OpenAI’s GPT-6 Sol and Luna Put Model Efficiency at the Center

OpenAI introduced GPT-6 Sol and GPT-6 Luna as additional tiers in its GPT-6 family. [12] The company says both models were trained with methods similar to GPT-6 Astra, extending Astra-related advances in professional work, factuality, coding, computer use, and alignment to faster and more affordable models. [1] OpenAI attributed the lower serving costs to improvements in caching and inference. [1]

GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. [1][4][9] An OpenAI spokesperson told VentureBeat that these are permanent rather than promotional or introductory prices. [9] OpenAI said Sol and Luna are available through its API and are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. [5]

AWS describes Sol as a daily model for recurring complex tasks and software development, including feature implementation, debugging, code review, refactoring, data analysis, and multistep workflows. [8] Luna is positioned for focused, high-volume work such as extraction, summarization, classification, and routing. [8][9] Both models support context windows of up to one million tokens. [8]

On an internal OpenAI factuality evaluation, Sol made roughly half as many mistakes as GPT-5.6 Sol. [8] Luna offers adjustable reasoning effort so developers can balance quality, speed, and cost for each request. [8] OpenAI continues to position Astra as its strongest model overall and recommends it for projects requiring the best results without compromise. [1]

Anthropic released Claude Opus 5.5 on the same day, targeting long-running coding agents, research, and professional knowledge work. [2][10] Its API pricing is $4 per million input tokens and $20 per million output tokens, 20% below Opus 5’s per-token rates. [10] Anthropic says typical workloads cost about 40% less overall because Opus 5.5 also completes tasks with fewer tokens. [10][12]

Anthropic reduced cache-write pricing from $6.25 to $5 per million tokens and cache-read pricing from $0.50 to $0.20 per million tokens. [10] AWS says Opus 5.5 uses adaptive thinking on every request, deciding how much reasoning effort a task requires. [7][11] The model also includes safety classifiers similar to Claude Fable 5.1 for biology, cybersecurity, and AI development, with more frequent refusals than previous Opus versions. [7]

On OpenAI’s AutomationBench test of business workflows across applications, GPT-6 Sol at xhigh effort outperformed Claude Opus 5 at max effort while costing 9% as much per task. [1] VentureBeat reported that Claude Opus 5.5 exceeded Fable 5.1 on several agentic coding, knowledge-work, scientific, and multidisciplinary reasoning benchmarks, despite Fable 5.1 carrying a higher API price. [10] Some of the cited results are vendor-selected customer results rather than independent evaluations, while OpenAI also describes one assessment as an “internal factuality evaluation.” [1][10]

GPT-6 Sol and Luna became generally available through Amazon Bedrock on September 22, joining Astra in the service’s GPT-6 lineup. [8] Claude Opus 5.5 is also available through Amazon Bedrock and Claude Platform on AWS, with Bedrock providing zero-data-retention support by default, regional data residency, Guardrails, and Knowledge Bases. [11] CNBC reported that both labs face pressure from lower-cost open-weight competitors while customers seek to restrain AI spending. [12]

OpenAI and Anthropic are both turning efficiency into a product feature: cheaper tokens, cheaper cache reads, and models tuned for recurring or long-running work. GPT-6 Sol is priced below Claude Opus 5.5, and OpenAI says Sol beat Claude Opus 5 on AutomationBench at a fraction of the task cost, while Luna occupies a much cheaper high-volume tier. For practitioners, the releases make model routing more consequential because Astra, Sol, Luna, and Opus 5.5 offer distinct balances of capability, token consumption, and price. What the sources don't address: how these models compare in independent production evaluations after accounting for latency, retries, and task-specific failure rates.