---
type: "post"
id: "subq-snap-1-1-ties-for-first-on-b63bf5"
slug: "subq-snap-1-1-ties-for-first-on-b63bf5"
title: "SubQ Snap 1.1 ties for first on AADI, with 384K context and a 20% price cut"
description: "Subquadratic released SubQ Snap 1.1, a post-training update to its decision model that retains a 384K-token context window while reducing input pricing by 20%."
retrieval_nugget: "Subquadratic released SubQ Snap 1.1, a post-training update to its decision model that retains a 384K-token context window while reducing input pricing by 20%. It tied for first on Artificial Analysis’s Decision Index, but the result reflects abstention and calibration rather than leading raw accuracy—a distinction that matters when choosing a model for routing or classification."
published_at: "2026-11-04"
updated_at: "2026-11-07"
record_date: "2026-11-04"
date_kind: "published_at"
topics: ["agent-interfaces"]
entities: ["Subquadratic","SubQ Snap 1.1","SubQ Snap 1.0","Artificial Analysis","d1-ultra","Elephant 9B","Jev 1.14"]
related_patterns: ["model-answers-become-task-interfaces"]
novelty_label: "same_event"
prior_refs: []
source_url: "https://www.orcarouter.ai/blog/subq-snap-1-1-decision-index-first"
source_urls: ["https://www.orcarouter.ai/blog/subq-snap-1-1-decision-index-first"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:subq-snap-1-1-ties-for-first-on-b63bf5"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/subq-snap-1-1-ties-for-first-on-b63bf5/","markdown":"https://newruntime.com/posts/subq-snap-1-1-ties-for-first-on-b63bf5.md","json":"https://newruntime.com/posts/subq-snap-1-1-ties-for-first-on-b63bf5.json"}
markdown_url: "https://newruntime.com/posts/subq-snap-1-1-ties-for-first-on-b63bf5.md"
---

# SubQ Snap 1.1 ties for first on AADI, with 384K context and a 20% price cut

## Retrieval answer

Subquadratic released SubQ Snap 1.1, a post-training update to its decision model that retains a 384K-token context window while reducing input pricing by 20%. It tied for first on Artificial Analysis’s Decision Index, but the result reflects abstention and calibration rather than leading raw accuracy—a distinction that matters when choosing a model for routing or classification.

## Summary
Subquadratic released SubQ Snap 1.1, a post-training update to its decision model that retains a 384K-token context window while reducing input pricing by 20%. It tied for first on Artificial Analysis’s Decision Index, but the result reflects abstention and calibration rather than leading raw accuracy—a distinction that matters when choosing a model for routing or classification.

## What happened
Subquadratic announced SubQ Snap 1.1 on October 22, 2026. The source describes it as a post-training release, not a new model: the Subquadratic Sparse Attention backbone, 384K-token context window, and text-plus-image input path remain unchanged from Snap 1.0. The model is designed for bounded decision tasks, where an input is matched to a fixed set of allowed answers—for example, routing to a queue, deciding whether to call a tool, or choosing whether to escalate.

The change in version 1.1 is how the model handles uncertainty. Subquadratic says it added explicit uncertainty calibration and distils multi-step reasoning into a single forward pass, rather than generating a chain of thought before committing to an answer. The company reports that expected calibration error on held-out probability benchmarks fell from 0.108 to 0.014. On unanswerable questions, it reports that confident errors fell from 294 out of 600 to 57 out of 600, an 81% reduction. These are vendor-reported results in the supplied source, not a guarantee of equivalent performance on a buyer’s own data.

Artificial Analysis’s Decision Index v1.0 gave Snap 1.1 a score of 65, tying it with d1-ultra and Elephant 9B for first place. The same index write-up, as described in the source, gives Snap 1.1 accuracy of 20%, compared with 32% for d1-ultra and 12% for Elephant 9B. Its leading position therefore should not be read as a claim that it answers more questions correctly. The source attributes the distinction to its stronger willingness to abstain when uncertain. Jev 1.14 scored 57, eight points lower.

Snap 1.0’s preview went live on October 12 at $0.05 per million input tokens. Snap 1.1 is listed at $0.04 per million input tokens, with output tokens free; the source characterizes this as a 20% input-price cut. The supplied material says the product page and API documentation list Snap 1.1 as the default public-preview decision model at that price. It does not provide further details about service-level guarantees, rate limits, or the evaluation protocol beyond the figures summarized above.

## What the archive adds
[SubQ Snap’s earlier product proposition: bounded agent decisions over longer evidence](https://newruntime.com/posts/this-ai-lab-takes-on-jev-with-12x-0746e5/) is directly related. That earlier Field Note discussed the model’s long-context positioning; this release adds a specific calibration claim, a Decision Index result, and updated pricing. This is a new release on a known story, not a restatement.

## Trend signal
This supports, narrowly, [model answers becoming task interfaces](https://newruntime.com/patterns/model-answers-become-task-interfaces/): Snap is presented as returning a choice from a fixed answer set, with a probability and the option to abstain, rather than producing an unrestricted chat response. The source describes a model capability, not evidence that this interface pattern is becoming standard across deployed products.

## Why now
The release follows Snap 1.0’s October 12 preview. Subquadratic’s announcement makes a specific change in uncertainty behavior measurable and pairs it with a lower input price. That creates a timely reason for teams already considering the preview to reassess its price and test its decision behavior, without implying that the update has been independently validated for their use case.

## Why it matters
A model used to route work or choose among tools can cause operational cost not only by making a wrong choice, but also by making that choice confidently when the input does not support it. A system that abstains more reliably could make escalation or human review easier to design around. That is the practical hypothesis behind Snap 1.1’s reported calibration and unanswerable-question results.

The price change is also concrete: teams pay less for input tokens than at the Snap 1.0 preview rate, while output tokens are listed as free. But the Decision Index tie is not a general-purpose quality ranking. The reported 20% accuracy should discourage teams from choosing the model solely because it shares the top score. It is most relevant to builders with bounded decisions who can define acceptable errors, abstentions, and escalation paths, then measure those outcomes on representative examples. The source does not establish that the model improves a complete workflow or reduces its total cost.

## Limits and uncertainty
The supplied source reports benchmark scores and calibration figures but does not establish how the evaluation sets were assembled, how well they represent production traffic, or whether results were independently reproduced. It also does not show comparative performance on a buyer’s own routing or tool-selection tasks. A Decision Index tie should not be treated as evidence of equal accuracy: the listed accuracy values differ materially. The material confirms public-preview availability and token pricing, but does not specify rate limits, availability guarantees, or other usage terms.

## Reader takeaway
SubQ Snap 1.1’s notable claim is not that it gets the most answers right; it is that it can be more calibrated about when not to answer. Its 384K context and lower input price may make it worth testing for fixed-choice decisions, but teams should compare accuracy, abstention, and confident errors on their own data before relying on it.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Interfaces](https://newruntime.com/topics/agent-interfaces.md): Explore the agent-interfaces topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)