Microsoft Launches Decision-1 as H2O.ai Challenges Its Speed Comparison

Microsoft launched Microsoft-Decision-1 on October 9, with a self-reported median response time of 85 milliseconds and a speed comparison that rival H2O.ai disputes. The model assigns probabilities to fixed answer options in a single pass instead of generating text. It is intended to reduce delay in tasks such as routing support requests and checking an agent’s proposed action.

Fixed-option scoring

Decision-1 is built on Alibaba’s Qwen3.5-9B, with additional training by Microsoft. Developers supply the input, a question and possible answers. The Foundry catalog lists uses including routing support requests and grading an agent’s proposed action against supplied criteria.

The model is available through Microsoft Foundry and OpenRouter. At launch, input costs $0.042 per million tokens, with output free. Applications can use the returned probabilities to set thresholds for automatic action or human review, and include a “cannot tell” option when evidence is insufficient.

The disputed speed comparison

Microsoft’s launch comparison put Decision-1’s median response time at 85 milliseconds, against 210 milliseconds for H2O-Lightning-4B. Its chart footnote identifies its own result as a same-region Foundry measurement and rivals’ results as adjusted JevBench medians checked October 7. Microsoft also claimed it was 35 times faster than GPT-6 Sol in its own benchmarking.

H2O.ai, the developer of a competing decision model, challenges the treatment of self-hosted rivals. It says Microsoft used JevBench’s network-adjusted estimate rather than raw local execution time. JevBench measured H2O-Lightning-4B v1.1 at about 29 milliseconds locally; doubling that median and adding 150 milliseconds produces approximately the figure in Microsoft’s chart.

JevBench times hosted APIs from its measuring host through the network and back. Self-hosted models are timed on-GPU, without network delay. Their raw measurements cannot be directly compared with API response times.

JevBench applies the same estimate to every self-hosted row. The adjustment is an assumption, rather than a measured network cost or hardware normalization. Its method notes also describe a planned calibration of the network allowance against real end-to-end measurements.

An independent Foundry measurement

On October 10, JevBench recorded a native Azure Foundry median of 459 milliseconds for Decision-1 across a 124-request speed subset. JevBench timed a network-inclusive API round trip, while Microsoft measured within the same Foundry region.

The independent measurement does not reproduce Microsoft’s original serving setup or its full benchmark suite.

Microsoft’s quality results

In its launch evaluation, Microsoft reported the highest accuracy across 36 benchmarks covering nearly 150,000 questions, ahead of models including Jev, Quyet-1.0-Large and H2O-Lightning-4B. The company says those benchmarks were kept blind from training and covered tasks including routing, long-context decisions and safety.

JevBench’s independent evaluation uses its own observations and scoring method, so it does not verify Microsoft’s claim to lead that separate comparison.

Microsoft also reported an internal Xbox Research trial that sorted more than 10,000 pieces of player feedback and reviews into themes defined by researchers. In that trial, the company reported quality competitive with GPT-6 Sol while running more than 14 times faster.

Limits and future versions

Decision-1 provides no written rationale for its scores. The catalog excludes using it as the sole automated decision-maker for consequential personal decisions, including credit, employment and healthcare. Confidence calibration is strongest on familiar task types, and wording or option order can change scores.

Microsoft’s deployment guidance calls for testing representative workloads and validating confidence thresholds on labeled examples before automating decisions. It also advises placing Foundry resources near applications and users to reduce network delay.

Microsoft says it plans to soon rebase this model on other model families, including its own MAI models and OpenAI models.

Musk Plans to Use Claude Opus 5.5 and Other Services in Grok Bot

Elon Musk announced on October 7 plans for SpaceXAI's Grok Bot to use external AI providers to handle work for customers. The product would select a provider by task, with Claude Opus 5.5 and outside

The Implicator

Reflection AI Unveils Beam, an Open-Weight Model It Says Matches China's GLM-5.2

“DeepSeek and Qwen and all these models are our wake-up call because if we don't do anything about it, then effectively, the global standard of intelligence will be built by someone else,” Misha Laski

The Implicator

The plumbing war: How enterprises are building the routing layer for multi-model AI

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai