Fastino GLiDE brings adaptive reasoning to agent decisions
Fastino GLiDE combines adaptive reasoning with structured decisions. Its September 30 release brings new routing options and practical limits for developers.
Fastino introduced GLiDE on September 30, adding a model that can spend extra reasoning on uncertain decisions before returning a structured answer. The launch announcement positions it for tasks such as choosing an agent tool, routing a request or evaluating a proposed action. GLiDE is available through the Fastino API.
The practical distinction is where the reasoning goes. An application supplies the situation and permitted outcomes. GLiDE evaluates that choice internally, then returns probabilities and a decision that software can read directly.
Defined choices with room for uncertainty
The API reference describes three question types. Noul answers a yes or no question with a probability. Choice selects among named options. Score evaluates an ordered set of levels, returning both a winning level and a probability weighted expectation.
Each question can use up to 40,000 tokens of context, including the supplied state. Choice and Score support as many as 255 outcomes. Multiple questions can share one state in a request, but each asks for a single answer.
Confidence also has a specific meaning. For Choice and Score it measures the gap between the top two probabilities. Fastino recommends tuning action thresholds on representative data instead of treating a default value as the right tolerance for every business.
Reported benchmark gains need their context
Fastino reports a Decision Index 0.2.1 score of 64.81, compared with Jev’s published 57.91. It says GLiDE leads across all five evaluation areas and 31 of 38 benchmarks. The company attributes the results to a design that starts with a quick assessment and allocates more reasoning when the decision is uncertain.
These are Fastino’s own results using the official scorer. The announcement says GLiDE has not been added to the public leaderboard, where submissions to that version are paused. The figures provide a starting point for evaluation rather than evidence that an application will avoid routing mistakes.
Cost and integration details matter
Fastino’s current price list charges $0.15 per million input tokens for GLiDE and no charge for output tokens. Developers should measure usage on their actual requests. The reference says reported input usage sums internal passes across questions and can exceed the context limit for an individual question.
The inference guide also documents limits that affect implementation. Requests cannot stream or contain a batch of independent states. Teams migrating from Jev need to update score handling because GLiDE’s winning level and expected level are separate fields.
A useful trial would keep existing actions unchanged while comparing GLiDE decisions with reviewed examples. Measure mistaken selections, uncertain cases, latency and billed usage together. A structured answer makes integration easier, but the team still owns the rule that decides when to act.
Our coverage of Cloudflare Clef and Clef Flash looks at another approach to bounded agent decisions. GLiDE adds adaptive reasoning to that growing category, with the practical tradeoffs best tested on the decisions a team actually needs to make.
Illustrative server racks photographed by alq666 on Wikimedia Commons, used under Creative Commons Attribution ShareAlike 2.0. Reduced resolution rendition. The photograph and resized versions retain that license. It does not show a Fastino facility.
