OpenAI, Microsoft and Cloudflare Joined In: Why That Matters

Image placeholder

Editorial, product-forward hero illustration on an off-white background. Four square tiles stand in a horizontal row, each a different soft shade: pale slate, pale emerald, pale blue, pale sand. Each tile has one small wordmark-style label — 'JEV', 'OPENAI', 'MICROSOFT', 'CLOUDFLARE' — set in a flat sans-serif. Behind all four tiles runs a single thin horizontal line marked with four tick points dated 'SEP 15', 'SEP 29', 'OCT 1', 'OCT 9' — the shared infrastructure spine. One short caption-like tagline runs under the row: 'four decision APIs in twenty-four days'. Clean, flat, geometric. Palette: off-white background, pale slate borders, one emerald accent on the shared line, navy ink for all labels. No AI robots, no neural networks, no brains, no faces.

• Aspect ratio: 16:9
• Source size: 1600 × 900 px (sharp on retina at 760 px display width)
• Format: WEBP (preferred), PNG fallback
• Target file size: under 300 KB

Save to /public/images/blog/<slug>/, then set src and alt on this <MediaFigure /> to swap it in. Keep the ratio prop so the browser reserves the right space.

By Dr. Sanju Abraham, Fractional CTO — 180+ projects shipped, 34 funded companies. Published October 4, 2026.

Key takeaways

• Within 24 days of TypeSafe AI’s Jev launch (September 15, 2026), OpenAI, Microsoft, and reportedly Cloudflare all shipped their own decision APIs. OpenAI previewed its Decisions API at DevDay 2026; Microsoft launched Decision-1 on October 9; Cloudflare reportedly shipped the Clef family around October 1.
• That speed is the signal. Big companies do not clone a new category in three weeks unless they already have the model in a drawer. The drawer was open because the shape of the problem is now obvious to everyone building AI products.
• I call this the three-week consensus: when a credible startup launches a category and the majors ship competing products inside roughly three weeks, that category has stopped being an experiment and started being infrastructure.
• The four implementations share a grammar: typed outputs, confidence or probability per answer, parallel questions in one call, structured schemas instead of prose. Vendor differences are real; the primitives are converging.
• Builders should plan for standards, not bets. Treat decision APIs the way you treat SQL databases or HTTP: pick a vendor on cost, latency, and privacy, but write code against the shared shape so switching is cheap.

The short answer: Between September 15 and October 9, 2026, four different vendors shipped the same kind of product — a typed, structured “decision model” API that returns labels, scores, and probabilities instead of prose. TypeSafe AI started it with Jev. OpenAI’s Decisions API followed at DevDay. Microsoft’s Decision-1 landed in Azure AI Foundry. Cloudflare reportedly joined with its Clef model family. When every major platform ships a competing version of your product inside three weeks, you are not witnessing a trend. You are watching a category become a standard.

A friend texts you: “Everyone is suddenly talking about this ‘decision API’ thing. Is it hype?” You could point them at a benchmark. You could point them at a demo. But the honest answer is simpler: look at the calendar. In 24 days, four of the biggest names in cloud and AI shipped the same shape of product. That does not happen by accident. This piece is about what that calendar means and what you do with it.

What exactly happened in the 24 days after Jev launched?

Four decision APIs shipped in rapid succession. Here is the public record, with the parts I can verify and the parts I am hedging.

