The archive · AI & Models · Strategic decision · 2020–2025
Humanloop built the first LLM eval platform, then Anthropic's acqui-hire ended it in 2025
UCL-spinout Humanloop made LLM evaluation tooling for enterprises, raised $7.9M, then wound down when Anthropic hired its team in August 2025.
Humanloop
What the business is
Humanloop was a London startup whose platform let engineering teams version prompts, run human-in-the-loop evaluations, and monitor LLM-powered applications in production.
Starting capital：$7.91M seed across two rounds led by Y Combinator and Index Ventures (PitchBook, via TechCrunch).
How it started
Founded in 2020 as a University College London spin-out by Raza Habib, Peter Hayes and Jordan Burgess, Humanloop went through Y Combinator and the Fuse incubator. It started before ChatGPT made LLMs mainstream: the founders' bet was that developers would need versioning, evaluation and monitoring tooling to ship reliable AI applications, and would buy it from a dedicated platform company.
What happened
Humanloop raised $7.91M in two seed rounds led by YC and Index Ventures and won enterprise customers including Duolingo, Gusto and Vanta. It called itself 'the first development platform for LLM applications' and helped define what became LLMOps. As OpenAI, Anthropic and Google pushed agents and longer contexts at enterprises, model labs began building evaluation and monitoring directly into their platforms, and the acqui-hire wave for AI tooling teams intensified.
How it ended up
In July 2025 Humanloop told customers it was shutting down in preparation for an acquisition. On 13 August 2025 the three co-founders and roughly a dozen engineers and researchers joined Anthropic in a talent deal — Anthropic said it did not buy Humanloop's assets or IP — and the platform was sunset on 8 September 2025.
Background
Humanloop was founded in 2020 as a University College London spin-out by Raza Habib, Peter Hayes and Jordan Burgess, and went through Y Combinator and the Fuse incubator. Its bet predated ChatGPT: developers shipping LLM-powered features would need dedicated tooling — prompt versioning, human-in-the-loop evaluation and production observability — and enterprises would buy that tooling from a standalone company rather than build it or wait for model labs.
The platform won early enterprise adopters including Duolingo, Gusto and Vanta, and raised $7.91M in two seed rounds led by Y Combinator and Index Ventures. Humanloop branded itself 'the first development platform for LLM applications' and helped turn evaluation into its own discipline, positioning the tooling as an enterprise necessity on par with observability or security.
The ChatGPT boom validated the need but changed who could supply it: OpenAI, Anthropic and Google began shipping evaluation, monitoring and compliance features directly with their models and enterprise platforms, squeezing the standalone market just as a war for AI tooling talent erupted. In July 2025 Humanloop told customers it would shut down in preparation for an acquisition.
On 13 August 2025 Anthropic confirmed it had hired the three co-founders and around a dozen engineers and researchers — explicitly without acquiring Humanloop's assets or IP. Billing had stopped on 30 July, and on 8 September 2025 the platform and all customer data became permanently inaccessible, ending the startup with no product left and investors not bought out.
What has to be true
- Humanloop was early and right about the problem: before ChatGPT it built prompt versioning, evals and observability, and enterprises like Duolingo, Gusto and Vanta paid to use them.
- Its product sat directly beneath the model layer, so once OpenAI and Anthropic bundled evaluation and monitoring into their own platforms, the standalone pitch lost its moat.
- The AI talent war made team value outstrip company value: acqui-hires such as Windsurf-to-Google and Humanloop-to-Anthropic priced engineers above assets.
- Anthropic needed enterprise-trust features — evals, guardrails, bias monitoring — faster than it could build them, so hiring Humanloop's team beat waiting.
- Because Anthropic bought no assets and no terms were disclosed, investors including YC, Index, Albion, LocalGlobe and UCLTF had no clear financial exit.
What can be applied
When a category you create becomes a free feature of the platforms above you, you are a feature too: once Anthropic bundled evaluation into its stack, Humanloop's only sellable asset was its team.
Aftermath
After 8 September 2025 the Humanloop platform and all customer data are permanently inaccessible; billing had stopped on 30 July. The co-founders — Raza Habib, Peter Hayes and Jordan Burgess — and most of the engineering and research team joined Anthropic, where Habib said their evaluation workflows would support safety research and enterprise readiness. Anthropic stressed that it did not acquire Humanloop's assets or intellectual property, so the startup's investors were not bought out. Humanloop's own sunset notice remains online as the public record of the shutdown.
Sources
- Anthropic nabs Humanloop team as competition for enterprise AI talent heats up
- AI acqui-hiring spree continues with Anthropic snapping up Humanloop team
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card