Tech • AI • Robotics • Game
OpenAI Decisions API launches as Claude $200 users hit weekly token limits
Briefing
OpenAI debuts Decisions API
OpenAI introduced a Decisions API aimed at software workflows that need a direct machine-actionable signal rather than generated prose. The product is designed for routing, classification, prioritization and other business logic tasks where probabilities matter more than explanations. It evaluates both text and images, widening its use across support, moderation and operations stacks. The launch marks a clear product split between generation-first interfaces and decisioning-focused infrastructure.
GPT-6 Luna powers decisioning
The new API currently runs on a single model, GPT-6 Luna. OpenAI says it supports three output formats: predicate, score and choice, covering binary checks, ordered ratings and fixed-option selection. That structure maps closely to common enterprise automation patterns such as spam detection, urgency triage and ticket routing. By constraining outputs, the system is built to be easier for applications to consume reliably.
OpenAI promises 10x faster calls
Because the Decisions API skips text generation, OpenAI positions it as roughly 10 times faster than the Responses API for simple evaluation tasks. The company also removes output-token charges, a notable shift for teams running large volumes of lightweight checks. Pricing is listed at $0.10 per million input tokens, making small classification jobs extremely cheap on paper. That cost profile could make the API attractive for high-frequency backend workflows.
Predicate, score and choice split workloads
The API's three decision types suggest a deliberate attempt to standardize model outputs for automation. Predicate estimates whether a single statement is true, score ranks an input across ordered levels, and choice selects the best option from a predefined list. In practice, that means developers can use one endpoint for tasks as different as image labeling, policy checks and customer-support triage. The format reduces the need for brittle prompt parsing and post-processing.
Claude users confront weekly caps
On the consumption side, Claude users on the $200 subscription tier are being pushed to manage usage around weekly rather than session-based limits. Session caps may reset within hours, but weekly all-model limits can halt work for days once exhausted. A practical pacing rule cited is about 14% per day across a seven-day cycle. The emphasis reflects how flat-fee AI plans increasingly depend on token budgeting discipline rather than headline access alone.
Tokens, not context, drive spend
The main pressure on premium Claude usage comes from token volume, especially generated output, rather than simply from large context windows. Output is costly because reasoning-heavy operation produces substantial hidden generation under the hood. That shifts optimization away from obsessively trimming input context and toward reducing unnecessary model deliberation. For many users, the biggest savings come from choosing lighter settings for routine tasks.
Fable emerges as quota killer
Among the models discussed, Fable stands out as the most expensive option for everyday use. Its pricing is cited at about $10 per million tokens, roughly 2.5 times the cost of Opus on both input and output. That makes it poorly suited to routine drafting, coding and chat if preserving weekly allowance is the goal. The implication is that premium reasoning models remain best reserved for narrow, high-value moments.
Sonnet becomes the practical default
For regular work, Sonnet is presented as the budget-conscious baseline within Claude. Pricing is cited around $2 input and $10 output per million tokens, making it materially cheaper than heavier alternatives. Users are also advised to avoid high-reasoning modes unless the task truly requires them, because those settings consume quota quickly. The broader takeaway is that model selection and reasoning level now function like resource-allocation decisions, not just quality preferences.
Videos covered
Decisions API Full Tutorial for Beginners: How to Use & Set up Decisions API
- •A shift from generation to decisioning
- •Faster and cheaper for simple workflows
- •One model, three decision types
Comment AVOIR Claude en illimité et arrêter de cramer tes limites
- •Weekly limits matter more than session caps
- •Tokens drive the real cost
- •Fable is the costliest option