Articul8 Bivia: A Decision Model That Can Abstain — and Has No Price List

The most consequential line in Bivia's API documentation is not a benchmark. It is status. An answered call comes back with a value, a probability between 0 and 1, and a distribution over the options you offered. A refused one comes back with value: null, an abstain_reason, and the same distribution — the model telling you, in a typed field, that the evidence did not support a commitment. Articul8 Bivia, launched from Santa Clara on 28 October 2026, is a multimodal decision model rather than a generative one: one HTTP POST in, typed JSON out, zero output tokens. The abstention field is the feature the launch leans on hardest, and it is the feature you can switch off with a single parameter. Whether that appeal holds is a question about calibration, and the reference point is TypeSafe AI's Jev-1.13 — the decision model this blog has covered most, and the one you can already call through OrcaRouter.

Put the two side by side and the difference is what each model will accept as input. Jev answers typed questions about text. Bivia takes text, images, documents, code, sensor signals, audio and video, and grounds its answer in the artefact: a bounding box, a character offset, a video timestamp. What the two vendors share is a business problem, not an architecture. Both are sold as a way to stop paying frontier-model rates for judgments that resolve to a yes, a no, a label, or a number. Articul8's version of that pitch arrives with a working API, a free-token offer, and no published per-token price, which makes the first decision Bivia forces on you an accounting one.

What Bivia actually returns

The product page is unusually specific about the contract, which is worth reading closely because the press release is not. A decision is a POST with your API key in BIVIA_API_KEY; the response is the same JSON shape every time, which is the entire point of a non-generative output. You pick one of nine primitives — Decide, Choice, Boolean, Score, Set, Span, Ground, Extract, and "Data science" — and the answer is coerced into one of five typed shapes. No prose, no parse step, no truncated brace at token 4,000.

• Inputs — text, images, documents, code, signals, audio and video, seven modalities through one shared architecture rather than separate pipelines per format.

• Outputs — Choice, Boolean, Score, Set and Span, each returned with a calibrated probability and a distribution across the options you supplied.

• Modes — Fast runs one decision in a single pass; Deep decides and then challenges its own decision before committing; Fast can be upgraded with think: true. The page's summary strip reads "Fast Deep Thinking" while the detail section describes "Fast · Deep · Abstain", so whether Thinking is a fourth mode or a toggle on Fast is left unresolved by the vendor.

• Refusal — optional. Send allow_abstain: false and Bivia commits to an answer every time, which removes the safeguard and leaves you with a probability that may be near zero.

• Grounding — regions, offsets and timestamps ride along with the answer, so a downstream check can re-open the evidence at the exact pixel or second the model used.

• Language — 50+, with the page naming English, Japanese, Korean, Hindi, Tamil, Spanish, German and Portuguese among them.

The integration surface is broader than a decision API usually gets. MCP servers for Claude Code, Codex, Cursor and VS Code, a CLI and a Claude skill, a floating desktop app for macOS and Windows, Chrome and Edge extensions that let you select a region of the screen and ask about it, a Google Workspace add-on with =BIVIA.DECIDE, =BIVIA.CHOICE and =BIVIA.SCORE formulas, and SDKs in Python, Node.js, Go and Rust. None of that changes what the model is. It does mean the trial costs you a command rather than an integration project.

The abstain switch is the whole product, and it is the least measured thing about it

Abstention is the right idea and the harder half of the claim. A decision layer that guesses is worse than a generative model that guesses, because a typed answer looks auditable whether or not it was earned. So the interesting question is not whether Bivia can return abstained — the API proves it can — but whether its confidence is calibrated enough that the refusals land on the genuinely ambiguous cases and the commitments land on the clear ones. That is a measurable property, and Articul8 has not published a measurement of it.

What exists instead is independent work on the category. JEVal, a benchmark introduced in arXiv 2610.03935, tests 25 model configurations on 11,257 instances drawn from 36 datasets across 10 application domains. Its finding is pointed: general decision models perform best when the answer can be resolved from the evidence in front of them, and get less reliable when specialist knowledge is required or when uncertainty has to be estimated faithfully. The abstract's phrase is that these models can pick the most likely result while "substantially overstating its probability". The same paper runs τ-bench and finds that faster local decisions shorten episodes — but overall task success drops as errors compound over a long trajectory.

