents Spend Your Money? - Stanford study of token spend in agentic coding across frontier models on SWE-bench Verified tasks. paper
  • OpenCost Inference Cost Tracking - OpenCost documentation on turning GPU and shared-infrastructure spend on vLLM and llm-d deployments into cost per million tokens. tool: Apache-2.0 last commit
  • tokview - Local, zero-config proxy that shows a coding agent's token spend by session, model, and tool call. tool: MIT last commit

Transcript Analysis

Whole Bill Accounting

  • FOCUS 1.5 and AI Cost - Tokenomics Foundation post on the scope of FOCUS release 1.5 for AI cost, including what it leaves out.
  • FOCUS Specification - Linux Foundation billing-data schema specification, with invoice detail and billing period datasets for reconciling spend against invoices. tool

Related lists

Footnotes

Maintained by the team at Quesma.