Galileo AI Evaluate

Cisco announced intent to acquire Galileo Technologies, Inc. on April 9, 2026 (https://www.networkworld.com/article/4156855/cisco-to-acquire-galileo-for-ai-observability.html), and the deal closed within Cisco's Q4 FY2026 — Splunk's own May 28, 2026 blog post already says 'Cisco has completed its acquisition of Galileo' (https://www.splunk.com/en_us/blog/leadership/splunk-observability-galileo.html), and Galileo's GitHub org is now literally titled 'Galileo (now part of Cisco)' (https://api.github.com/orgs/rungalileo, checked 2026-08-27). Terms were not disclosed (https://www.crn.com/news/networking/2026/cisco-to-snap-up-ai-observability-startup-galileo-technologies-to-expand-splunk-platform). As of August 7, 2026, Galileo's own release notes state the product 'is now Splunk Agent Observability' for anyone onboarding after that date, with new documentation moved to agent-observability-docs.splunk.com (https://docs.galileo.ai/release-notes). The galileo.ai marketing site, self-serve signup, and free tier are all still live and independently operable as of this check (verified by loading app.galileo.ai/sign-up directly), but a buyer should assume the product will keep converging into the Splunk/Cisco Observability Cloud line rather than continuing as a standalone roadmap.

Galileo's core technical bet is running evaluation with small, purpose-distilled 'Luna' models instead of a frontier LLM-as-judge for every trace. This isn't just a marketing claim — Galileo's engineering team published it as a peer-reviewable arXiv paper (Luna-2: Scalable Single-Token Evaluation with Small Language Models, Feb 2026, arXiv:2602.18583), which reports over 80x lower inference cost and over 20x lower latency than frontier-LLM judges at comparable accuracy on content-safety and hallucination benchmarks, and claims (self-reported, not independently audited) that Luna-2 is protecting 100M+ AI sessions and >100B tokens/month in production with $30M+ in annual eval-cost savings for customers. Whatever one makes of the self-reported production numbers, the architecture itself — LoRA/PEFT heads on a shared SLM backbone, enabling many concurrent specialized metrics on one GPU — is a real, technically legible answer to the 'evaluation is too expensive to run on everything' problem that most LLM-as-judge tools don't solve. Galileo also publishes and maintains a public, still-updated Agent Leaderboard on Hugging Face ranking LLMs on agentic tasks (453 likes, https://huggingface.co/spaces/galileo-ai/agent-leaderboard) — an artifact independent of the vendor pricing pitch that a buyer can inspect directly.

The pricing page (https://galileo.ai/pricing, checked 2026-08-27) publishes concrete numbers for two of three tiers: Free ($0/month, 5,000 traces/month, unlimited users and custom evals) and Pro ($100/month billed yearly, 50,000 traces/month, standard RBAC, advanced analytics, Slack support) — but the Pro card is explicitly annotated '*Pricing scales based on number of traces,' with no published rate card for what a trace costs above 50,000/month. Enterprise is 'Contact us' with no published floor. A buyer evaluating this for meaningful production volume (the whole pitch is evaluating 100% of traffic) needs a sales call to find the real cost curve — the headline $100/month tells you almost nothing about what running it on real production traffic costs.

Galileo Technologies was founded in 2021 by Vikram Chatterji, Yash Sheth, and Atindriyo Sanyal — alumni of Google AI (BERT), Google speech recognition, and Uber AI/Apple Siri respectively (https://galileo.ai/about). It raised a $45M Series B in October 2024 led by Scale Venture Partners with Databricks Ventures and ServiceNow Ventures participating, bringing total funding to $68.1M per Crunchbase (https://www.crn.com/news/networking/2026/cisco-to-snap-up-ai-observability-startup-galileo-technologies-to-expand-splunk-platform; corroborated at https://www.forbes.com/sites/richardnieva/2024/10/15/galileo-series-b-45-million-scale-venture-partners/). Customers named in vendor-published case studies include Comcast, HP, NTT, and Twilio, per Cisco/CRN reporting and Galileo's own case studies page — these are vendor claims, not independently verified, and should be read as such (https://galileo.ai/case-studies).