Milliseconds.ai Review: A $0.04 Decision API That Answers in 90 Milliseconds

Most AI APIs sell you words. Milliseconds.ai sells you answers your code can branch on — a label, a score, a yes or no, a filled-in form field — for $0.04 per million input tokens, with output tokens free. The standalone API platform officially launched on October 25, 2026, according to launch coverage from Wisevoter, though its public footprint is older: its engineering blog dates to September 19, and the underlying models were already running inside its parent company CloudRaker's document product. That matters: this is not a two-week-old startup API, but a repackaging of infrastructure that was already serving paying document workloads.

In short: milliseconds.ai's decision-machine-1 gives your application eight typed operations over text and images — classify, extract, verify, and more — at $0.04 per million input tokens, with a free tier of 125 million tokens a month and production unlocked by a $10 credit bundle. It is slightly cheaper than the two best-known rivals in the new "decision model" category, and it does more of the work for you, including image input and audio transcription. The catches are real but small: extraction is capped at five fields per call, there are no published accuracy benchmarks of its own, and the cheapest price in the category differs by two-tenths of a cent per million tokens — not a reason to switch on its own.

What milliseconds.ai actually is

The pitch is easiest to understand by looking at what it refuses to do. Send decision-machine-1 a support message and it will not write you a reply. It will tell you the message is about billing, with a probability of 0.74, and that it is urgent, with a probability of 1.00. Send it an invoice and it will not summarize it. It will return the vendor, the invoice number and the total as structured fields. The model never generates prose. Software does not want paragraphs. It wants values it can put in an if statement.

There are eight operations, each with its own response contract: yes-no (is a statement about the text true?), classify (pick a label from your set, with scores for every alternative), classify-tree (walk a label hierarchy), rate (place the text on a scale you describe), answer (find a passage and return it with character offsets, or null), extract (fill a JSON Schema), entities (find people, organizations, dates and other typed values) and verify (check a proposed value against the source). Everything goes through one base URL, https://api.milliseconds.ai, with one API key, and every call is synchronous — no job to poll, no webhook to wait for.

There is also a second model in the same account. rakeaudio transcribes audio with word timestamps and speaker labels, in 25 European languages, with up to four speakers. TypeScript and Python SDKs, a command-line tool, and OpenAI- and TypeSafe-compatible surfaces are also available, so code written for a rival decision API can be pointed at milliseconds.ai with less rewriting.

The price, with the fine print

The headline number is simple: $0.04 per million input tokens, and output tokens are free. That is the whole rate card for decision-machine-1. There are no per-capability prices and no per-model tiers. Every operation costs the same, which makes estimating a bill unusually easy for an AI API. What counts as input is the validated request body — your text plus the labels, questions and schemas you send — and repeated text in separate calls is billed again. Responses carry an x-input-tokens header so you can audit each call.

The free tier is generous by the standards of the category. Test keys come with 125 million free input tokens a month, no card required, and the allowance covers every capability. The limits on test keys are 30 inference requests per minute and 500,000 input tokens per minute. Production is unlocked by buying a credit bundle of at least $10. Auto top-up is optional — there is no saved card requirement and no subscription. Production keys start at 1,000 requests a minute and 1 million tokens a minute, and the limits rise with lifetime spend and account age — reaching 10,000 requests a minute with no token cap at $500 and 30 days.

Three pieces of fine print are worth knowing before you budget. First, test and production balances are separate: test keys use only the monthly free allowance, production keys use only purchased credits, and buying credits does not refill the test allowance. Second, the free allowance is shared across your whole organization — every test key draws from the same 125 million tokens. Third, when the free allowance runs out, test requests pause until the next monthly renewal. None of this is hidden; it is all on the pricing page. But the pause-on-exhaustion detail is the one that will surprise someone mid-demo.

