Signing up grants $1, once per account. Older entries below describe grants that applied at the time; Pricing and credits is the current source.
DeepSeek V4 Pro is public
DeepSeek V4 Pro (deepseek-v4-pro) is available through OpenAI-compatible chat completions. Use the model id in the same POST /v1/chat/completions request shape as the other public hosted models. GET /v1/models remains the authoritative live list.APIBillingCreditsModel APIs
Idempotency-Key covers streaming, credit top-ups that cannot shrink or double charge, and a named milestone when you cross one
Idempotency-Key now covers streaming
A dropped stream retried with the sameIdempotency-Key no longer generates or charges twice. Streaming chat completions used to ignore the header. They now register the key before dispatch, so a duplicate arriving while the stream is open gets 409 idempotency_conflict and starts no second generation, and a duplicate arriving after the original settled gets the terminal usage and cost as JSON with X-RunInfra-Idempotent-Replay: true. We do not store the tokens we deliver, so the text itself is never replayed. A streaming key is held for the response deadline plus a short settlement grace, 270 seconds at the outside, then clears on its own.This changes behavior for a client that sends one constant key. See the warning at the top of this page.Credit top-ups
Promotion codes no longer apply to credit top-ups. Credits are stored value, so a discount code reduced the credits you received by exactly what it took off the price: it appeared to work and bought you nothing. The amount you select is now the amount you are charged and the amount that lands on your balance.A reload cannot charge you twice. A top-up carries a durable identifier for that attempt, so a refresh, a back button, a second submit, or a return from a bank verification screen resumes the same payment instead of starting another.A confirmation that does not settle is reported as unknown, never as declined. When confirmation does not finish inside its bound, the screen says the result is not yet known and waits for the credits to land, rather than reporting a failure on a payment that may have succeeded.A captured payment now ends in credits or a retry. A top-up settles only when the amount quoted, the amount the payment provider recorded, and the amount actually captured all agree in US dollars.Crossing a spend milestone says so
The purchase that moves you up a capability milestone names the milestone you reached and the one you left. It reads the same ladder that enforces your limits, so it cannot congratulate you on capacity your workspace has not been granted, and it only moves forward.Model pages state how a number was measured
A model’s throughput figure now names its measurement path in the unit, reading output tokens per second, end to end when measured through the API you call, and output tokens per second, model only when measured at the model itself. The two are different measurements, and printing them under one unit let a model-only figure read as the speed a request receives.Model APIsModelsAPI
Model APIs go live: hosted models behind one OpenAI-compatible endpoint, a second model, and reasoning-model budgets
Model APIs are live
RunInfra now hosts and serves models behind one OpenAI-compatible API. Point an OpenAI client athttps://api.runinfra.ai/v1, create a workspace key, and call the model id. No GPU to provision and nothing to deploy, priced per million tokens at the rates each model’s page publishes. Start at the Model APIs quickstart.The Model Library is the public catalog. Every hosted model’s page publishes its context window as served by our own deployment, its prices, its capabilities, and its measured performance with the conditions of the measurement attached.Nemotron 3.5 Lightning 30B (nemotron-3-5-lightning-30b) joins DeepSeek V4 Flash (deepseek-v4-flash), listed ahead of its public opening with its served facts published. A paused model answers 503 hosted_model_paused with its scheduled return time and stays listed in GET /v1/models, so nothing about your integration changes when it opens.Reasoning models budget correctly out of the box
Reasoning models spend billed output tokens thinking before they answer, so an under-budgeted request can return empty content. Each reasoning model’s page now publishes its recommended minimummax_tokens and bakes it into the code examples it generates. See Reasoning models.Generated kits say where every value came from
Every value in a generatedconfig/vllm-config.yaml carries an inline source label, and the header counts measured values, derived values, recorded configuration, defaults, and values whose provenance is unavailable. GPU memory utilization now has one answer across the config, serve script, Dockerfile, Compose file, and Kubernetes manifest, and when no provenance-backed value exists the kit omits the number and says the runtime will choose its default rather than fabricating one.OptimizationQualityDeploymentsModels
Comparable before and after, lossless-first optimization, named deploy failure causes, and model access refused at accept
A run ends with a comparable before and after
A speedup is published only when both sides were measured the same way. Every measured result records its role, the instrument that produced its latency, the unit of its throughput, and the model and modality it belongs to, and two results are compared only when all four agree. When they do not, no winner is crowned and the verdict names the axis that differed.The baseline is measured in the run, on the same instrument as the optimized results and before any technique is applied. It is never synthesized from a catalog figure or an earlier run.Every completed run records one quality verdict, and there are exactly three: measured, no quality risk when nothing output-affecting was applied, or not measured with the reason stated. Where several stages carry gates, the lowest score decides. See Run outcomes.Lossless first
