Cost-economics & token-governance coverage

Which products and frameworks ship cost-control features? Per the May 2026 volumetric agent, cost / token-economics is the #2 most-asked question in agent infrastructure (171 HN mentions in the last 12 months, Sequoia's "60-70% gross margins to inference is unsustainable" framing, LangChain's State-of-Agents survey logging cost as the second-biggest blocker). The practitioner question has shifted from "how expensive is this?" to "how do I govern spend?" — this matrix pivots the catalog's seven cost-* columns into a governance-score view.

Datadog State of AI Agents 2026 found that 69% of input tokens go to system-prompt overhead — the cost-governance crisis is real, but the picture below shows which frameworks have the features practitioners can actually pull. Token budgets, prompt caching, and cost attribution are the load-bearing trio; semantic caching is still almost nowhere.

Coverage: 110/1002 products (11%) have at least one cost-economics cell filled in. Of those analyzed, 53.6% declare support for at least one of the seven tracked governance features. The rest of the catalog (mostly research papers and depth-floor rows) is surfaced as unknown by default — the matrix explicitly shows where research has bottomed out rather than papering over the gap.

Out of scope: per-call pricing data (e.g. $0.003 / 1k input tokens) — pricing changes too fast and goes stale within weeks. This view tracks the governance features that let you control spend, not the spend itself.

Per-feature support

Out of 110 analyzed products. Bars scale to the leader; numbers are absolute supportedCount.

Token budget 25 (22.7%)
Prompt caching 19 (17.3%)
Semantic caching 4 (3.6%)
Batch API 7 (6.4%)
Model routing 27 (24.5%)
Streaming-only 12 (10.9%)
Cost attribution 24 (21.8%)

Leader: Model routing with 27 supported products (24.5% of analyzed).

Legend

yes feature documented — citation in the cell tooltip
no feature absent / not exposed for this product class
unknown not yet researched, depth-floor reached, or not applicable

