> ## Content Index
> Fetch the complete content index at: https://bytevyte.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Price War Erupts as OpenAI Halves GPT-6 Sol and Luna API Costs Within 90 Minutes of Claude Opus 5.5
- URL: https://bytevyte.com/price-war-erupts-as-openai-halves-gpt-6-sol-and-luna-api-costs-within-90-minutes-of-claude-opus-5-5/
- Published: 2026-09-23T17:46:43.000Z
- Updated: 2026-09-23T17:46:43.000Z
- Description: OpenAI cut GPT-6 Sol and Luna API prices by 50% about 90 minutes after Anthropic shipped Claude Opus 5.5, deepening the AI price war over cost per task.
- Author: Bytevyte Editorial
- Tags: ai-beats

OpenAI has halved the list price of its newest models, launching **GPT-6 Sol and Luna** at roughly 50% below the promotional rates of the GPT-5.6 generation they extend. Sol enters at $2 per million input tokens and $10 per million output tokens; Luna lands at $0.10 and $0.50\. The release came on September 22, about 90 minutes after Anthropic shipped Claude Opus 5.5, and both new tiers are available on Amazon Bedrock alongside the existing Astra model.

Two frontier labs cut costs inside the same hour, and neither led with a benchmark record. Both led with price per token and price per finished task, which is a different contest from the capability race the industry has run for three years.

## What OpenAI Shipped

Sol occupies the middle rung of what is now a three-tier GPT-6 family, sitting below Astra and above Luna. OpenAI positions it as the daily model for recurring complex work: software development, multi-step agentic coding, long analytical tasks and automation pipelines that a team runs continuously rather than once a quarter. On an internal factuality evaluation, OpenAI reports that Sol makes roughly half as many mistakes as GPT-5.6 Sol.

Luna targets the other end of the workload curve. It handles focused, high-volume jobs such as summarisation and request classification, and it exposes adjustable reasoning effort so a team can spend more compute on hard queries and less on routine ones. Both models accept context windows of up to 1 million tokens, the ceiling enterprise retrieval pipelines have come to expect from a flagship API.

Astra stays at the top of the family, and this launch repriced only the two new tiers, leaving OpenAI selling three capability levels against Anthropic's single new Opus release. The tiering gives procurement teams a cleaner way to segment spend by workload, and it gives OpenAI room to discount the middle and bottom without devaluing its frontier model.

OpenAI attributes the reduction to gains in caching and inference efficiency that it says are being passed through to customers rather than held as margin. Cache pricing is where the two vendors diverge. OpenAI's cut applies to the published input and output rates for both new models, while Anthropic layers a lower cache-read rate on top of a model that already consumes fewer tokens. For workloads dominated by repeated context, such as a code assistant that re-reads a large repository on every turn, the second approach can produce a bigger effective reduction than the headline percentage suggests.

On Amazon Bedrock, both models arrive with AWS security controls, governance tooling and auditing features. That packaging matters for regulated buyers, who frequently reach frontier models through a cloud provider rather than a lab's own endpoint.

Context size and price interact in a way the headline rates hide. A 1 million token window is only useful if feeding it is affordable, and at $0.10 per million input tokens Luna makes bulk document processing routine where it previously required batching and sampling. Sol's rates matter more for agentic coding, where a single task can consume hundreds of thousands of tokens across planning, editing, testing and retrying. The pricing floor now determines how ambitious an agent architecture a team can afford to run in production.

## GPT-6 Sol and Luna Pricing, Compared

The published rates put Sol at half of Claude Opus 5.5's uncached input and output pricing.

| Model      | Input ($/1M tokens) | Output ($/1M tokens) | Change vs GPT-5.6 |
| ---------- | ------------------- | -------------------- | ----------------- |
| GPT-6 Sol  | 2.00                | 10.00                | about -50%        |
| GPT-6 Luna | 0.10                | 0.50                 | about -50%        |

OpenAI also states that Sol outperforms Claude Opus 5 on the AutomationBench business evaluation while using about 9% of Opus 5's cost per completed task. That reframes the comparison from cost per token to cost per unit of finished work, the metric finance teams budget against. Both figures are vendor-reported rather than independently verified.

## Anthropic's Counter: Fewer Tokens, Not Just Cheaper Ones

Anthropic's answer is structural rather than a headline discount. Claude Opus 5.5 is the first release in the Claude 5.5 family, and Anthropic says it completes tasks using fewer tokens than Claude Opus 5 while charging less per token, with cheaper cache reads stacked on top of that gain. The combined effect is a smaller bill for the same job without a straight list-price cut.

