Helicone

Helicone's pitch is one control plane instead of a stack of half-connected tools, and for engineering teams that already route across multiple providers, that pitch mostly holds up. The gateway side is where the economics live: passthrough billing at 0% markup means you pay the provider's rate and nothing on top, which matters more the larger your token spend gets. Caching and automatic fallbacks cut both latency and spend without you writing provider-specific retry logic, and rate limiting is enforced at the gateway rather than scattered across your services. The observability side is deeper than a simple request log. You get sessions and users tracking, so you can reconstruct a full user journey instead of staring at isolated requests. The Playground supports reasoning effort control — minimal, low, medium, high — with a visual thinking display for reasoning models, which is a genuinely useful debugging aid when a model's answer goes sideways for non-obvious reasons. Datasets let you version prompt data and run regression checks, HQL gives you SQL-like queries over your logs, and Prompt Management V2 supports typed variables inside system prompts, messages, and tool schemas, with version control and rollback that don't require a code deploy. Country-based filtering, added in July 2025, uses the Cloudflare edge server receiving the request, so geographic analytics come for free if you're already on the gateway. Where Helicone trails is deep prompt evaluation and experiment tooling — that's LangSmith and Braintrust territory, and Helicone's own comparison table concedes limited evaluation and experiments next to those products. If your bottleneck is systematic A/B testing of prompts and scoring outputs, you'll likely pair Helicone with something else. Single-provider shops also won't get much from the multi-model routing, though the logging and cost tracking still apply. The scale claims are unusually concrete: 14.2 trillion tokens processed across three years, and 1M context window support for Claude Sonnet 4 and 4.5 across Anthropic, Bedrock, and Vertex. In September 2025 the AI Gateway moved onto the cloud with passthrough billing from the dashboard and an OpenAI-compatible API to any model. The open question is roadmap: Helicone was acquired by Mintlify in March 2026, and while the platform still ships — changelog entries through November 2025 and blog posts into 2026 — buyers making a multi-year commitment should factor in that ownership change. Non-profits, students, open-source projects, and sub-$5M-funded startups under two years old get discounts, so check the discount page before paying list.