---
title: "GPT-6 Sol is 50% cheaper. That makes routing discipline more important, not less."
description: "OpenAI cut Sol API prices in half while AWS brought GPT-6.1 Sol to Bedrock. Teams should optimise for cost per accepted outcome, not default every workload to the newest model."
url: "https://devencelab.com/insights/2026/09/29/gpt-6-sol-is-50-cheaper-that-makes-routing"
date: "2026-09-29"
section: "Insights"
tag: "Model Releases"
author: "Devence Lab"
reading_time: "2 min read"
site: "Devence Lab"
license: "Readable and quotable with attribution to the canonical URL."
---

# GPT-6 Sol is 50% cheaper. That makes routing discipline more important, not less.

OpenAI cut Sol API prices in half while AWS brought GPT-6.1 Sol to Bedrock. Teams should optimise for cost per accepted outcome, not default every workload to the newest model.

OpenAI prices GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, 50% below GPT-5.6 Sol promotional pricing. AWS made GPT-6.1 Sol generally available on Amazon Bedrock on 29 September for coding, computer use and professional workloads. The deployment lesson is not simply that frontier reasoning got cheaper. Lower unit cost makes poor routing easier to hide.

OpenAI still positions GPT-6 Astra as its strongest model. Sol is the model for teams that need high capability repeatedly without paying the Astra premium on every call. That makes routing policy more important, because the cheapest successful model for a task is not necessarily the cheapest model per token.

## Route on accepted outcomes, not benchmark rank

OpenAI reports GPT-6 Sol at xhigh effort scoring 33.2% on AutomationBench at $0.27 per task, while its published comparison places Claude Opus 5 at 26.9% and 11.1 times the cost per task. Those are vendor evaluation results, not a production guarantee, but the metric is the useful part: cost per completed task captures both capability and spend.

Production routing should copy that unit of measurement. Define task classes, record the selected model and reasoning effort, then measure accepted completion, retries, latency and human escalation. A low-cost call that needs three attempts is not cheap. A stronger model that completes a high-consequence task once may be.

> The routing target is not the cheapest token. It is the cheapest accepted outcome at the required risk level.

## Reasoning effort belongs in the budget

GPT-6 Sol's published results vary by reasoning effort, which means model selection alone is an incomplete control. Teams should treat reasoning depth as part of the route: routine extraction and classification can start lower, while ambiguous planning, code changes or consequential analysis can earn more compute.

Log reasoning effort beside the task result. Then establish escalation rules from evidence: repeated failure, low confidence, policy-sensitive actions or expensive downstream consequences can move a task upward. The default should be justified by the workload, not inherited from a model launch.

## Keep the control plane stable when models change

Bedrock availability gives AWS teams a familiar deployment path, but a new model should not create a new governance exception. Identity, data handling, logging, tool permissions and approval boundaries should stay attached to the workload while the underlying model changes.

The practical rollout is a controlled substitution test. Pick representative production tasks, run GPT-6 Sol against the current route, compare accepted outcomes and total cost, and promote it only where the evidence holds. Cheaper frontier intelligence is useful because it expands the routing options. It does not remove the need to route.

## Sources

- [Bring near-Astra intelligence to everyday work with GPT-6.1 Sol on Amazon Bedrock](https://aws.amazon.com/blogs/machine-learning/bring-near-astra-intelligence-to-everyday-work-with-gpt-6-1-sol-on-amazon-bedrock/) - AWS Machine Learning Blog
- [Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) - OpenAI