RightNow System Zero: AI Decisions at $0.00028 a Call

Decision models have a billing problem that nobody talks about. Every serious use case, a robot adjusting its grip, a phone agent deciding whether to escalate, a fraud check on a live video feed, runs thousands of tiny judgments a day, and the classic per-token pricing model bills each one like a small essay. RightNow, a Y Combinator Fall 2026 startup, thinks the unit should be the decision itself. On October 30 it launched System Zero, a closed-source API that streams decisions with per-option probabilities while the input is still arriving, and charges you for each answered question. The headline number is USD 0.00028 for a simple call. Here is what that actually buys, what the fine print says, and how the pricing compares to the rest of the decision-model field.

What System Zero actually is

System Zero is not a chatbot, and it is not trying to be one. It is a purpose-built decision endpoint: you send it some state, a JSON object, a piece of text, a PDF, an image, a voice clip or even robot sensor data, and you attach named questions like "damaged": {"type": "boolean", "instructions": "Did the item arrive damaged?"}. Back comes a structured answer for each question, with a probability on every option, a chosen decision and a confidence value. There are three question types: boolean, choice and score.

The product's central pitch is streaming. Most decision models, in the founder's own words posted on X at launch, "read text, wait until the whole input has arrived, and then give you an answer." That breaks down for inputs that never finish: a robot, a phone call, a video feed. System Zero returns decisions as the input arrives, and the live homepage shows a recorded demo where the refund-or-replace judgment updates as more text and images stream in. The company backs the claim with a promised technical paper on what it calls Reflex Training and Reflex Architecture. That paper is not published yet, so for now the streaming mechanism is a description, not a verifiable architecture.

The inputs, per the official docs, are text, JSON state (including robot state), codebases, PDFs, images, voice and video, with video explicitly in preview. Supported file formats are PNG, JPEG, WAV, PDF and MP4. Standard text context is 38,000 tokens, and a 1M-token window can be enabled in console settings, though the docs note that long-text state only reaches 131,072 tokens when that band is served. The API is REST at api.rightnowai.co, with TypeScript and Python SDKs, idempotency keys, workspaces, and a playground for trying calls without code.

What it costs

RightNow's pricing page, updated October 29, 2026 under "price card G" (October 27, 2026), charges three ways at once: for the tokens it reads, for each answered question, and a small per-call fee. Output is free. The table below carries the official rates, with GBP and CAD figures converted at October 31, 2026 market rates, verified at publication.

The company's own worked example makes the math concrete: a 500-token call with one answered question costs USD 0.00028 (about £0.00021 or C$0.00039). A million of those calls would run USD 280 (about £209 or C$391), though that arithmetic is ours, not the company's. Note where the money goes in that example: the per-question fee is the largest of the three charges. The more questions you ask per call, the more the billing behaves like a per-decision tax rather than a token meter.

The fine print has real teeth. Only answered questions are billed; refused questions cost nothing, and an all-refused request carries no text, question, file or call charge at all. The per-call fee is charged once per call, never on refusals, zero-answer, free, promotional or replay calls. Several files each pay their own fee, and the same text token is never charged twice. Text caching is automatic and cuts the token rate by 90%, but the cache lifetime, minimum size and eviction policy are not published, so cached savings are best treated as a bonus rather than a plan. New accounts get USD 5 of free credit at signup, subject to eligibility checks, one per verified email on the first workspace. One credit costs USD 1, and the minimum top-up is USD 5. There are no subscriptions, no tiers and no published volume discounts on the pricing page.

The obvious question, how much does a million decisions cost, is the one the company has not answered. Its own site lists the cost per million tasks as pending. Until that figure is published, the per-call arithmetic above is the only honest basis for comparison.

How it compares

RightNow's homepage places its model on the Artificial Analysis Decision Index next to established rivals. The table below reproduces that framing, with the honesty flag the company itself includes: RightNow's 63.20 was measured by the company on the public held-out half of the benchmark on October 28, 2026. It is not an independent Artificial Analysis evaluation, and the company notes the ordering is "indicative, not an Artificial Analysis ranking."

On price, the comparison is deliberately hard to do because the billing models differ. Using the same 500-token, one-question worked example as the yardstick, and our own arithmetic flagged as ours: TypeSafe's Jev at USD 0.042 per million input tokens would put the input side at roughly USD 0.000021 for the call, while OpenAI's Decisions API at USD 0.10 per million would come to roughly USD 0.00005. RightNow's USD 0.00028 is higher, and the difference comes from the per-question fee, which dominates at small input sizes. That is the trade the company is making: simpler, predictable per-decision billing in exchange for a higher floor on tiny calls. For long-running streams where one call carries many decisions, the balance shifts the other way, since additional questions each cost only USD 0.00013 without a new call fee.

For context on the market, OpenAI's Decisions API went generally available on October 22 at USD 0.10 per million input tokens with free output, as covered in our earlier OpenAI reasoning-models launch piece, and the category was largely created by TypeSafe's Jev. For another per-decision comparison point, see our Milliseconds.ai review, which covers a rival decision API at roughly USD 0.04 per decision with a 90-millisecond latency claim.

