OpenAI Decisions API Hits Beta: $0.10/1M Tokens [2026]

OpenAI opened its Decisions API to every developer in public beta on October 6, 2026, a quiet but significant move that pushes the company further into the plumbing of everyday software. Instead of generating a conversational reply, the new endpoint takes a block of text or an image and returns a typed answer: a yes-or-no call, a pick from a fixed list of options, or a score against a rubric. For engineering teams building AI agents, support triage systems, or content moderation pipelines, that is a different shape of product than the chat-style API most developers have spent the last three years building against.

The announcement landed first in OpenAI’s developer community forum on October 6, then spread through technical blogs and trade press over the following two weeks. It arrives at a moment when the AI developer tools cluster is unusually crowded. Anthropic just slashed Claude Haiku 5.5 pricing by 90 percent, GitHub Copilot and Cursor are trading blows over which platform wins the AI coding race, and DeepSeek just opened its own API to the public. The Decisions API is OpenAI’s answer to a specific gap in that race: a dedicated, cheaper, faster interface for the bounded decisions that sit inside almost every AI-powered application, not the open-ended generation that dominates headlines.

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

What the OpenAI Decisions API actually does

The Decisions API lives at a new endpoint, POST /v1/decisions, and it is built around a narrow idea. Developers submit evidence, which can be plain text, user messages, or inline base64-encoded images, along with a question that fits one of three templates. A predicate question asks whether a condition is true or false. A choice question asks the model to pick from a fixed set of labeled options. A score question asks for a rating against ordered rubric levels. Choice and score questions return probability distributions across the possible outcomes rather than a single flat answer, which gives developers a confidence signal to build logic around instead of just a label.

OpenAI is explicit about the intended use cases: content classification, request routing, work prioritization, and selecting which model, tool, or action an AI agent should invoke next. That last category matters most for the coding-tools audience. As AI agents move from answering questions to taking actions inside codebases, deployment pipelines, and customer systems, someone has to decide which tool the agent calls, how urgent a task is, or whether a generated output is safe to ship. OpenAI is betting that developers would rather call a cheap, fast, typed endpoint for that decision than route it through a full conversational model and then parse the response themselves.

At launch, only one model powers the endpoint: gpt-6-luna. There is no multi-model selection in the public beta, so developers evaluating the tool are testing a single, fixed backend rather than choosing between tiers. OpenAI also shipped a Decisions Playground, a browser-based tool for drafting and testing decision questions before writing any code, mirroring the prototyping workflow Google already offers through AI Studio for Gemini models.

Pricing: $0.10 per million input tokens, no output charge

The headline number developers are reacting to is price. OpenAI is charging $0.10 per 1 million input tokens for gpt-6-luna through the Decisions endpoint, with no separate charge for output tokens, cache reads, or cache writes. That is a deliberately aggressive number for a decision-making tool, and it is designed to make high-volume, low-stakes classification cheap enough to run on every request rather than sampling a subset.

There is a catch worth flagging for anyone budgeting a production rollout. Reporting on the launch, including a technical breakdown published by Mixed News, notes that inputs above 272,000 tokens are billed at double the base rate, and that routing requests through US or EU regional processing adds a 10 percent surcharge. Neither of those modifiers is prominently stated in OpenAI’s own published guide, so teams should confirm the applied rate against their account settings before committing to the endpoint for anything approaching production-scale traffic.

OpenAI also claims Decisions answers return roughly 10 times faster than running the same evaluation through its general-purpose Responses API. That is a meaningful claim if true, because latency is often the real bottleneck in agent pipelines that need to make dozens of small routing calls per user action. But OpenAI has not published a benchmark methodology, percentile breakdown, or hardware configuration behind that figure, so it should be read as a marketing claim rather than an independently verified result until outside testing catches up.

Detail	Decisions API, public beta
Public beta launch date	October 6, 2026
Endpoint	POST /v1/decisions
Supported model	gpt-6-luna only
Base price	$0.10 per 1 million input tokens
Output token charge	None
Long-context surcharge	2x base rate above 272,000 input tokens (per Mixed News)
Regional processing surcharge	10 percent for US or EU processing (per Mixed News)
Claimed speed advantage	Roughly 10x faster than the Responses API, per OpenAI (unverified independently)
Image input format	Inline base64 only, no hosted URLs or file IDs
Decision types	Predicate (yes/no), choice (fixed set), score (rubric)

Why OpenAI built a separate decision layer instead of reusing chat

It is worth stepping back to see why this product category exists at all. Classification and routing logic inside software used to be handled by deterministic rule engines, regex pattern matching, decision trees, and hand-tuned machine learning classifiers trained on labeled datasets. Those systems are cheap to run and predictable, but brittle: they fail silently on inputs the rules did not anticipate, and updating them requires retraining or rewriting logic rather than adjusting a prompt.