Images are billed as a fixed number of input tokens, set by the detail level and the capability. A classify call bills an image as 1,000 tokens at low detail (512 px), 2,000 at medium (768 px, the default) and 4,000 at high (1,024 px). Answer costs 1,500, 3,000 and 6,000 tokens; extract, entities and verify cost 2,000, 4,000 and 8,000. In dollars, 1,000 images through classify at high detail cost $0.16, and 1,000 through extract at high detail cost $0.32. Images bill once per request; you send bytes, a data URL or base64 (JPEG, PNG or WebP, up to 5 MB), and the company stores none.

Transcription has its own rate card. rakeaudio costs $0.12 per audio hour with speaker labels and $0.10 per hour without them. That is still the $0.04 rate in disguise — an audio hour converts to 3,000,000 input tokens with speakers, 2,500,000 without — and the same credits pay for both models. The free test allowance covers roughly 41 hours of audio with speakers, or 50 hours without.

For scale: one million decision calls a day at 400 input tokens each is 400 million tokens, or $16 a day; the company's estimator puts one billion tokens at $40 a month. The free tier — about $5 of production value a month — is a genuine evaluation budget, not a teaser.

How fast is it, really?

The name is a promise, so the latency numbers deserve a close look. The pricing page lists model time at medium image detail as roughly 90 milliseconds for a decision, 0.3 to 0.9 seconds for an answer call, and 2.5 to 3.5 seconds for an extract call. Those are the company's own figures, measured on its own infrastructure, and they describe model time — not the round trip from your server. The honest reading: the "milliseconds" in the name applies to simple decisions — route this, flag that, check this value — not to extraction.

For comparison, TypeSafe says Jev answers in 70 to 500 milliseconds and Microsoft says Decision-1 is 35 times faster than GPT-6 Sol on decision tasks — both vendor figures, with no independent head-to-head yet. Treat every latency figure in this category as a vendor claim until someone publishes one.

Decision models are suddenly a category

Milliseconds.ai did not invent this product shape. TypeSafe's Jev launched it in September 2026, and Microsoft followed with Decision-1 on October 9. What was a curiosity six weeks ago is now a four-way race, and the price points have already converged:

Two things stand out. First, $0.042 per million input tokens looks like the market-clearing price for this category: Jev and Decision-1 both charge it, to the third decimal. Milliseconds.ai undercuts it by $0.002, about 4.8%. That is a real difference on the rate card and a negligible one on any actual bill: at one billion tokens a month, it saves you $2.

Second, the products differ more in what they do than in what they cost. Milliseconds.ai is the only one of the four that bundles transcription in the same account and credit balance, and the only one with a dedicated extraction operation returning source spans with offsets. Jev's distinctive feature is its three question types — Noul, Choice and Score — evaluated in one parallel pass over shared state, which suits workflows asking several bounded questions about the same context at once. Decision-1's edge is distribution: a 9-billion-parameter model, post-trained from Alibaba's Qwen3.5-9B, inside Microsoft Foundry. OpenAI's Decisions API is reported at $0.10 per million input tokens, and the one most likely to already sit inside your OpenAI billing.

What's under the hood — and why the company tells you

Milliseconds.ai is unusually candid about its internals. The name decision-machine-1, its engineering blog says, "should not be read as a claim that every operation uses one newly trained foundation model." The choice path uses Qwen2.5-1.5B-Instruct, constrained to an option token, with scores from the option log probabilities. The span path uses GLiNER2.5. Transcription runs on parakeet-ultra from Moondream and NVIDIA's Sortformer v2, both under a CC BY 4.0 licence. Decision-machine-1 is an assembly of small open models behind a carefully designed interface, not a new foundation model — which the company presents as a feature, since a narrow interface lets it swap backends without users redesigning integrations.

That assembly has a pedigree: the same models and infrastructure powered CloudRaker's Paperwork API, a document parsing product, before they were split out into milliseconds.ai. But milliseconds.ai publishes no accuracy benchmark for decision-machine-1 — no equivalent of Microsoft's 36-benchmark claim or TypeSafe's four-workflow evaluation. You cannot compare decision quality across these products from published numbers today; your own labelled data is the benchmark that matters, and the free tier exists so you can run it.