Matrix — showing 110 of 1002

Product	Tier	Section	Token budget	Prompt caching	Semantic caching	Batch API	Model routing	Streaming-only	Cost attribution	Score
LangGraph (orchestration) Token budget	T1	Agent frameworks (no first-party memory layer)	●	●	●	●	●	○	●	6 / 7
Anthropic Claude (foundation models) Token budget	T1	Foundation models (substrate reference)	●	●	○	●	○	●	●	5 / 7
Claude Code (Anthropic) Token budget	T1	Agent IDEs & coding harnesses	●	●	○	○	●	●	●	5 / 7
Vertex AI Memory Bank (now Agent Platform Memory Bank) Token budget	T2	Platform-provider memory	●	●	○	●	○	●	●	5 / 7
Alibaba Qwen 3 family Token budget	T1	Foundation models (substrate reference)	●	●	○	●	○	●	○	4 / 7
Anthropic Computer Use Token budget	T1	Computer-use & desktop agents	●	●	○	●	○	●	○	4 / 7
AutoGen / AG2 Token budget	T1	Agent frameworks (no first-party memory layer)	●	●	○	○	●	○	●	4 / 7
Cline (framework) Token budget	T1	Agent frameworks (no first-party memory layer)	●	●	○	○	●	○	●	4 / 7
OpenAI Agents SDK Token budget	T1	Agent frameworks (no first-party memory layer)	●	●	○	○	●	○	●	4 / 7
OpenAI Codex CLI Token budget	T1	Agent IDEs & coding harnesses	●	●	○	○	●	●	○	4 / 7
Sourcegraph Cody (agent) Token budget	T1	Agent IDEs & coding harnesses	●	○	○	○	●	●	●	4 / 7
Continue.dev Token budget	T2	Agent IDEs & coding harnesses	●	●	○	○	●	○	○	3 / 7
CrewAI Token budget	T1	Agent frameworks (no first-party memory layer)	●	○	○	○	●	○	●	3 / 7
DeepSeek R1 / V3 family Token budget	T1	Foundation models (substrate reference)	●	●	○	○	○	●	○	3 / 7
Mistral Vibe (Remote Agents) + Mistral Medium 3.5 Token budget	T1	Coding-agent memory	●	○	○	●	○	●	○	3 / 7
OpenAI Realtime API Token budget	T1	Voice agent platforms	●	●	○	○	○	●	○	3 / 7
OpenHands Token budget	T1	Agent IDEs & coding harnesses	●	●	○	○	●	○	○	3 / 7
Windsurf (now Devin Desktop) Prompt caching	T1	Agent IDEs & coding harnesses	○	●	○	○	●	○	●	3 / 7
Browser Use Token budget	T1	Computer-use & desktop agents	●	○	○	○	●	○	○	2 / 7
ElevenLabs Conversational AI (now ElevenAgents) Model routing	T1	Voice agent platforms	○	○	○	○	●	●	○	2 / 7
JetBrains AI Assistant Model routing	T1	Agent IDEs & coding harnesses	○	○	○	○	●	○	●	2 / 7
Letta / MemGPT Token budget	T1	Dedicated memory layers	●	○	○	○	●	○	○	2 / 7
agentmemory (rohitg00) Token budget	T3	Coding-agent memory	●	–	–	–	●	–	–	2 / 7
OpenAI Codex Memories + Chronicle Prompt caching	T3	Coding-agent memory	–	●	–	–	●	–	–	2 / 7
TrueFoundry AI Gateway Token budget	T3	Inference platforms & gateways	●	●	–	–	–	–	–	2 / 7
Amazon Q Developer Cost attribution	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	●	1 / 7
Arize Phoenix Cost attribution	T1	Evaluation & observability platforms	○	○	○	○	○	○	●	1 / 7
Augment Code Prompt caching	T1	Coding-agent memory	○	●	○	○	○	○	○	1 / 7
Bolt.new (StackBlitz) — harness Cost attribution	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	●	1 / 7
Braintrust Cost attribution	T1	Evaluation & observability platforms	○	○	○	○	○	○	●	1 / 7
Browserbase Cost attribution	T1	Browser-agent memory	○	○	○	○	○	○	●	1 / 7
Cognee Model routing	T1	Dedicated memory layers	○	○	○	○	●	○	○	1 / 7
Cognition Devin v2 / Spec Mode Cost attribution	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	●	1 / 7
Datadog Bits AI Cost attribution	T1	Use-case-specific agent harnesses	○	○	○	○	○	○	●	1 / 7
Devin (Cognition) Cost attribution	T1	Coding-agent memory	○	○	○	○	○	○	●	1 / 7
GraphRAG (Microsoft) Model routing	T1	Retrieval-as-memory hybrids	○	○	○	○	●	○	○	1 / 7
Lovable.dev — harness Cost attribution	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	●	1 / 7
Mem0 Model routing	T1	Dedicated memory layers	○	○	○	○	●	○	○	1 / 7
Replit Agent v3 Cost attribution	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	●	1 / 7
Salesforce Agentforce Cost attribution	T1	Use-case-specific agent harnesses	○	○	○	○	○	○	●	1 / 7
v0 (Vercel) Cost attribution	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	●	1 / 7
Zep & Graphiti Model routing	T1	Dedicated memory layers	○	○	○	○	●	○	○	1 / 7
ACE — Agentic Context Engineering Token budget	T3	Recent method papers — theorized, no distinct product	●	–	–	–	–	–	–	1 / 7
Agent Router (formerly Envoy AI Gateway) Cost attribution	T3	Inference platforms & gateways	–	–	–	–	–	–	●	1 / 7
agentgateway Semantic caching	T3	Inference platforms & gateways	–	–	●	–	–	–	–	1 / 7
Bifrost (Maxim AI) Semantic caching	T3	Inference platforms & gateways	–	–	●	–	–	–	–	1 / 7
Docker MCP Catalog, Toolkit & Gateway Model routing	T3	MCP gateways & registries	–	–	–	–	●	–	–	1 / 7
EvoMemBench Batch API	T3	Memory benchmarks & evaluation	–	–	–	●	–	–	–	1 / 7
Gemini CLI memory (save_memory / GEMINI.md) Model routing	T3	Coding-agent memory	–	–	–	–	●	–	–	1 / 7