Quantization is no longer applied by default. It runs only when you ask for it, when base-precision weights do not fit the resolved GPU, when the target is local hardware, or when the declared output is a constrained on-device format. The plan states the decision either way. See Lossless first.Deploy failures name a real cause
A worker that never starts now says why, from its own log. RunInfra captures a bounded, redacted tail and classifies image and registry failures, storage exhaustion during startup, GPU driver and runtime incompatibility, entrypoint crashes, and out-of-memory kills, each with the action that resolves it. When the log evidences none of them the message stays generic and retryable instead of guessing.Model access is refused at accept
A model your workspace provably cannot read is refused when you accept the plan, before a GPU starts and before billable work begins, instead of failing partway into a paid run. Each refusal names one concrete next action. Hugging Face timeouts and server errors stay retryable and are never treated as access refusals. See Model access.Failures are itemized. When a run step fails, the result lists each failed step with its reason, bounded and redacted so long logs and secrets never flood the card.Hardware is disclosed. If a run could not get the GPU you requested, the result says which GPU it requested and which it actually ran on, and comparisons label these plainly as run hardware.Older numbers are dated. A measurement carried over from an earlier run is stamped with when it was measured, so a fresh-looking number is never silently stale.The daily billing digest is opt-in. Nobody receives it unless they turn it on at Settings > Notifications. Product notifications such as low-credit and budget alerts stay on by default and can be switched off. Security and account emails are always delivered.
OptimizationBilling
Runs settle to delivered evidence: a downgraded phase reduces the charge instead of failing the run
Runs settle to what they proved. Every phase in an accepted plan declares the outputs it must produce, for example measured latency and memory for a benchmark phase, or a quality score for a quantization phase. A declared phase that finishes without those outputs no longer fails the run and no longer charges as if the evidence existed: the run completes with that phase marked downgraded.Downgrades reduce the charge, removing the portion tied to each downgraded technique at settlement, so the final charge reflects delivered evidence rather than the plan. They are also visible and gating: the result lists each downgraded phase with its missing outputs, and a run with downgrades is not marked deployable.
BillingOptimization
Large-model optimization pricing, advisory capacity checks, and settlement only on durable completion
Large-model optimization quotes. New quotes keep the existing size pricing through 70B parameters, then rise more steeply for larger models. Work already admitted keeps the price basis recorded when it started.Capacity at intake. RunInfra warns before paid consent when a model exceeds managed capacity. You can still optimize and export it for your own infrastructure. The check is advisory, so an unavailable capacity assessment never blocks pricing or consent.Charges settle only on durable completion. An optimization run’s final charge settles only when the run has durably finished. A run interrupted mid-flight keeps its usage record pending until it reaches a real terminal state, instead of settling early on an in-flight snapshot.
Gateway key env name. The dashboard reveal panel and generated snippets suggest storing the managed gateway key as
RUNINFRA_GATEWAY_KEY. A self-hosted export kit keeps RUNINFRA_API_KEY for its own self-minted secret, so the two never share a name. See Authentication.Persistent worker sizing. Worker counts requested at deploy time persist across restart, start, and GPU changes. An explicit request value wins, then the deployment’s original sizing, then the workspace default, capped at 32 per deployment.BillingAccountSecurity
One honest credit balance, self-serve account deletion, and a hardened sign-in
One balance, everywhere. The navbar, billing page, and session views read the same live available balance, updating after credit purchases and settled usage, so every surface reports the same amount.The full session money story. Before a run starts you see the estimated charge and available balance. During the run, one live estimated-charge figure. At the end, actual settled usage. Failed or cancelled runs cost nothing, and agent chat messages settle to actual usage.Signup credits are visible from day one. The grant and every transaction against it appear on Settings > Cost, and the live balance shows in the top-bar credits chip on every page.Self-serve account deletion. Settings > Workspace has a danger zone. The dialog previews exactly what will be removed, lists blockers such as live deployments and running jobs with inline actions to clear them, and requires a typed confirmation plus your password.Sign-in hardening. Passwords require 8 characters minimum. Email confirmation and password-reset links work when opened in a different browser, and expired links say so with a resend path.