Large language models changed that equation by letting developers describe a classification task in plain language instead of building a dataset. The problem is that a general chat completion endpoint is overkill for a yes-or-no question. It returns free-form prose that has to be parsed, it costs money for output tokens the developer does not actually need, and it is slower than a narrowly scoped call has to be. The Decisions API is OpenAI’s attempt to package the semantic flexibility of an LLM into an interface that behaves like the old rule engines: fast, typed, and bounded, while keeping the ability to handle messy real-world inputs that rules choke on.

That framing explains why OpenAI positions this squarely as a developer tool rather than a consumer feature. The use cases OpenAI names, content classification, request routing, work prioritization, and tool selection inside agents, are all backend plumbing decisions that happen thousands of times per hour in a busy application. None of them need the full generative range of a conversational model. They need a fast, cheap, reliable verdict.

How this fits the broader AI coding tools race

The Decisions API lands in the middle of an unusually busy stretch for AI developer infrastructure. Anthropic has been aggressively repricing its lineup, most recently cutting Claude Haiku 5.5 pricing by 90 percent in a move widely read as the opening shot in a broader AI pricing war. Google has been pushing Gemini deeper into agentic workflows, and the ongoing comparisons between Gemini 4 Argon and Claude Opus 5.5 on coding benchmarks show how much of the competitive fight now centers on developer-facing capability rather than consumer chat polish.

Meanwhile the coding-agent ecosystem itself keeps fragmenting and then consolidating. Claude Code recently shipped a plugin ecosystem that crossed 341 available tools, OpenCode 2.0 opened API access to more than 75 AI models in a single interface, and a new coordination platform called BAND launched specifically to let four AI coding agents work together on the same task without stepping on each other. DeepSeek, for its part, opened its V4.1-Flash-1017 model’s API to developers directly, undercutting the established players on price the same way Anthropic did with Haiku.

Against that backdrop, the Decisions API is less a flashy new model and more an infrastructure bet. OpenAI is not trying to win a benchmark leaderboard with gpt-6-luna. It is trying to become the default routing and classification layer underneath agent frameworks that are multiplying every month. If an agent coordination platform like BAND needs to decide which of four coding agents should handle a given subtask, or an IDE plugin needs to decide whether a generated pull request is safe to auto-merge, that is exactly the kind of bounded decision OpenAI wants developers reaching for its new endpoint to resolve.

Competitive positioning against Anthropic, Google, and Microsoft

There is no publicly confirmed, directly equivalent decision endpoint from Anthropic or Google at the time of this launch, so a precise feature-for-feature comparison is not possible yet. What can be compared is market positioning. Microsoft has already staked out adjacent territory with Decision-1, an AI model the company says scores choices dramatically faster than conventional approaches, aimed at a similar bounded-decision niche inside enterprise workflows. The two products are not identical: Microsoft’s Decision-1 is pitched as a scoring model for enterprise decision workflows, while OpenAI’s Decisions API is a general-purpose typed endpoint aimed at application developers building classification and routing logic directly into their own software and agents. But the fact that two major vendors are independently building toward the same category in the same month signals that specialized decision-making interfaces are becoming a recognized product line of their own, not a one-off experiment.

Forbes contributor Ron Schmelzer framed the launch as part of a broader wave of AI decision models emerging in late 2026, noting that OpenAI’s per-token price sits above at least one smaller rival in the space while OpenAI compensates with the pull of its existing developer ecosystem and tooling. That is a familiar pattern in this market: newer, smaller vendors compete on price, while the incumbents with the largest existing developer base compete on convenience, documentation, and the fact that teams are often already paying for other OpenAI products.

Vendor	Product	Positioning	Status as of Oct 19, 2026
OpenAI	Decisions API	Typed predicate, choice, and score endpoint for app and agent logic	Public beta since Oct 6, 2026
Microsoft	Decision-1	Scoring model for enterprise decision workflows	Previously announced, separate product line
Anthropic	Claude Haiku 5.5	Low-cost general model, not a dedicated decision endpoint	Priced down roughly 90 percent in recent repricing
DeepSeek	V4.1-Flash-1017	General-purpose model with newly opened public API	API opened to developers
Google	No confirmed dedicated decisions endpoint	Agentic workflows built around Gemini models generally	No direct equivalent confirmed

What is actually confirmed versus still unverified

Because this is a beta launch barely two weeks old, it is worth being precise about what is solid and what is not. The launch date of October 6, 2026, the endpoint name, the three decision types, the single supported model, and the headline $0.10 per million input token price are all confirmed directly from OpenAI’s own developer documentation and changelog. The 10x speed claim, the long-context surcharge, and the regional processing fee come from secondary technical reporting rather than OpenAI’s primary documentation, which is why they carry more uncertainty.

