GPT-6 Luna Benchmarks: Official Scores and Real-World Limits

OpenAI introduced GPT-6 Luna alongside GPT-6 Sol to expand the GPT-6 model tier with a lower-cost option designed for high-volume tasks. The release establishes GPT-6 Luna as a lightweight model priced at $0.10 per million input tokens and $0.50 per million output tokens, which reduces Application Programming Interface (API) rates by 50% compared to GPT-5.6 Luna. The primary keyword focus in technical evaluations centers on the GPT-6 Luna benchmarks, which document specific gains on multi-app workflow tests and internal factual error tracking.

Unlike full-sized flagship systems such as GPT-6 Astra or Claude Opus 5 from Anthropic, GPT-6 Luna targets high-frequency, budget-constrained operations rather than unbounded complex reasoning. On the AutomationBench 1.0.6 evaluation, OpenAI reports that GPT-6 Luna at high effort improves on its predecessor by 5.4 percentage points while lowering cost per task by 58%. Additionally, on internal factual error assessments using de-identified user conversations, OpenAI reports that GPT-6 Luna at higher effort levels matches the factual reliability of the larger GPT-5.6 Sol model.

For small and mid-sized business (SMB) teams running structured automation pipelines, these numbers provide a specific trade-off between execution speed, accuracy, and operational overhead. Operators evaluating whether to transition legacy classification, extraction, or customer support pipelines to GPT-6 Luna must weigh the published evaluation gains against the inherent gap between synthetic agent test suites and live, messy production environments.

Official GPT-6 Luna Benchmarks and Published Metrics

OpenAI published specific benchmark figures for GPT-6 Luna across workflow execution, factual precision, and API pricing. The released documentation centers primarily on AutomationBench 1.0.6 and an internal factuality evaluation based on historical error logs from user conversations.

On AutomationBench 1.0.6, which evaluates artificial intelligence (AI) agents across 47 tools spanning sales, marketing, operations, customer support, finance, and human resources (HR), GPT-6 Luna at high effort scored 5.4 percentage points higher than GPT-5.6 Luna. OpenAI also reported that this configuration lowered the direct cost per task by 58% compared to the predecessor model.

The vendor did not publish standalone percentage scores for GPT-6 Luna on the Agents' Last Exam evaluation in this release note, reserving that metric for GPT-6 Sol, which scored 56.4% at maximum effort. For factual accuracy, OpenAI evaluated de-identified conversations where users previously flagged factual mistakes; under higher reasoning effort configurations, GPT-6 Luna matched the factual reliability of GPT-5.6 Sol while running at approximately one-hundredth of the computing cost.

• AutomationBench 1.0.6 performance: A gain of 5.4 percentage points over GPT-5.6 Luna at high effort settings.
• Cost per task reduction: A 58% decrease in cost per task compared to GPT-5.6 Luna on the same workflow test suite.
• Factuality evaluation: Parity with GPT-5.6 Sol factual reliability during targeted error-inducing test prompts at higher effort levels.
• Standard API token pricing: $0.10 per million input tokens and $0.50 per million output tokens, representing a 50% reduction from promotional pricing.

API Pricing and Unit Economics for GPT-6 Luna

Published evaluation results must be weighed against input and output token expenses to calculate real operational return on investment (ROI). OpenAI cut the token rates for GPT-6 Luna by half relative to the promotional pricing of the prior generation.

The current published rate stands at $0.10 per million input tokens and $0.50 per million output tokens. In comparison, the mid-tier GPT-6 Sol runs at $2.00 per million input tokens and $10.00 per million output tokens, making GPT-6 Luna twenty times cheaper on input and output processing.

Because prices and token tier limits can change without notice, developers and system architects should verify current rates directly on the official OpenAI pricing page before deploying large-scale pipelines. When calculating total cost, teams must also account for the token overhead incurred when setting GPT-6 Luna to higher effort levels to achieve the benchmarked factual accuracy numbers.

How GPT-6 Luna Benchmarks Map to Business Tasks

Benchmark scores only provide business value when engineers understand which synthetic tests correspond to specific daily office operations. High scores on synthetic tool-use suites do not guarantee that an agent handles messy input files or incomplete database records without failure.

The AutomationBench test suite measures how effectively an LLM issues function calls, reads schemas, and updates simulated records across business software categories like customer relationship management (CRM) software and enterprise resource planning systems. A 5.4 percentage point gain on AutomationBench indicates improved consistency when parsing structured fields, routing tickets, or populating standard tables.

Similarly, the internal factuality metric predicts performance on retrieval-augmented generation (RAG) tasks, such as querying an employee handbook or summarizing vendor contracts. When GPT-6 Luna matches GPT-5.6 Sol on factual retention at higher effort levels, it suggests that automated document summarization pipelines will experience fewer invented facts when provided with clear reference context.

Why GPT-6 Luna Benchmarks Overstate Everyday Performance

Standard laboratory benchmarks systematically overstate practical model performance due to rigid evaluation designs, sanitized data inputs, and prompt tuning. The published GPT-6 Luna benchmarks reflect controlled test suites rather than unpredictable enterprise working conditions.

OpenAI noted that its factuality test data derived from de-identified user conversations containing known historical errors, meaning the test set is intentionally hostile rather than representative of standard office queries. Furthermore, the test runs on AutomationBench rely on cleanly structured API interfaces with unambiguous tool definitions, whereas real-world business databases frequently present null values, ambiguous column headers, and conflicting permissions.

Another operational variable is reasoning effort. Achieving the published factual accuracy and workflow execution numbers requires running GPT-6 Luna at higher effort levels, which increases processing latency and total output tokens. In production workflows requiring sub-second response times, running at lower effort configurations will yield lower accuracy than the headline benchmark suggests.

Who This Model Serves and Deployment Boundaries

GPT-6 Luna is built for high-throughput, structured data processing where unit cost dominates total system viability. It provides an efficient engine for classification, sentiment analysis, entity extraction, data cleansing, and initial customer support routing.

This model is not built for open-ended legal contract drafting, high-liability regulatory compliance analysis, or deep technical architecture decisions. Organizations attempting to handle multi-jurisdictional compliance or bespoke financial audits should bypass lightweight models entirely and deploy frontier models such as GPT-6 Astra or Claude Opus 5 from Anthropic, where deeper contextual reasoning minimizes structural oversights.

Our answer on model selection would change if OpenAI introduced native self-verification mechanisms that achieved maximum benchmark factuality scores at minimum reasoning effort without token inflation. Until such capabilities appear in production, operators must treat GPT-6 Luna as an economical execution tool rather than an autonomous decision-maker.

Practical Next Steps for Validating GPT-6 Luna Benchmarks

Teams should not accept vendor evaluation metrics as a substitute for internal testing on proprietary datasets. Validating whether the advertised gains hold true requires establishing an isolated testing sandbox using historical production logs.

Engineers should curate a representative sample of 200 real transactions containing noisy edge cases, misformatted records, and ambiguous requests. Running these records through GPT-6 Luna at various effort configurations allows organizations to measure real task success rates alongside actual token consumption and latency numbers.

Compare the measured accuracy and latency directly against your existing production baselines to determine whether migrating your pipeline to GPT-6 Luna benchmarks justifies the development and integration effort.