Pros and cons

What works: the lowest published rate in the category, as a single flat rate across every capability, so bills are predictable. The free tier is genuinely usable — 125 million tokens a month with no card — and production starts at $10 with no subscription. Image input works out of the box, transcription lives in the same account, and the eight operations map cleanly onto the jobs developers actually build: routing, triage, extraction, verification. The docs are thorough, with worked recipes for triage, extraction, moderation and PII detection.

What doesn't: extract is capped at five fields per call, so a document with more fields needs multiple calls or a different product — and the company's answer is its own Paperwork API, an upsell rather than a fix. There are no published accuracy benchmarks for decision-machine-1, so you are buying on price, latency and interface, not measured quality. The cheapest price in the category saves about 5%, which is noise on any real bill. And the launch story is slightly muddled: blog posts date to September 19 and a Product Hunt listing circulated weeks ago, so October 25 is a coming-out moment for a service already running.

Who should try it

The clearest fit is a team already chaining decisions inside an agent or support pipeline — routing tickets, grading agent actions, checking values against a source — where every decision currently goes through a general-purpose LLM at ten to fifty times the cost. The second fit is a team that needs image understanding and transcription from the same vendor and credit balance, which none of the three rivals offers in one place.

Start on the free tier with the one workflow you actually run — ticket triage is the obvious first candidate — and measure it against your own labelled examples. If the decisions are right and the latency holds, a $10 production bundle is a cheap experiment. If they are wrong, no price fixes that, and the free tier lets you find out before spending anything.

FAQ

How much does milliseconds.ai cost?

$0.04 per million input tokens for decision-machine-1, output free. Test keys include 125 million free input tokens a month with no card. Production needs a credit bundle of at least $10; auto top-up is optional. Transcription is $0.12 per audio hour with speaker labels, $0.10 without.

Is milliseconds.ai free?

For evaluation, yes. Test keys renew 125 million free input tokens every month and cover every capability, including transcription — roughly 41 hours of audio with speakers or 50 hours without. Production use is pay-as-you-go from purchased credits, and the free allowance does not apply to production keys.

What can decision-machine-1 do?

Eight typed operations over text or images: yes-no, classify, classify-tree, rate, answer, extract, entities and verify. It returns labels, scores, source spans and structured fields — never prose — with OpenAI- and TypeSafe-compatible surfaces as well.

How is milliseconds.ai different from TypeSafe Jev and Microsoft Decision-1?

All three are decision-only models that return probabilities instead of text, and all three charge nothing for output. Milliseconds.ai is slightly cheaper at $0.04 per million input tokens against $0.042 for Jev and Decision-1, accepts images, bundles transcription in the same account, and offers dedicated extraction and span operations. Jev evaluates several question types in one parallel pass; Decision-1 is a 9-billion-parameter model in Microsoft Foundry.

Key takeaways

• Milliseconds.ai's decision-machine-1 turns text and images into labels, scores, fields and yes/no answers for $0.04 per million input tokens, with output free.
• Free test keys include 125 million input tokens a month with no card, and a $10 credit bundle unlocks production with no subscription.
• Transcription is bundled in the same account at $0.12 per audio hour with speaker labels and $0.10 without, billed as input tokens.
• It undercuts Jev and Microsoft Decision-1 by about 5% on the published input-token rate, but the bigger differences are image input, transcription and dedicated extraction operations.
• Extract is capped at five fields per call, there are no published accuracy benchmarks for decision-machine-1, and the product is an assembly of small open models rather than a new foundation model.

Decision models went from a single product to a four-way category in six weeks, and milliseconds.ai's entry is the cheapest and the most complete of the four on paper. The price is not the reason to pick it — two-tenths of a cent per million tokens is not a moat. The reason is that it does the most work for you before you write any code. Try it on your own data, on the free tier, and let the decisions speak for themselves.

For context on where decision APIs fit in the agent stack, see our coverage of 1Password for Agents and Walmart letting AI agents shop with spending limits. The official product pages are milliseconds.ai, its pricing page and its documentation.