What is not yet confirmed is just as important. There is no public list of named companies or production developers using the Decisions API at scale yet. OpenAI has not published numeric rate limits, a dedicated SDK, or a language-by-language client library beyond its general API ecosystem. Enterprise privacy claims, including zero data retention and HIPAA eligibility, appear in some secondary coverage but are not spelled out with eligibility conditions in OpenAI’s own guide. Developers evaluating the endpoint for anything beyond a prototype should treat those gaps as open questions to resolve with OpenAI directly rather than assumptions to build a roadmap around.

Risks developers should weigh before adopting it

The most immediate risk is over-trusting the probability scores the endpoint returns. A choice or score question returns a probability distribution, which looks precise, but a probability is still a model’s best guess dressed up in a number. Teams building moderation or safety-critical routing logic on top of these scores need calibration testing of their own rather than assuming the returned confidence maps cleanly to real-world accuracy.

The second risk is scale without review. A $0.10 per million token price point is explicitly designed to make it economical to run a decision on every single request rather than a sample. That is useful, but it also removes a natural cost-based brake that might otherwise prompt a team to add human review at high volume. Cheap automated decisions at massive scale raise the stakes of any systematic bias or blind spot in the underlying model, because there is no longer a budget constraint forcing anyone to look closely.

Finally, there is vendor lock-in risk baked into a single-model beta. With only gpt-6-luna supported, any application built against the Decisions API today is implicitly betting on that one model’s behavior staying stable, or on OpenAI’s willingness to preserve backward compatibility as it expands model support later. Teams that have already built provider-agnostic routing layers around tools like OpenCode’s 75-model API access may be reluctant to introduce a single-model dependency for a core piece of application logic.

Market reaction so far

Reaction in the two weeks since launch has been measured rather than explosive. Developer community threads on OpenAI’s own forum show engineers testing the Playground and comparing early latency against the Responses API, but no major production case study has surfaced publicly yet. Technical outlets including Apidog and StackOne published explainer posts within days of launch, a sign that the developer tools press sees this as a durable new category rather than a minor API addition, even if adoption data has not caught up to the coverage.

That gap between press attention and confirmed production adoption is normal for a two-week-old beta, but it also means the real test of the Decisions API is still ahead. The products that will prove or disprove OpenAI’s bet are the agent frameworks and coding tools racing to ship faster this quarter, including the expanding Claude Code plugin ecosystem and multi-agent coordination platforms like BAND, which need exactly the kind of fast, cheap routing decisions this endpoint is built to provide.

What this means for developer workflows going forward

For teams building with AI agents today, the practical takeaway is to treat the Decisions API as a candidate for the narrow, high-frequency decisions buried inside a pipeline rather than a replacement for a full model call. A support ticket router, a content moderation pre-filter, or a step that decides which of several coding agents should pick up a task are reasonable first experiments. A decision with legal, financial, or safety consequences is not, at least not without independent calibration and a fallback path when the model’s confidence is low.

It is also worth watching whether competitors respond with their own dedicated endpoints. If Anthropic, Google, or a smaller specialist vendor ships a comparable typed-decision API in the coming months, pricing on this entire category could compress quickly, echoing what already happened with general-purpose model pricing after Anthropic’s Haiku cut and DeepSeek’s API opening. A category this new rarely stays a one-vendor market for long once the use case is validated.

Five predictions for the next six months

At least one major competitor, most plausibly Google or Anthropic, will announce a comparable typed-decision endpoint within two to three quarters, following the same pattern as pricing wars that have already played out across general-purpose models.
OpenAI will expand model support for the Decisions API beyond gpt-6-luna once the beta stabilizes, likely adding a smaller or cheaper variant aimed at even higher-volume routing use cases.
Agent coordination frameworks, including multi-agent platforms like BAND, will become the primary adopters, since tool and task selection inside agent pipelines maps almost exactly to the predicate and choice question types OpenAI has shipped.
Independent benchmark data on the claimed 10x speed advantage will surface from third-party developers within a few months, either validating or complicating OpenAI’s marketing claim.
Pricing on this endpoint will come under pressure faster than OpenAI’s general-purpose API pricing has, because classification and routing are commodity tasks that smaller, specialized vendors can undercut without matching OpenAI’s broader model capability.