OptimizationRunbooksReliability
Measured comparisons, durable runbook executions with live re-attach, and plan-edit honesty
Measured comparisons. The compare models, compare engines, and compare techniques chat modes run real benchmarks instead of advisory summaries. Compare 2 to 5 models sequentially on the same GPU target with the same serving configuration, each with its own measured baseline, or benchmark one model across serving backends. A winner is declared only from measured rows on matching hardware.Durable runbook executions. Executions survive interruption. A refresh, timeout, crash, or redeploy converges the run to a blocked, retryable state with resume and restart actions instead of stranding it at running. Reloading mid-execution re-attaches within about a second, restoring phases, live cost, and the Stop control. Cancel stops the underlying GPU work, and a cancelled run can never promote a version afterward.Plan-edit honesty. Saving a plan edit reports whether execution actually changed, and edits that would corrupt the executable graph are rejected with an actionable message at save and at accept. Agent GPU and cost lines are validated against feasibility data when a snapshot exists, and labeled as estimates otherwise.
Malformed Hugging Face model ids are rejected before a runtime starts, and RunInfra checks whether the resolved repository is reachable, blocking missing, gated, invalid-token, and access-denied models before starting GPU work. Valid raw
owner/repo ids resolve even when they are not in the static catalog, while proprietary or malformed ids fail clearly.Responses adapter. A dedicated
/v1/responses reference for the Responses-shaped chat-completions adapter, including streaming, instructions, response_format, and supported tool pass-through fields.Error reference. The OpenAPI spec and error guide now cover the public gateway statuses developers should handle: rate limits, credit exhaustion, idempotency conflicts, replay-unavailable responses, upstream failures, and gateway timeouts. See Errors.Latency clarification. A prompt that mentions latency without a target now asks whether to use a specific budget or optimize for the cheapest compatible GPU first.Cheapest-compatible GPU search. Cost-first prompts prioritize GPUs that fit the model and constraints before spending time on higher-cost options.Serving-first runbooks. LLM optimization tunes serving settings earlier, so batch size, prefix cache, and GPU utilization become the baseline for later quantization and kernel work.Plan transparency. Runbooks show open questions, quality gates, estimate assumptions, and duration ranges directly in the plan.
Measured metrics. Optimization surfaces highlight measured latency, throughput, cost, and quality from real runs, and quantization comparisons use real model-output checks against a baseline.Traceable requests. Chat, deploy, inference, and optimization requests carry request and workspace trace headers so support can correlate issues quickly.Privacy controls. Analytics respect Do Not Track, avoid client IP collection, clear identity on signout, and scrub secrets from captured URLs.
PlatformAPIDeployments
Instant Start weight caching, workspace-scoped keys, and additional OpenAI-compatible routes
Instant Start. RunInfra’s weight-caching layer reduces cold starts for compatible scale-to-zero deployments.Workspace-scoped keys. A workspace key reaches every verified deployment, selecting the target with
model in the request body, or discovering what is available with GET /v1/models.More routes. Deployed embedding models answer POST /v1/embeddings, and speech-to-text and text-to-speech deployments expose /v1/audio/transcriptions and /v1/audio/speech. Pipelines can target vLLM, SGLang, TensorRT-LLM, or vLLM Omni where the model and modality support it.RunInfra launched with chat-driven pipeline building, optimization, and deployment: describe a use case in plain English, and RunInfra selects models, optimizes the pipeline, and deploys an OpenAI-compatible endpoint. Benchmark GPU choices, quantization options, serving settings, and kernel optimizations against latency, throughput, cost, and quality goals, then ship with managed GPU hosting, autoscaling, and monitoring.