RightNow wants you to run open weights at the speed frontier labs squeeze out of their own GPUs. On November 17, 2026, Y Combinator’s Launch YC post introduced RightNow as a GPU kernel lab now selling hosted open-model inference through a single API key, with OpenAI- and Anthropic-compatible endpoints and published per-token prices.
That is a different bet from renting the same stock vLLM stacks every neocloud lists. RightNow claims it writes kernels per model, cites Revolut moving 2,000 engineers to GLM 5.3 on dedicated hardware, and invites teams spending $1k+/month to port a heavy workload in one day. Whether those economics apply to your traffic needs measurement — but the launch is a real routing option in the open-model wave alongside Together prefix caching and Hugging Face plus OpenRouter.
Own your AI at full speed: what RightNow actually ships
The company homepage and open-models docs describe a single workspace key (rp_...) that reaches every hosted model through:
| Connection detail | Value |
|---|---|
| API base | https://api.rightnowai.co/v1 |
| Chat completions | POST /v1/chat/completions |
| Messages API | POST /v1/messages |
| Auth | Workspace key in RIGHTNOW_API_KEY |
RightNow open model inference capacity tiers on one API base URL
Image: RightNow AI
Open model inference here means open weights on RightNow metal, not downloading weights into your VPC. You still own the model choice — GLM, DeepSeek, Qwen, Nemotron, and others listed on the pricing page — while RightNow owns uptime, kernels, and billing.
The YC launch narrative contrasts closed labs (“rent your intelligence”) with open models that are good enough if someone makes them fast enough. That matches teams already chasing Qwen Code and Baseten’s efficient frontier guide, but adds vertically integrated kernel work as the moat.
Published pricing and cache economics
RightNow posts USD per million tokens with live cache-hit rates — unusual transparency for a brand-new host. Snapshot from its pricing table on November 17:
| Model | Input | Cached input | Output | Cache hit (listed) |
|---|---|---|---|---|
| Nemotron 3.5 Lightning 30B | $0.05 | $0.01 | $0.15 | 95.0% |
| DeepSeek V4.1 Flash | $0.14 | $0.03 | $0.58 | 99.6% |
| GLM 5.3 Flash | $0.11 | $0.03 | $0.45 | 98.8% |
| GLM 5.3 | $1.30 | $0.13 | $4.20 | 98.4% |
| Qwen3.8 27B | $0.10 | $0.01 | $0.40 | 93.0% |
Prompt caching is automatic — no special header in the docs — which matters for agent loops that resend the same system prompt. Compare with explicit caching products like Together’s prefix cache when you model long-horizon coding agents.
Capacity modes on the pricing page:
- Pay as you go on shared capacity (available now).
- Reserved capacity and dedicated deployments marked not offered yet.
- Batch jobs below standard rates on idle capacity.
Limits scale with spend: 10,000 requests/minute per key at signup, 50,000 after first purchase, 200,000 after $200 in top-ups, with optional monthly spend caps per key.
How to try RightNow with existing SDKs
RightNow’s docs target drop-in clients. Conceptually:
from openai import OpenAI
client = OpenAI(
api_key="rp_your_workspace_key",
base_url="https://api.rightnowai.co/v1",
)
response = client.chat.completions.create(
model="glm-5.3-flash",
messages=[{"role": "user", "content": "Summarize KV-cache wins in one paragraph."}],
)
print(response.choices[0].message.content)
Swap the base URL and key; keep your application code. For Anthropic-style agents, point the Messages client at the same host per open-models docs.
Before production traffic:
- Run a side-by-side latency and quality check against your current host on identical prompts.
- Confirm model slug strings on the pricing page match what the API accepts — hosts sometimes use internal IDs.
- Read data retention defaults; RightNow claims zero retention unless you opt in elsewhere in the contract flow.
Kernel lab backstory: why YC cares now
GPU stack tuning from custom kernels rather than stock serving software
RightNow’s YC company profile lists founders Jaber Jaber and Osama Jaber, batch Fall 2026, team size 2, based in San Francisco. The launch post says frontier labs already use their optimization tools on private clusters; the API product productizes the same stack for everyone else.
That story rhymes with NVIDIA’s developer-kit pushes and the wider “picks and shovels” selloff — capital markets freak out while serving costs still dominate unit economics for coding agents. A kernel lab selling open weights is betting builders will trade closed-model prestige for bill size, especially after price shocks on Claude usage limits and DeepSeek free-tier caps.
The Revolut anecdote in the launch post — ~90% coding spend reduction moving 2,000 engineers to GLM 5.3 — is marketing, not audited FinOps. Treat it as a benchmark to replicate, not a guarantee.