That second finding cuts directly at the argument for putting a decision model at every step of an agent. Compounding is the failure mode a per-step accuracy number hides: a step that is right 90% of the time is right 59% of the time after five dependent steps. A separate pre-registered test on Zenodo (SILT-RP-010) put seven decision models through 799 unpublished yes/no decisions and found only one configuration consistent with its own calibration claim; Jev came out at an ECE of 0.052, which the authors recorded as inconclusive rather than confirmed. None of this is about Bivia, and none of it should be read as a verdict on Bivia. It is the standard the category is being held to, and Bivia has not yet been held to it.

Where hands-on data does exist, it is thin enough to be honest about. One developer ran 42 labelled decisions from his own overnight coding agent through Jev-1.13 and Clef-Flash, and both handled the ten computer-use action choices perfectly, including two safety traps — one of which required an abstention — while both missed the same four supervision cases. He describes it himself as "a 42-case benchmark from one person's agent, not a paper". Bivia was not in it.

The headline benchmark names its comparator, which is more than most launches do

Articul8's internal tool-routing test scores Bivia at 0.91 against Jev at 0.84, and the release says so by name rather than hiding behind "a leading competitor". That deserves credit; named comparators are falsifiable in a way that anonymous ones are not. The rest of the benchmark block needs heavier labelling:

• Tool routing — Bivia 0.91 vs Jev 0.84, vendor-internal, methodology not published.

• End-to-end workflow completion — roughly 73% vs 37% for the comparison model, a figure that appears in Articul8's launch posts rather than the press release.

• Energy and Verilog tasks — 100% accuracy on selected tasks after domain tuning. The release itself flags these as in-distribution, meaning the tasks resemble the tuning data; it is a statement about fit, not about generalisation.

• Answer stability — Articul8 claims competitors, Jev among them, changed answers in about one run in eight when options were reordered, while Bivia stayed consistent. No harness, sample or option set is disclosed.

The asymmetry worth noticing is on the other side of the comparison. Jev has been through independent evaluation: 37 datasets in one arXiv study, the JEVal suite, the pre-registered calibration test, and several developer-run harnesses. Bivia, one day old, has a press release and a product page. That is normal for launch day and not a criticism of the model — it is a reason to treat 0.91 as a number to reproduce rather than a number to plan around.

Pricing: you pay for input tokens, and there is no rate card to compute it with

The commercial terms are simple in shape and empty where it matters. No subscription. You pay only for input tokens, which follows from a model that emits none. The first 10,000 users get up to 1 billion tokens free. Access requires registering at articul8.ai/bivia. What is missing is the actual rate: no dollars per million input tokens appears on the product page, in the release, or in the launch posts. The page's cost graphics — a "¢$$$ 1 forward pass" comparison — are illustrations, not prices.

Without a rate you cannot do the one calculation the entire category rests on, which is whether a decision call is cheaper than the generative call it replaces. Articul8's own demo numbers make the appeal visible without making it computable: 10,058 decisions against data-centre cooling telemetry in roughly 15 seconds, 13 of which abstained and were handed to an operator; 200 wafer lots and 1,000 decisions across one 12-hour shift, five calls per lot over eight lithography scanners; a single run quoted at $0.40 without saying what that run contained. Each of those is a demonstration captured and replayed by the vendor, and several examples on the page are explicitly labelled illustrative.

The free tier is the practical answer for now. A billion input tokens is enough to run a real pilot — a few hundred labelled cases through Fast and Deep, with abstention on and off — before anyone has to quote you a price. That pilot is the only way to learn your abstention rate on your own data, and the abstention rate is what decides whether the economics work. A strict threshold sends more cases to humans; a loose one lets confident mistakes through. That trade-off between coverage and risk is the core of selective classification, and no vendor benchmark substitutes for your own threshold.

