Choose by the job Helicone did for you. For tracing and evaluations with cost alongside, Langfuse. For a gateway that routes requests and blocks spend over a limit, including Claude Code and Codex, Portkey or LiteLLM. For per-customer budgets enforced in a hosted proxy with a margin view, Spendline. For cost per customer and feature compared with what each customer pays, without a proxy and including self-hosted GPUs, Ganivra.
This comparison is written by Ganivra, one of the five tools, so weigh it accordingly. Every competitor fact links to that vendor's own pages, checked November 14, 2026. Prices and features change; confirm them on each vendor's site. If something here is wrong, tell us and we will correct it.
Helicone alternatives compared on cost
| Cost question | Langfuse | Portkey | LiteLLM | Spendline | Ganivra |
|---|---|---|---|---|---|
| How it connects | SDKs and OpenTelemetry, outside the request path; optional AI Gateway (beta) | Gateway in the request path; open-source gateway or hosted | Proxy in the request path; open source (MIT), self-hosted | Hosted proxy: change your base URL | Python SDK, REST events or OpenTelemetry; no proxy |
| Cost per customer | By user, tags and metadata in dashboards and the Metrics API | By _user and any metadata key | By end user, key, team and customer; the new Customer Portal adds customer organization, team and member (open source up to 5 of each) | By x-customer-id header, enforced by default | By customer, plan, feature, workflow, agent and execution |
| Margin against revenue | Not documented | Not documented | Budgets sized against what a customer paid; prices and revenue stay in your billing system, so no measured margin | Yes: monthly revenue per customer against attributed cost | Yes: revenue from plans, credits or records against AI and delivery costs |
| Claude Code and Codex | Tracing integrations and gateway setups for both | Routed through the gateway, with per-developer limits | Routed through the proxy; subscription traffic costed at list price (Claude) or $0 (ChatGPT) | Not documented for their own traffic | From their OpenTelemetry export, by developer, team and project |
| Self-hosted models and GPUs | Custom per-token prices; no GPU-hour costs | Custom per-token prices; no GPU-hour costs | Per-token or per-second prices; no GPU-hour costs | Not documented; hosted APIs only | Daily GPU bill split across calls by GPU time; idle cost and utilization |
| Budgets | Spend alerts; no blocking | Limits that block requests, with alerts | Budgets that block requests at many levels, including customer organization, team and member | Strict budgets refuse requests before the provider | Alerts by team, developer or customer; no blocking |
| Price | Free tier; Core $29/mo; Pro $199/mo; Enterprise $2,499/mo; self-host free (MIT) | Free tier; Production $49/mo; Enterprise custom; gateway free to self-host | Free and open source; Enterprise custom | No public price; 30-day free pilot | $29 to $199/mo by calls; 14-day free trial |
"Not documented" means we did not find it in the vendor's documentation, not that it is impossible to build. A proxy can block a request before it is sent; a tool outside the request path cannot, but it adds no latency and has nothing to fail open or closed.
What changed at Helicone
Helicone announced an acquisition by Mintlify in March 2026 and later said it was no longer part of Mintlify. When we checked on November 14, 2026, Helicone's homepage said it is now part of Experiential Labs, with details to come. Its pricing page still listed a free Hobby plan, Pro at $79 a month and Team at $799 a month, and its documentation describes the project as open source under Apache 2.0.
Helicone offers a proxy or AI gateway and asynchronous logging, with per-user cost from a Helicone-User-Id header and custom properties. If you are moving, list which of those you rely on: gateway features such as caching and fallbacks, request logs, or per-customer cost. That list decides which alternative fits.
Langfuse: tracing and evaluation, with cost alongside
Langfuse is an open-source LLM engineering platform for traces, prompts and evaluations. It records usage through SDKs and OpenTelemetry, outside the request path, and offers an optional AI Gateway in beta. Costs can be viewed across models, tags and users, and custom model definitions set prices per usage type, which works for self-hosted and fine-tuned models.