• September 15, 2026 — TypeSafe AI launches Jev. The product is a “System One” model that returns typed answers (Choice, Score, Noul) in a single parallel pass. Launch coverage is in VKTR’s write-up. For a plain-English explanation see what Jev is.
• September 29, 2026 — OpenAI previews a Decisions API at DevDay. Secondary coverage reports a new POST /v1/decisions endpoint, three output types called Predicates, Choices, and Scores, and a claim of roughly 10× faster responses than OpenAI’s standard chat path. One outlet reports public beta on October 6, 2026; dates across sources differ by a week. Model name is reported as either GPT-6 Luna or GPT-6.1 Sol; treat the exact name as unsettled. Reporting: ProPakistani launch coverage and firecrawl’s teardown. The deep-dive on OpenAI’s entry lives in our companion piece, OpenAI enters the decisions market.
• Around October 1, 2026 — Cloudflare reportedly ships Clef models. A third-party model directory describes two sizes (27B and 9B), a single-pass scoring interface, and up to four images per request. I could not find an official Cloudflare page confirming the dedicated “decision” branding, so treat the Clef details as reported but unverified. What is on the record is that Workers AI supports structured JSON outputs through its JSON mode, per Cloudflare’s own changelog.
• October 9, 2026 — Microsoft launches Microsoft-Decision-1 in Azure AI Foundry. Microsoft’s announcement describes a model that returns calibrated probability scores for a fixed set of answer options instead of generated text. Third-party listings put the context window at 32,768 tokens and note there are no open weights. Source: Microsoft’s Foundry blog post.

From Jev’s launch day to Microsoft’s launch day: 24 days. From Jev to the OpenAI preview: 14 days. That is faster than most companies ship a configuration flag.

Why does it matter when big companies move this fast?

Three reasons the timing is the message. Taken individually, each of these launches is a product. Taken together, inside a three-week window, they are a verdict.

• Nobody ships a frontier-adjacent model in two weeks from a cold start. The engineering timeline — training, evaluations, product packaging, pricing, documentation, legal review, launch assets — is months, not weeks. If OpenAI, Microsoft, and Cloudflare were able to put a decision product on the shelf inside 14 to 24 days of Jev’s launch, each one of them already had the work in a drawer. The drawer was open because internal product teams had already decided the shape was obvious.
• Fast-follow signals category validation, not category theft. When three majors clone the shape of a startup’s product inside a month, they are not racing to steal customers from the startup. They are racing to prevent being the only major without that shape in their catalog when enterprise buyers start asking for it. The follow is a bet that the category will be big enough that absence from it is expensive.
• Fast-follow starts commoditization immediately. The second a category has three or four implementations with roughly the same primitives, the market stops rewarding novelty and starts rewarding the usual stack of boring virtues: price per million tokens, latency percentiles, data residency, auditability, SOC 2 posture. That is a very different competitive game, and it starts the day the second entrant ships.

If you have been to this movie before — object storage after S3, serverless after Lambda, vector databases after Pinecone — you know the next act. The category gets a boring name, the primitives converge, the price curve bends toward zero, and the thing that was a competitive advantage last year becomes a line item on an architecture diagram. That is the arc decision APIs are on now.

What is the three-week consensus?

The three-week consensus is a market signal I am naming for this pattern. It fires when a credible startup launches a new API category and the major cloud and model providers ship competing versions inside roughly three weeks. When it fires, that category has stopped being an experiment and started being infrastructure. Builders keep misreading the pattern as hype, which is why it needs a name.

The signal has three pieces. First, the follow has to be from vendors who could credibly have ignored the startup for a quarter. Second, the competing products have to share the core primitives, not just the marketing language — in this case: typed outputs, confidence per answer, parallel questions per call, structured schemas. Third, the window has to be short enough that pre-existing internal work is the only explanation. All three are true for decision APIs this month.

When you see this pattern, you have gotten a free piece of information: the category is going to be a standard, and your main job as a buyer or builder shifts from “is this real?” to “which implementation and under what portability contract?” For the broader category background, see not every AI needs to talk and chatbot vs decision model.

Image placeholder

Infographic: a four-column comparison grid of decision APIs. Columns labeled 'TypeSafe Jev', 'OpenAI Decisions', 'Microsoft Decision-1', 'Cloudflare Clef'. Rows labeled 'Primitives', 'Context window', 'Multi-image input', 'Hosted / Open', 'Status'. Cells carry short labels like 'Choice / Score / Noul', 'Choices / Scores / Predicates', '32K tokens', 'reported', 'hosted'. Palette: off-white background, pale slate borders, emerald accent for header row, navy ink. Clean, flat, product-forward. No AI robots.