Video and documents are where Bivia stops looking like the other decision models

The textual decision models — Jev, Clef, the various open rebuilds — take typed questions about text and return probabilities. Bivia's shared architecture over seven modalities is a different product category wearing the same label, and the demo that shows it is a 52-minute 4K documentary of the International Space Station fed in frame by frame in a single request, surfacing seconds-long events such as night-side city lights and an aurora. Sampling frames would have missed them; the claim is that Bivia does not sample.

That is a genuinely different capability, and it points at the industrial use cases Articul8 actually sells: inspection footage where the event is three seconds in forty minutes, engineering drawings, tables and charts, sensor traces. The named customers are Hitachi Energy, CTC, AZing and Megazone, and the platform around Bivia connects to 25-plus enterprise and operational systems with deployment on public, private, sovereign and air-gapped clouds. Domain tuning, though, is sales-led — the page's own instruction for it is "Talk to us about your domain". So the 100% energy and Verilog figures describe tuned deployments built with Articul8's help, not a model you download and point at your plant on a Tuesday.

Where a decision model belongs in an agent — and why that is a routing question

The architecture that makes sense is not "replace the LLM with Bivia". It is a three-part chain: deterministic rules absorb the cases they can settle, the decision model takes the bounded judgments that remain and returns a label with a probability, and the application validates the label against policy and permissions before anything happens. Only then does the workflow choose between a cheap action, a frontier model call and a human. A decision model is a recommendation step, not a source of authority — it cannot authorise a merge, a spend, or a skipped approval, and a well-formed label chosen for the wrong reason is still wrong.

Which means the operational question is routing, and it has three parts: which steps go to the decision layer, which escalate to a generative model when the probability is low or the answer is abstained, and what happens when the decision endpoint itself is slow or down. That is the shape of a routing policy, and it is what OrcaRouter's routing DSL is for — composing a decision call and a generative fallback into one call, with automatic failover behind it so an abstention or an outage on either side has a defined next hop rather than a stuck agent. The generative half of that chain is where the pass-through pricing matters: one API for more than 200 models at provider list price with 0% markup, so when a vendor cuts a price the rate you route at moves the same day.

One honest caveat on availability. Bivia is not on OrcaRouter's catalogue — it is reachable through Articul8's own API and the vendor's platform, and nowhere else that we can vouch for. TypeSafe AI's Jev-1.13 is on the catalogue, which makes a side-by-side trial of the hosted decision model we do serve a one-key exercise; trying Bivia means a second contract and a separate credential. If you want to see whether a decision layer earns its place before signing anything, that asymmetry is the cheapest experiment available.

Who should try Bivia now, and who should wait

Try it if your decisions live in artefacts rather than sentences — inspection video, drawings, scanned claims, sensor logs — because that is the part of Bivia no competitor in the decision-model wave currently matches, and the free token allowance covers the pilot. Try it if you are already an Articul8 customer with the platform deployed, because Bivia also runs inside Bivia Studio alongside the pipeline you have; or if you need air-gapped or sovereign deployment, which the company-level support covers. Try it if a coding agent is where you would use it, because the MCP servers and the curl installer make that a ten-minute test rather than a project.

Wait if you need a unit price to model spend, because there isn't one. Wait if procurement requires an independent benchmark or a published abstention rate, because neither exists yet. Wait if you need weights you can host, licence terms you can read, or a latency SLA — none of those was released. And wait if the domain tuning is the whole value: the gap between Bivia out of the box and Bivia tuned on your domain is undisclosed, and that gap is exactly what you would be buying.

Three things would settle it, and all three are the vendor's to publish: a per-token rate card, an abstention rate measured on a harness someone else can run, and results on a benchmark that predates the launch. Until then the fair reading of Bivia is a well-specified API attached to an unusually strong modality story, sold on benchmarks its own team generated, with the most interesting behaviour — knowing when to refuse — left as the thing you should measure yourself. The abstain switch is free to try and, on the evidence available, still the best reason to.