It documents Claude Code and Codex integrations, both as traces and through its gateway. Alerts can notify you when spend crosses a threshold. We did not find a revenue or margin comparison, budgets that stop requests, or cost per GPU-hour.
Price: Hobby is free for 50,000 units a month; Core is $29 a month and Pro $199 a month, each with 100,000 units then $8 per 100,000; Enterprise is $2,499 a month. Self-hosting is free under the MIT license, with some features behind a license key.
Good fit if Helicone was mainly your request log and you want deeper tracing, prompt management and evaluations, with cost by user as part of that.
Portkey: a gateway with budgets and coding-agent governance
Portkey is an AI gateway in the request path, hosted on edge workers or self-hosted from its open-source gateway. Its site says it is now part of Palo Alto Networks' Prisma AIRS. Cost is grouped by any metadata key, with _user powering user-level analytics. Custom models, including self-hosted ones, take your own per-token prices.
Budgets can be set in dollars or tokens and stop usage when reached, with email alerts before the limit. Its coding-agent setup routes Claude Code and Codex through the gateway, logs cost by team and project, and blocks requests at per-developer limits. We did not find a revenue or margin comparison or cost per GPU-hour. Models without a price show as $0, and budgets do not apply to them.
Price: a free tier with 10,000 recorded logs a month; Production is $49 a month with 100,000 logs; Enterprise is custom and includes granular budgets and private deployment.
Good fit if you used Helicone as a gateway and want routing, fallbacks, guardrails and hard spend limits, including for your developers' coding agents.
LiteLLM: open-source proxy with budgets everywhere
LiteLLM is an open-source (MIT) proxy you host yourself, putting many providers behind one OpenAI-compatible API. It records spend by end user, key, team and organization; grouped spend reports and request tags are marked Enterprise. Budgets apply at global, team, member, user, key and customer level and block requests once exceeded. Self-hosted models can be priced per token or per second.
It documents routing Claude Code and Codex through the proxy. On subscriptions, Claude traffic is costed at Anthropic's API list price and ChatGPT traffic at $0, so neither shows what a plan actually costs you. A margin setting adds a markup for chargebacks, which is not the same as comparing cost with revenue.
On November 13, 2026, LiteLLM announced a Customer Portal for companies selling agents built on the gateway. Each request carries the customer's organization alias and member ID as headers, and spend and budgets apply to the customer organization, team and member, with webhooks when a budget is crossed and temporary increases. Its example sets a $60 spend budget for a member who bought $100 of credits, to protect margin. Prices, payments and revenue stay in your own billing system, so it caps cost against what you expect to earn rather than measuring margin per customer. The open-source version covers up to 5 customer organizations, teams per organization and members per team; larger hierarchies and per-model budgets are Enterprise. The launch post noted that a release with the complete flow had not yet been verified, so check availability first.
Price: free and open source; Enterprise is custom-priced, with a 30-day trial.
Good fit if you want to run your own gateway, keep everything in your infrastructure, and enforce budgets per key, team or customer.
Spendline: per-customer budgets in a hosted proxy
Spendline is a hosted proxy for B2B software: you change your base URL and send x-customer-id, agent and tag headers, and it checks budgets before forwarding supported requests. Over a strict budget, a request is refused with HTTP 402 before it reaches the provider. It covers chat, messages and responses routes for about ten providers; embeddings, images and audio are not proxied.
It is the closest to Ganivra on margin: where you share monthly revenue per customer, it compares it with attributed AI cost. Hosting, support and other delivery costs are outside that figure. We did not find coverage of Claude Code or Codex traffic or of self-hosted GPU costs.
Price: no public price. It offers a 30-day free pilot on live traffic; paid terms are a platform fee plus a capped charge on managed AI spend, agreed before the pilot.
Good fit if you want hard per-customer budgets enforced in the request path and a margin view, and you are comfortable routing production traffic through a hosted proxy.