• Aspect ratio: 16:9
• Source size: 1280 × 720 px (sharp on retina at 760 px display width)
• Format: PNG
• Target file size: under 500 KB

Save to /public/images/blog/<slug>/, then set src and alt on this <MediaFigure /> to swap it in. Keep the ratio prop so the browser reserves the right space.

Are these four products actually the same thing?

Not quite. The grammar is converging; the details still differ. Here is the honest reading, cross-checked against vendor-adjacent coverage.

	TypeSafe Jev	OpenAI Decisions	Microsoft Decision-1	Cloudflare Clef (reported)
Launch	Sep 15, 2026	~Sep 29 / Oct 6, 2026	Oct 9, 2026	~Oct 1, 2026
Primitives	Choice / Score / Noul	Choices / Scores / Predicates	Fixed-option prob. scores	Choice / yes-no / score
Context	32K tokens	not confirmed	32,768 tokens	not confirmed
Image input	text-only at launch	reported	not emphasized	up to 4 images (reported)
Hosted / Open	hosted	hosted	hosted (Foundry)	hosted (Workers AI)
Open weights	no	no	no	no

• Primitives. Jev calls them Choice, Score, Noul. OpenAI calls them Choices, Scores, Predicates. Microsoft describes “calibrated probability scores for fixed answer options.” Cloudflare’s reported Clef family is described similarly. Different names, same three jobs: pick one, rate on a scale, estimate a probability.
• Confidence model. All four are reported to return either a confidence score or a calibrated probability per answer. Calibration quality is the thing you cannot read from a datasheet; see why 70% sure beats yes for the pattern that depends on it.
• Context window. Jev launched at 32K tokens. Microsoft Decision-1 is reported at 32,768 tokens. OpenAI and Cloudflare numbers are less clear in the coverage I checked. Call this “comparable.”
• Input modality. Cloudflare’s reported Clef family supports up to four images per request. OpenAI’s Decisions API is reported to accept image inputs. Jev launched text-only at September pricing. Microsoft’s entry is described in text-and-structured-data terms.
• Hosting and privacy. All four are hosted APIs. None of them is open-weight as of this writing. If “we need to run this in our own VPC” is a hard requirement for a specific workload, this generation of decision APIs is the wrong shelf, and the open-weight question becomes a different conversation — one I will come back to in open-weight decision models.

A full market map — including the smaller independent entrants that tend to arrive right after the majors — lives in the thirteen decision APIs market map.

What should builders actually do with this signal?

Treat decision APIs as infrastructure, write code against the shared shape, and defer vendor lock-in decisions until price and latency stabilize. Three concrete moves.

• Pilot now, but pilot the pattern, not the vendor. The highest-value thing you can learn in Q4 2026 is which workflows in your product benefit from a typed decision call. Pick one — ticket routing, lead scoring, moderation triage — and ship it against whichever vendor minimizes your friction. The pattern transfers; the SDK does not. See your first decision model experiment and five decision API business tasks.
• Write a thin adapter. Hide the vendor behind a one-file interface with three functions: choose, score, verify. Every current implementation can be shimmed into that. When the price curve bends in six months, you swap the file. The compatibility question is covered more carefully in Jev-compatible decision APIs.
• Delay the lock-in decision. You do not need to pick “our decision API vendor” today. You need to pick one for pilot one, one for pilot two if it is a better fit, and let the operational data tell you where the long-run bet should land. For the structured version of that decision, see choosing a decision API in 2026 and reading decision model benchmarks.

The one thing not to do is to interpret “four majors shipped” as “pick Microsoft or OpenAI because they are the biggest.” Size has never been the determining factor in a commoditized API category. Price, latency, calibration, and privacy are. On those, the real answers are weeks away, not days.

Is this hype or is this real?

It is real, and the three-week consensus is the strongest evidence you will get this year. A healthy skeptic should still ask three questions.

