Experiential Open AI Gateway: Pricing Impact

Experiential launched on Hacker News on August 27 as an open-source model gateway that aims to do more than proxy requests. The “open OpenRouter” description is a category comparison: Experiential is an independent project that can use OpenRouter as one upstream provider. It gives agents one API for hosted, bring-your-own-key, and local models, then offers an opt-in path for turning production traces into a custom router or fine-tuned model.

This is not a new foundation model or a provider price cut. The pricing story is the gateway layer: Experiential says it adds no markup, while its optimizer tries to reduce cost per successful task by learning which model should handle each request.

What launched

The Apache-licensed project exposes OpenAI-compatible Chat Completions and Responses APIs, plus Anthropic Messages support. Its catalog covers OpenAI, Anthropic, Gemini, Azure, Bedrock, Fireworks, OpenRouter, local models, and generic OpenAI-compatible endpoints.

Teams can run the gateway locally or use Experiential’s hosted control plane. The gateway adds identities, model permissions, budgets, spend limits, usage accounting, provider waterfalls, and aliases so an application can call a stable model name while the serving route changes behind it.

The launch post claims a Rust-native data plane, broad model coverage, and low gateway-added latency. Those are project-reported claims, not independent AI Pricing Guru benchmarks.

Pricing comparison: three deployment paths

Path	Gateway charge	Who pays the model bill?	Main extra cost
Self-hosted Experiential	Open-source software	Operator through connected providers or local inference	Infrastructure, operations, and evaluation runs
Hosted Free plan with BYOK	Advertised as no-markup pass-through	Provider bills the user’s key directly	Evaluation and optimization usage
Hosted Free plan with platform credits	Advertised as public-catalog pricing with no markup	Experiential credits are drawn down	Evaluation and optimization usage

The live table and chart above compare representative upstream models from the AI Pricing Guru dataset. A no-markup gateway does not make those models free; it means the routing layer says it does not add a percentage to their catalog rate.

Use our token calculator to test your actual input, cached-input, and output mix. Current provider cards are on the OpenAI pricing page, Anthropic pricing page, and DeepSeek pricing page.

Our maintained Experiential Labs pricing page records the $0 hosted plan, 0% token markup, Apache-2.0 self-hosting license, and custom Enterprise status. Experiential adds no model row to the canonical pricing dataset or public pricing API: it is a routing and control layer whose payable rate follows the selected upstream model or local infrastructure. Encoding it as a zero-dollar model would incorrectly imply free inference.

How usage becomes a better route

Experiential imports standardized OpenTelemetry traces, mines representative tasks, and builds a simulation of the workload. It can then run candidate models against simulated tasks, apply an LLM judge, and fit a nearest-neighbor classifier over prompt embeddings. The fitted router selects a model for each new request.

After collecting more routed traces, users can also invoke a managed fine-tuning path for an open model through Tinker. That is the sense in which usage can become a “better model”: it is an optimization workflow around a user’s own traffic, not a universal checkpoint automatically improved by every gateway call.

The economic target is a better cost-quality frontier. Routine tasks can go to a cheaper model while difficult tasks escalate. But savings are not guaranteed: simulations, judging, embeddings, retries, and fine-tuning all consume resources, and poor routing can multiply calls.

Who benefits—and who loses

Agent teams with diverse, repeated workflows benefit most. They can centralize provider keys and budgets, compare local and hosted models, then learn a route from real task distributions instead of maintaining hand-written rules forever.

Teams with one stable workload may discover that a single model remains the best route. That is still useful evidence, but it weakens the case for dynamic switching. In the HN discussion, the project’s creator advised switching rarely or at task boundaries because moving between models can forfeit cached-input savings.

Security-sensitive teams also need to choose the deployment mode carefully. Local trace ingestion keeps the workflow under operator control, while uploading traces to a hosted service changes the data boundary. Experiential says its anonymous product telemetry excludes prompts, traces, model names, credentials, and raw content, and it can be disabled.

Labs decision: router and training blocker

Experiential does not enter the Cost-per-Task Labs leaderboard as one model. The current suite compares fixed model routes; relabeling an upstream result as Experiential would hide which model answered and what it cost.

A valid router study needs a pinned Experiential commit and price catalog; immutable fitting, validation, and held-out traces; independent quality labels; raw route decisions; identical provider and retry settings; token usage; latency; failures; and billed cost. Fine-tuning also needs the exact base model, training budget, adapter artifact, serving rate, and paired held-out outputs. Until those inputs exist, Labs publishes an explicit availability and reproducibility blocker rather than copying deterministic simulation evidence into the live ranking.

What to do now

Do not replace a production router on the launch claim alone. Mirror a representative trace sample, define an accepted-result metric, and compare the proposed route against your current single-model baseline.

Track total calls, cached tokens, retries, latency, judge cost, and human correction—not just the chosen model’s list rate. Set a hard optimization budget and require held-out evaluation before activation. Our AI API pricing comparison provides a current candidate shortlist; the vLLM Micro-Agent routing analysis covers a related open routing approach.

Bottom line

Experiential is a credible new open gateway option with an unusually ambitious feedback loop. Its zero-markup billing model removes one obvious gateway premium, but the real test is whether trace-trained routing lowers cost per accepted result after caching, evaluation, and operations are counted. Run it in shadow mode first, and promote only the routes that beat a fixed baseline on held-out work.

Sources: the Experiential GitHub repository, official Experiential pricing, Experiential billing documentation, Experiential gateway documentation, and the Hacker News launch discussion. Repository, license, pricing, billing language, and documentation verified August 28, 2026.