Ganivra: cost per customer against what they pay, without a proxy
Ganivra records model calls, MCP tool calls and delivery costs through its Python SDK, REST events or OpenTelemetry, outside the request path. It prices 162 models automatically and connects cost to revenue from customer plans, credit packs and top-ups or confirmed revenue records, showing what is left by customer, plan, feature, workflow and agent. It does not store prompts.
It reads Claude Code's and Codex's own OpenTelemetry export, so coding-agent cost appears by developer, team and project without changing where requests go. For models on your own GPUs, it splits each day's GPU bill across the calls served by measured GPU time and shows idle cost and utilization.
Limits: Ganivra cannot block a request; budgets alert but do not stop spend. It is not a gateway, so it does not route, cache or fall back between providers, and it is not a tracing or evaluation tool. It does not invoice customers.
Price: $29 to $199 a month for 10,000 to 300,000 calls, with a 14-day free trial.
Good fit if your question is which customers and features cost more than they pay. Moving from Helicone? The Helicone to Ganivra migration guide covers field mapping, history and cutover step by step.
How to choose a Helicone alternative
- 01Do you need to stop spend, or explain it?
Blocking a request needs a proxy: Portkey, LiteLLM or Spendline. If alerts are enough, a tool outside the request path adds no latency and no new point of failure.
- 02Is the question engineering or commercial?
Debugging prompts and quality points to Langfuse. Pricing plans and finding loss-making customers points to Spendline or Ganivra.
- 03Do coding agents count?
If your developers' Claude Code and Codex use is a large share of spend, check how each tool records it, and whether subscription traffic shows what you actually pay.
- 04Do you run models on your own GPUs?
A per-token price works for an estimate. If you want the GPU bill, including idle hours, spread across customers and features, check that the tool supports it.
Several teams run two of these together, for example a gateway for routing and limits and a separate tool for customer economics. Use the same customer IDs in both so the numbers reconcile.
Helicone alternatives: frequently asked questions
What is the best Helicone alternative for cost tracking?
It depends on what you used Helicone for. For tracing and evaluations with cost alongside, look at Langfuse. For a gateway with routing, budgets that block requests and coding-agent governance, look at Portkey or LiteLLM. For per-customer budgets enforced in a proxy, look at Spendline (hosted) or LiteLLM's Customer Portal (self-hosted). For cost per customer compared with what each customer pays, without a proxy, look at Ganivra.
Is Helicone still maintained?
Helicone's homepage said it is now part of Experiential Labs, with details to come, when we checked on November 14, 2026. Earlier in 2026 it announced an acquisition by Mintlify and later said it was no longer part of Mintlify. Check Helicone's own site for the current status before deciding.
Which Helicone alternatives work without a proxy?
Langfuse records usage through its SDKs and OpenTelemetry, outside the request path, and offers an optional gateway. Ganivra uses its Python SDK, REST events or OpenTelemetry and has no proxy. Portkey, LiteLLM and Spendline sit in the request path, which is what lets them block requests over a budget.
Which alternatives track Claude Code and Codex costs?
Portkey and LiteLLM document routing Claude Code and Codex through their gateways. Langfuse documents tracing integrations and gateway setups for both. Ganivra reads Claude Code's and Codex's own OpenTelemetry export, by developer, team and project, without changing where requests go. Spendline did not document coding-agent traffic when we checked.
Can these tools show margin per customer?
Of the five, Spendline and Ganivra compare AI cost with customer revenue. LiteLLM's Customer Portal, announced in November 2026, sets spend budgets per customer organization, team and member sized against what the customer paid, but prices and revenue stay in your billing system, so it caps cost rather than reporting margin. Langfuse and Portkey break cost down by user or customer without a revenue comparison.
See your own numbers
Find which customers cost more than they pay.
Start a 14-day Ganivra trial, connect one workflow, and compare each customer's AI cost with their revenue. No proxy and no credit card required.