Open models vs closed models: where RightNow fits
| Axis | Closed frontier API | RightNow-style open host |
|---|---|---|
| Weights | Proprietary | Open-weight families on card |
| Differentiation | Model quality and tools | Price, speed, caching, residency |
| Switching cost | Prompt/tooling lock-in | Lower if clients stay SDK-compatible |
| Risk | Policy and safety changes | Host reliability and kernel bugs |
RightNow does not replace GPT-6 Astra healthcare pilots or Claude Opus PCB agents where frontier reasoning wins. It competes on high-volume, price-sensitive paths — batch evals, internal codegen, routing previews — where open weights already pass bars.
Side-by-side with other open-model hosts
Builders evaluating open model inference providers should compare list price, measured tokens/sec, and operational maturity:
| Provider shape | What you get | Trade-off |
|---|---|---|
| RightNow | Kernel-tuned short list, automatic caching, dual protocol | Young company; reserved/dedicated not GA |
| OpenRouter + HF | Huge catalog, fast routing experiments | Less visibility into kernel stack per model |
| Together | Strong agent/coding marketing, prefix cache story | Different model mix and enterprise contracts |
| Self-host vLLM | Maximum control | You own kernels, capacity planning, and on-call |
Our Baseten efficient frontier guide walks the self-host vs managed decision tree. RightNow is a bet that managed + kernel lab beats generic managed for specific open families — test with your prompts, not vendor tok/s rows alone.
Published per-million-token rates beside measured inference throughput
Teams routing Chinese open models should also read Alibaba Qwen4 Max guardrails and export-control context before moving regulated workloads.
Billing, limits, and production hygiene
RightNow’s pricing page encodes FinOps-friendly knobs:
- Per-key monthly spend caps — set before you point a coding agent at the API.
- Auto-recharge with $5 minimum top-ups and $1M maximum — useful for bursty agent swarms if finance approves.
- Rate limits that scale with purchases — plan for throttling during first-week experiments.
Mirror the discipline we recommend for Nasiko agent gateways: proxy once, attribute spend per team, and stop runs at budget — even if RightNow is only the model backend.
For learning paths on switching hosts without rewriting agents, browse /workshops and the open generative AI directory.
Security and compliance checklist
Before sending customer data through any new inference host:
- Download the DPA and confirm retention matches marketing (zero by default on RightNow’s site).
- Log model version strings in your telemetry — open hosts rev weights without a frontier-style launch blog.
- Run red-team prompts you already use on Claude/GPT; open models change refusal behavior.
- Keep a fallback route via OpenRouter if the host errors during an incident.
Enterprise motions: reserved, dedicated, and custom kernels
The YC launch post markets three upsell paths:
- Reserved capacity — quoted per concurrent request hour on select models.
- Dedicated deployments — single-tenant serving (waitlist).
- Custom kernels — bring your own checkpoint; RightNow optimizes and hosts.
That is the same ladder neoclouds sell, but RightNow anchors it in kernel authorship rather than bare metal rentals. Until reserved tiers ship, enterprises should assume noisy-neighbor risk on pay-as-you-go and negotiate SLOs in writing.
What developers are saying (Hacker News)
The Launch YC thread had few comments at publish time on November 17. No substantive hands-on numbers appeared yet. When builders post latency benchmarks, we will fold them into an update section — do not treat launch-post claims as independent verification.
Risks and limits to watch
- New provider risk: uptime, support, and financial backing are unproven at megascale.
- Model coverage: only listed cards are kernel-tuned; arbitrary Hugging Face uploads are bring-your-own-model sales conversations.
- Reserved/dedicated tiers are not GA; heavy production may still need pay-as-you-go until those tiers ship.
- Compliance: zero-retention marketing must match your enterprise DPA if you handle regulated data.
When RightNow is the wrong default
Skip RightNow as primary if you need frontier multi-modal tools (computer use, deep research connectors), regulated industry attestations only frontier labs publish, or single-vendor procurement with an existing Azure or BedrockEA deal. In those cases, use RightNow as a cost burn-down lane for well-scoped open-model tasks — summarization, log triage, translation, bulk classification — where quality plateaus below the largest closed models.
The Launch YC post’s Revolut story is instructive: they allegedly moved coding workloads, not every customer-facing assistant. Copy the pattern, not the headline percentage.
Practical next step: create a shadow API key, replay one day of production prompts on a single open model RightNow lists, and compare cost, latency p95, and edit distance against your baseline. Ship only after finance and security sign off on the delta.