• Could the majors be chasing a false signal? Possible but unlikely at this scale. OpenAI, Microsoft, and Cloudflare each have the resources to kill a product that is not working; shipping on this timeline implies internal confidence, not just reactive panic.
• Could the primitives change? Yes, and probably will at the edges. Expect naming to shift (Noul will probably not stick as the industry term), modality to broaden (image, audio, document), and new output shapes to be added. The core three — pick, rate, verify — are unlikely to go away. For the taxonomy that is holding up, see the three questions a decision model answers.
• Could the fast-follow pattern reverse? It can, in categories where the leader has a defensible moat (proprietary data, network effects, distribution lock-in). Decision APIs do not obviously have one. Expect convergence, not divergence.

The honest version: this is early, it is fast, and almost nothing about the specific pricing, benchmarks, or product boundaries will look exactly the same in six months. What will look the same is the shape of the API. That shape is now a standard. Build to it.

Frequently asked questions

What decision APIs did OpenAI, Microsoft, and Cloudflare ship in 2026?

OpenAI previewed its Decisions API at DevDay 2026, with secondary coverage reporting a public beta around early October and three output types (Choices, Scores, Predicates). Microsoft launched Microsoft-Decision-1 in Azure AI Foundry on October 9, 2026, returning calibrated probability scores for fixed answer options. Cloudflare reportedly released the Clef model family around October 1, 2026; third-party sources describe a decision-shaped API, but I could not find an official Cloudflare page confirming the “decision” branding.

How soon after Jev did the big tech decision APIs ship?

Fourteen to twenty-four days. TypeSafe AI launched Jev on September 15, 2026. OpenAI’s Decisions API was previewed on September 29. Cloudflare’s Clef family reportedly shipped around October 1. Microsoft-Decision-1 launched on October 9. From Jev to Microsoft is 24 calendar days — a pace that only happens when internal product teams already have the model in a drawer waiting for a market signal.

What is the three-week consensus?

A market signal I coined for this piece. When a credible startup launches a new API category and the major cloud and model providers ship competing versions inside roughly three weeks, that category has stopped being an experiment and started being infrastructure. The pattern depends on three things: credible fast-followers, shared primitives, and a window too short to be explained by anything other than pre-existing internal work.

Are OpenAI, Microsoft, Cloudflare, and Jev interchangeable?

Not quite. The grammar is converging — typed outputs, confidence per answer, parallel questions in one call — but the specifics still differ. Context windows are comparable (around 32K tokens where published). Image input varies: Cloudflare’s reported Clef family and OpenAI’s Decisions API accept images; Jev launched text-only. All four are hosted; none is open-weight at launch. Vendor differences live in pricing, latency, privacy, and ecosystem — not in the shape of the API.

Should I pick a decision API vendor now?

Not for the long run. Pick one per pilot based on friction and fit, hide the choice behind a thin adapter with choose / score / verify functions, and let the operational data pick your long-term vendor over the next six months. The price and latency curves are moving fast enough that today’s winner is unlikely to be next year’s winner. Your job is to make the switch cheap, not to pre-commit.

Does this mean the decision API category is “won”?

The category is validated, not won. Validation and consolidation are different events separated by one to three years in most API categories. Expect more entrants, more primitives, more pricing shifts, and at least one open-weight entrant before any of this settles. The thing that is settled is that typed decision APIs are now a standard building block, in the same tier as structured JSON output or embeddings.

Where should I start if I want to try one of these?

Start with the pattern, not the SDK. Pick one small, repetitive judgment in your product — tagging, routing, scoring, moderation — and prototype a single decision call against it with whichever vendor you already have an account at. Measure accuracy, latency, and cost against your current approach. The learning transfers to any of the four, and the transfer is cheap if you kept the adapter thin. For the first-call walkthrough, see your first decision model experiment.

---

If you are weighing which decision API to pilot in Q4, or whether to build the thin adapter now versus later, that call-site-by-call-site judgment is exactly what a Fractional CTO is for. The three-week consensus just saved you the “is this real?” phase of the roadmap; the “which one and when” phase is where mistakes get expensive. One hour of the right conversation can save a quarter of engineering time. More pieces in this thread live in the AI Strategy archive, and the strategic frame for SMBs lives in an AI strategy for SMB owners that survives contact with reality.