GitHub Copilot Memory (agentic memory) Model routing	T3	Coding-agent memory	–	–	–	–	●	–	–	1 / 7
Honcho (Plastic Labs) Token budget	T3	Dedicated memory layers	●	–	–	–	–	–	–	1 / 7
Kong AI Gateway Semantic caching	T3	MCP gateways & registries	–	–	●	–	–	–	–	1 / 7
Obot Streaming-only	T3	MCP gateways & registries	–	–	–	–	–	●	–	1 / 7
OpenClaw memory Prompt caching	T3	File-backed / editor paradigms	–	●	–	–	–	–	–	1 / 7
SandBase Harness Token budget	T3	AI sandbox & runtime environments	●	–	–	–	–	–	–	1 / 7
Second Me (Mindverse) Model routing	T3	Personal AI / PKM / lifelogging memory	–	–	–	–	●	–	–	1 / 7
SimpleMem Token budget	T3	Recent method papers — theorized, no distinct product	●	–	–	–	–	–	–	1 / 7
Smithery Model routing	T3	MCP gateways & registries	–	–	–	–	●	–	–	1 / 7
Vercel AI Gateway Model routing	T3	Inference platforms & gateways	–	–	–	–	●	–	–	1 / 7
Abridge	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
AGENTS.md	T1	File-backed / editor paradigms	○	○	○	○	○	○	○	0 / 7
Amazon Neptune Analytics	T1	Knowledge-graph platforms	○	○	○	○	○	○	○	0 / 7
Anthropic Claude Memory	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
Character.ai	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
Charisma.ai	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
ChatGPT — Codex (cloud agent)	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	○	0 / 7
Claude Code auto-memory	T1	Claude Code memory mechanisms	○	○	○	○	○	○	○	0 / 7
Claude for Chrome	T1	Computer-use & desktop agents	○	○	○	○	○	○	○	0 / 7
claude-mem	T1	Claude Code memory mechanisms	○	○	○	○	○	○	○	0 / 7
CLAUDE.md	T1	File-backed / editor paradigms	○	○	○	○	○	○	○	0 / 7
Cursor Rules	T1	File-backed / editor paradigms	○	○	○	○	○	○	○	0 / 7
Decagon	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
Dia (Atlassian)	T1	Browser-agent memory	○	○	○	○	○	○	○	0 / 7
Fellou	T1	Browser-agent memory	○	○	○	○	○	○	○	0 / 7
GitHub Copilot Workspace	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	○	0 / 7
Google DeepMind Gemini Robotics	T1	Robotics foundation models & agent stacks	○	○	○	○	○	○	○	0 / 7
Google Gemini Memory	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
Granola	T1	Voice-first / wearable AI memory	○	○	○	○	○	○	○	0 / 7
Graphiti MCP Server (Zep)	T1	Claude Code memory mechanisms	○	○	○	○	○	○	○	0 / 7
Grok Memory (xAI)	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
Harvey Memory	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
Hippocratic AI Polaris	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
HubSpot Breeze AI Agents	T1	Use-case-specific agent harnesses	○	○	○	○	○	○	○	0 / 7
Hugging Face LeRobot	T1	Robotics foundation models & agent stacks	○	○	○	○	○	○	○	0 / 7
Kiro	T1	Agent IDEs & coding harnesses	○	○	○	○	○	○	○	0 / 7
LanceDB	T1	Vector-database infrastructure	○	○	○	○	○	○	○	0 / 7
MCP Servers (Reference)	T1	Agent frameworks (no first-party memory layer)	○	○	○	○	○	○	○	0 / 7
Mem0 MCP (official)	T1	Claude Code memory mechanisms	○	○	○	○	○	○	○	0 / 7
Mem0 Security / OpenMemory	T1	Memory governance, privacy & safety	○	○	○	○	○	○	○	0 / 7
Meta AI Memory	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
Microsoft Copilot Memory	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
Microsoft Edge Copilot Mode + Journeys (now Copilot in Edge)	T1	Browser-agent memory	○	○	○	○	○	○	○	0 / 7
Mistral Le Chat — Memories (now Vibe)	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
Model Context Protocol (MCP spec)	T1	Agent frameworks (no first-party memory layer)	○	○	○	○	○	○	○	0 / 7
Neo4j	T1	Knowledge-graph platforms	○	○	○	○	○	○	○	0 / 7
NVIDIA GR00T / Isaac	T1	Robotics foundation models & agent stacks	○	○	○	○	○	○	○	0 / 7
Official MCP Memory server	T1	Claude Code memory mechanisms	○	○	○	○	○	○	○	0 / 7
Omi	T1	Voice-first / wearable AI memory	○	○	○	○	○	○	○	0 / 7
OpenAI ChatGPT Memory	T1	Platform-provider memory	○	○	○	○	○	○	○	0 / 7
OpenAI Deep Research	T1	Use-case-specific agent harnesses	○	○	○	○	○	○	○	0 / 7
OpenAI Operator	T1	Computer-use & desktop agents	○	○	○	○	○	○	○	0 / 7
Perplexity Comet	T1	Browser-agent memory	○	○	○	○	○	○	○	0 / 7
Perplexity Deep Research	T1	Use-case-specific agent harnesses	○	○	○	○	○	○	○	0 / 7
Physical Intelligence (π)	T1	Robotics foundation models & agent stacks	○	○	○	○	○	○	○	0 / 7
Project Mariner	T1	Computer-use & desktop agents	○	○	○	○	○	○	○	0 / 7
Sierra	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
Suki AI	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
Supermemory	T1	Dedicated memory layers	○	○	○	○	○	○	○	0 / 7
Wayve GAIA-2 / GAIA-3	T1	Vertical / domain-specific AI memory	○	○	○	○	○	○	○	0 / 7
Windsurf Rules & Memories (now Devin Desktop)	T1	File-backed / editor paradigms	○	○	○	○	○	○	○	0 / 7