The model adds adaptive thinking, which scales reasoning effort to the complexity of the task at hand. Anthropic positions it as a collaborator for long-running coding and knowledge work that reports back on what it did, what it found and what it still needs. On AWS it ships through Amazon Bedrock with zero data retention by default, through Claude Platform on AWS for a native experience, and it is available in AWS GovCloud (US) for public-sector and regulated buyers.

Anthropic puts Opus 5.5 at 20% cheaper per token than Opus 5 before cache savings are counted. That is a smaller headline cut than OpenAI's 50%, but token efficiency and cache pricing can close the gap on workloads with long, repeated context, where most of the bill comes from reading the same documents again and again.

## Why the Cuts Landed Now

These are the first model launches from either lab since their leaders publicly argued for a slower pace of AI development. The rhetoric has not changed the release cadence, and it has not changed the direction of pricing.

Pressure is arriving from two directions. Cheaper open-weight models from Chinese labs have reset buyer expectations about what capable inference should cost, and enterprise demand has shifted toward the mid-tier models that carry most production traffic. xAI's Grok 4.7 shipped the same week, adding a third vendor to the same price bracket.

Both launches are bets that most enterprise traffic will sit in the cheap tier. Routing layers send the bulk of requests to a small, inexpensive model and escalate only hard queries upward, which means cuts at the bottom of the stack move total spend more than cuts at the top. On that logic, Luna's $0.50 output rate is the more consequential number of the two releases.

The 90-minute gap between the announcements is itself information. Anthropic set a new price and capability reference for the Opus line, and OpenAI matched and undercut it before the news cycle closed. Whatever the two companies say about development pace, their product calendars are now indexed to each other, and buyers should expect list prices to be reset whenever either lab ships.

For enterprise buyers, the GPT-6 Sol and Luna pricing change resets the floor on the cost of running an agent. Routing layers that send routine classification and summarisation to a cheap model and escalate only hard queries to a frontier tier get cheaper when the cheap tier costs $0.10 per million input tokens. The same arithmetic applies to retrieval pipelines that repeatedly feed long documents into context.

Committed-spend agreements signed against 2025 pricing are the immediate exposure. A team that locked annual volume discounts on GPT-5.6 or Opus 5 rates should re-run unit economics against the new list prices, because a 50% cut to the reference rate changes what a negotiated discount is worth. Internal chargeback models face the same problem: a platform team billing business units at last year's per-token rate is now over-recovering by a wide margin.

The unresolved question is whether 50% is a permanent reset or a land grab. If the reductions reflect durable inference and caching efficiency, they hold. If they are being funded by capital raised to buy developer share, list prices could firm once the contest over default model choice settles.

## Why this matters

The GPT-6 Sol and Luna cuts and Anthropic's token-efficiency answer point to the same conclusion: capability is no longer the scarce input in an enterprise AI budget, cost per completed task is. Buyers who once negotiated around benchmark leadership now have two vendors competing on unit economics, which strengthens their position at renewal. The larger consequence is that cheaper inference changes which applications are worth building at all, and that shift will outlast this week's release order.

## Sources

[OpenAI GPT-6 Sol and GPT-6 Luna are now generally available on Amazon Bedrock](https://aws.amazon.com/about-aws/whats-new/2026/09/openai-gpt-6-sol-luna-on-amazon-bedrock/?ref=bytevyte.com)

[Claude Opus 5.5 is now available on AWS](https://aws.amazon.com/about-aws/whats-new/2026/09/claude-opus-5-5-aws/?ref=bytevyte.com)

## Related Articles

- [Grok 4.5 Price War: SpaceXAI Undercuts Claude Opus by 75% to Dominate Coding Agents](https://bytevyte.com/grok-4-5-price-war-spacexai-undercuts-claude-opus-by-75-to-dominate-coding-agents/)
- [GPT-5.6 Sol price cut: OpenAI's 3-month defense](https://bytevyte.com/gpt-5-6-sol-price-cut-openais-3-month-defense/)
- [Anthropic's Claude Sonnet 5 Puts Opus-Class Agentic Power Within Enterprise Reach at $2/M](https://bytevyte.com/anthropics-claude-sonnet-5-puts-opus-class-agentic-power-within-enterprise-reach-at-2-m/)

✔Human Verified

---

*Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*