How this compares with older classification approaches

It helps to put the Decisions API next to the tools it is quietly trying to replace. A traditional rule-based router requires a developer to enumerate every input pattern in advance, which works well for narrow, stable inputs like file extensions but breaks down fast on messy natural language, like a support ticket that could be phrased a hundred different ways. A trained machine learning classifier solves some of that flexibility problem, but it requires a labeled dataset, a training pipeline, and retraining every time the categories change, which is slow and expensive for a fast-moving product.

The Decisions API collapses both of those steps into a single API call with a plain-language question, trading infrastructure complexity for per-request cost and a dependency on OpenAI’s model behavior. That tradeoff has been playing out across the industry for the last two years as LLM-based classification has gradually displaced purpose-built classifiers for teams that value speed of iteration over the lowest possible per-call cost at extreme scale. The Decisions API is a bet that at $0.10 per million input tokens, the economics finally tip in the LLM’s favor even for high-volume routing tasks that used to be too expensive to run through a full model.

Getting started with the Decisions endpoint

Developers who want to try the endpoint can start in the Decisions Playground before writing code, which lets a team draft predicate, choice, or score questions and test them against sample evidence directly in the browser. From there, the production call is a straightforward POST request to /v1/decisions with the evidence and question included in the payload. A minimal predicate-style request looks like this:

POST /v1/decisions
{
  "model": "gpt-6-luna",
  "evidence": "Customer message: My payment failed three times and I was still charged twice.",
  "question": {
    "type": "choice",
    "options": ["billing_error", "fraud_review", "general_support", "feature_request"]
  }
}

The response returns a probability for each option rather than a single flat label, which is the detail developers should design around rather than ignore. Routing logic that only reads the top-ranked option throws away the confidence signal that is arguably the most useful part of this API compared to a plain chat completion.

The bigger picture for AI developer tools

Step back far enough and the Decisions API is one more data point in a trend that has been building all year: AI vendors are no longer competing only on how good their flagship chat model is. They are competing on the surrounding infrastructure, the pricing tiers, the specialized endpoints, and the developer experience that determines whether a team builds its next product on top of their platform or a rival’s. Anthropic’s aggressive Haiku pricing, DeepSeek’s API opening, OpenCode’s 75-model access layer, and now OpenAI’s Decisions API are all moves in the same game: lower the friction and cost of building AI-native software badly enough that developers do not look elsewhere.

Whether the Decisions API becomes a default piece of the agent development stack or a niche tool used by a handful of teams will depend on something that cannot be settled from a launch announcement: whether the typed outputs and claimed speed hold up once real production traffic hits the endpoint at scale. That verdict is still months away, but the direction of the bet, specialized, cheap, fast infrastructure underneath the agent boom, is clear enough today.

Frequently asked questions

What is the OpenAI Decisions API?
It is a dedicated endpoint, POST /v1/decisions, that OpenAI opened to all developers in public beta on October 6, 2026. It evaluates text or images and returns a typed answer: a yes-or-no predicate, a pick from a fixed list of options, or a score against a rubric, instead of a free-form chat response.

How much does the Decisions API cost?
The base price is $0.10 per 1 million input tokens using the gpt-6-luna model, with no separate output token charge. Reporting from Mixed News indicates inputs above 272,000 tokens are billed at double the base rate, and US or EU regional processing adds a 10 percent surcharge, though those modifiers are not spelled out in OpenAI’s own published guide.

Which model powers the Decisions API?
Only gpt-6-luna is supported during the public beta. There is no multi-model selection available yet.

Is the Decisions API faster than OpenAI’s other APIs?
OpenAI claims it returns answers roughly 10 times faster than running the same evaluation through the Responses API. The company has not published a detailed benchmark methodology behind that figure, so it should be treated as an unverified product claim pending independent testing.

What can developers use the Decisions API for?
OpenAI points to content classification, request routing, work prioritization, and selecting which model, tool, or action an AI agent should take next as the primary use cases, all of which are bounded decisions rather than open-ended generation tasks.

Does the Decisions API support image inputs?
Yes, but only as inline base64-encoded data. Hosted HTTP or HTTPS image URLs and file IDs are not supported in the public beta.

Is there a competing product from Google, Anthropic, or Microsoft?
No directly equivalent public decision endpoint from Anthropic or Google has been confirmed as of this launch. Microsoft has a separate product called Decision-1 aimed at scoring choices in enterprise workflows, which overlaps conceptually but is not an identical offering.

What are the biggest risks of relying on the Decisions API?
The main risks are over-trusting probability scores without independent calibration, running automated decisions at high volume without human review since the low price removes a natural cost brake, and building core application logic around a single-model beta that could change before it reaches general availability.

OpenAI Decisions API Opens to All Devs at $0.10 per 1M Tokens [2026]