The honesty ledger

This is a one-day-old product, and the gap between company claims and verified facts is the story. Here is where each sits:

• Company claims, unverified: 30-millisecond decisions on text; up to 1,024 questions in one call; the 63.20 Decision Index score; "most decision models wait for the whole input" as the problem statement; SOC 2 Type 2, which the site footer names but which we could not independently verify.
• Company statements, taken at face value with caveats: the model "will stay closed source and available via RightNow API"; the team "spent half of our YC money training a model"; this is "our first checkpoint" and it is "still a little buggy," per the founder's launch post. The buggy-checkpoint admission is refreshing, and it means early adopters should treat outputs as draft-quality.
• Verified by reading the official docs: the full price card, the question types, the minimums and refusal rules, the supported inputs and formats, the 38,000-token standard context, the endpoint and SDK structure.
• Unknown: cost per million tasks (listed as pending); customer names; funding beyond the Y Combinator backing the site itself names; the promised Reflex Training paper; and how accuracy degrades on long streams versus one-shot calls.

One more flag that matters for searchers: there is a different, unrelated "RightNow AI" that optimizes CUDA kernels for GPUs, with listings on software comparison sites describing an AI-powered CUDA profiling platform. That company is not this one. This RightNow is the Y Combinator-backed decision-model startup at rightnowai.co, and mixing the two up will waste your time.

Pro tips and common mistakes

• Batch questions into one call. The per-call fee is charged once per call with at least one answer, so 100 questions in one call costs USD 0.013 in question fees plus one USD 0.000055 call fee, instead of 100 separate calls. Use sessions for streams that run indefinitely.
• Do not pay for refused questions, and design for them. Refusals are free, so clear question instructions that make the model answer confidently are not just quality control, they are cost control. An all-refused request bills nothing at all.
• Cache is a bonus, not a budget. Cached tokens are 90% cheaper, but the company publishes no cache lifetime, minimum size or eviction policy. Do not build a cost model that assumes every repeat request hits cache.
• Watch the audio minimum. Audio bills USD 0.00036666667 per started second with a 50-second minimum per file, so a 10-second voice clip costs about USD 0.0183, roughly the same as a full video file. Short clips are the worst value in the rate card.
• Do not compare the 63.20 to 65 as if they were measured the same way. One is company-run on half the test set; the others are independent evaluations. Treat RightNow's number as a floor for interest, not a ranking.

Frequently asked questions

What is RightNow System Zero?

It is a decision-model API from the Y Combinator Fall 2026 startup RightNow. You send it state, text, JSON, images, PDFs, voice, video or robot data, plus named questions, and it returns each decision with per-option probabilities and a confidence value, updating as streaming input arrives.

How much does RightNow System Zero cost?

The official rate card charges USD 0.19 per million uncached input tokens, USD 0.00013 per answered question and USD 0.000055 per answered call, with free output. The company's worked example of a 500-token call with one question totals USD 0.00028. New accounts get USD 5 of free signup credit, subject to eligibility checks.

Is System Zero open source?

No. The founder states the model "will stay closed source and available via RightNow API." The SDKs and docs are public; the model weights are not. A technical paper on the Reflex Training and Reflex Architecture is promised but not yet published.

How fast is RightNow System Zero?

The company claims 30-millisecond decisions on text, stated in its launch announcement. We did not test the model, and there is no independent latency measurement yet, so treat the figure as the company's target rather than a verified spec.

How does System Zero compare to Jev or OpenAI's Decisions API?

On RightNow's own benchmark table, its 63.20 Decision Index score sits just below Jev-family scores of 65, but that score is company-measured, not independently evaluated. On price, its per-question billing makes tiny calls more expensive than Jev's or OpenAI's per-token rates, while heavy multi-question streams are where its model gets cheaper.

Is RightNow the same as the CUDA optimization company called RightNow AI?

No. Software comparison sites list a separate "RightNow AI" that profiles and optimizes CUDA kernels for NVIDIA GPUs. The decision-model startup in this article is the Y Combinator-backed company at rightnowai.co. They share a name and nothing else.

Verdict

System Zero is the clearest attempt yet to sell decisions as a metered unit instead of renting tokens by the million. The streaming pitch is aimed squarely at robotics, phone calls and live video, where the current generation of decision models still struggles, and the per-question billing makes costs legible in a way per-token pricing is not. The caveats are just as clear: the flagship benchmark score is self-measured, the Reflex paper is promised rather than published, the founder calls the first checkpoint "a little buggy," and the company has not published the one number enterprise buyers will ask for, the cost per million tasks. With USD 5 of free signup credit, testing it is cheap. Trusting it in production is a decision you should make with your eyes open.

Sources: RightNow homepage; RightNow System Zero pricing; RightNow System Zero docs; Decisions reference; Inputs reference; FourWeekMBA; DailyCited; Zotpaper.