LLM observability consolidation, and what it does to your cost history

Start with the part that gets misread. Three of the six say plainly that the product continues. Helicone's post says services stay live in maintenance mode, with security updates, new models and bug fixes all still shipping. Langfuse's acquisition post says its goal continues to be building the best AI engineering platform. Traceloop says it will continue to sell the Traceloop functionality and honour its obligations to customers. Now the list, each read on 2026-08-03. ClickHouse announced on 2026-01-16 that it had acquired Langfuse. Helicone announced on 2026-03-03 that Mintlify had acquired it. Traceloop's co-founder posted in March 2026 that the company was joining ServiceNow. Humanloop's homepage is now an announcement that its team is joining Anthropic, and its own migration guide dates the sunset precisely: the platform closed on 2025-09-08 and account data was deleted that day. keywordsai.co resolves to respan.ai under a new name. Literal AI's own guide says the service remained available until 2025-10-31. Five of those six carry a date, three inside 2026 and two in 2025. Every row is the company's own statement or its acquirer's, which is the standard a claim like this has to meet. The point isn't a body count. It's that the question worth asking a vendor changed, from what the tool does to where your history lives when the logo does.

• Your cost history lives in a vendor's cloud and you have never exported it.
• You can read the last 90 days because that's your plan tier, not because you chose 90 days.
• A migration would reset your baseline, and your baseline is what every comparison is measured against.
• You picked the tool on its feature table and never asked who owns the company.
• Nobody can say what next month costs, so nobody notices when the history needed to say it goes missing.

1. Name the company that owns your cost tooling, and check whether it owned it a year ago.
2. Find your plan's data retention in writing, then say whether you chose that number.
3. Export your cost history now and open the file. Count the months that came out.
4. Ask how many months of history your forecast needs, and compare it to the count you just made.
5. Decide which of your data would survive a migration you had two weeks to complete.

Langfuse publishes 90 days of data access on Core at $29 a month and 3 years on Pro at $199, from its own feature table read on 2026-08-03. Both are fair offers. The question is which one your analysis needed, and whether anybody asked before the retention was set. Counted in whole calendar months, because whole months are what a comparison consumes.

The point isn't that 90 days is too short. It's that the number came from a pricing page rather than from the analysis you run, and a vendor changing hands is the moment you find out which one you got.

Here's the concession this page has to make. Local-first doesn't separate Culpa from this category. Eleven of the twelve tools whose feature sets were read (Helicone, Langfuse, LangSmith, Traceloop, Lunary, Braintrust, Arize, Vellum, PromptLayer, Athina, OpenLIT and Respan) publish a self-hosting option, and a team running a self-hosted Langfuse keeps its history through an acquisition exactly as a Culpa user does. The exposure belongs to managed-cloud users on a retention tier, which is most teams, because the managed option is the one that gets clicked. What Culpa adds past that point is the part none of those twelve contests. Culpa carries the revenue each customer pays you, so it reports margin rather than spend, and it forecasts next month from your own history with a range. A forecast is made of history, a migration is what shortens history, and zero of the twelve publish a forecast at all.

• LLM Retry Loop Cost: Why Failed Calls Still Bill You
• Prompt Version Cost Increase: Finding the Edit Behind It
• Unattributed LLM Spend: The Bucket That Quietly Grows
• What is a cost cross-check