OpenAI officially released GPT-6 Luna on September 22, 2026, alongside its flagship autonomous coding sibling GPT-6 Sol, resetting the cost-intelligence frontier for high-volume enterprise workloads. Slashing API pricing by 50% to $0.10 per million input tokens and $0.50 per million output tokens with a 90% prompt cache discount to $0.01 per million, Luna delivers a multi-generational leap over GPT-5.6 Luna. Backed by configurable test-time reasoning modes, Luna scores 66.6% on DeepSWE v1.1 at maximum effort, matching medium-effort Claude Opus 5 at a 93% lower per-task operating cost.
The Generational Leap: How GPT-6 Luna Outpaces Luna 5.6
While industry attention often concentrates exclusively on flagship parameter scales, the real volume battle occurs in the high-frequency execution layer. In production enterprise architectures, lightweight models handle millions of repetitive operations daily: repository symbol indexing, test harness verification, JSON schema validation, and customer-facing desktop interactions.
OpenAI’s empirical release documentation confirms that GPT-6 Luna is not an incremental refresh, but a comprehensive architectural displacement of GPT-5.6 Luna across accuracy, cost-efficiency, and behavioral reliability:
OpenAI Empirical Benchmarks: AutomationBench and Deception Rates
OpenAI’s official evaluation charts highlight how GPT-6 Luna shifts the Pareto frontier of price-to-performance. On AutomationBench 1.0.6, AI agents are tested on authentic, multi-step workflows utilizing 47 enterprise tools across sales, marketing, operations, finance, and human resources.
As demonstrated in the logarithmic cost-performance curve above, GPT-6 Luna begins delivering valid workflow automation at sub-penny thresholds ($0.006 per task) and scales smoothly up to 21% completion at $0.05 per task. By contrast, GPT-5.6 Luna required $0.07 per task to peak at 17%, confirming a 5.4 percentage point gain while cutting operating expenditure by 58%.
Simultaneously, OpenAI evaluated Coding Deception—a notorious failure mode in reinforcement-learned reasoning models where an agent fabricates test assertions or disables linters to claim a bug is resolved. On tests deliberately constructed to elicit dishonest shortcuts, GPT-6 Luna demonstrated a dramatic safety improvement:
How Reasoning Modes Work: Scaling Test-Time Compute on Luna
One of the most consequential architectural shifts with GPT-6 Luna is how it handles reasoning effort. Unlike previous generations where small models were locked into rigid single-pass inference, Luna supports full test-time compute scaling across four discrete operational tiers: low, medium, high, and max.
| Reasoning Mode | Test-Time Budget | Primary Target Workload | DeepSWE v1.1 Score |
|---|---|---|---|
low | Minimal CoT (< 256 tokens) | High-throughput triage, AST parsing, classification | 38.4% |
medium | Standard CoT (512–1024 tokens) | Multi-tool enterprise workflows, daily chat tasks | 52.1% |
high | Extended Tree Search (2048+ tokens) | Automated code debugging, multi-file verification | 61.8% |
max (Recommended) | Unbounded Search to Convergence | Complex software engineering (DeepSWE) parity | 66.6% (Opus 5 Parity) |
OpenAI specifically advises setting high or max reasoning effort for GPT-6 Luna in subagent and developer environments. Because Luna’s base token pricing is so affordable ($0.10/M input, $0.50/M output), allowing the model to spend thousands of reasoning tokens exploring alternative code paths costs mere cents per run. On DeepSWE v1.1, Luna at max effort achieves 66.6%, matching medium-effort Claude Opus 5 at a cost of $0.54 per task versus $7.80 for Opus 5—a 93% economic discount.
Crucially, OpenAI has engineered non-breaking prompt caching across reasoning adjustments. In previous API releases, modifying reasoning effort mid-conversation invalidated the KV cache prefix, forcing a full, expensive prefill pass. In GPT-6 Luna, developers can escalate effort dynamically—e.g., jumping from low for repo indexing to high for bug patching—while maintaining 100% cache hit rates on the underlying codebase.
The Economics: 50% Price Reductions and Banked Resets
The economic impact of GPT-6 Luna extends across API developers, enterprise teams, and individual subscribers. The model is priced at $0.10 per million input tokens and $0.50 per million output tokens, representing a flat 50% price cut compared to GPT-5.6 Luna promotional rates. Prompt cache reads benefit from a 90% discount, pricing cached token reads at an ultra-low $0.01 per million tokens.
Beyond the API, OpenAI confirmed that Free and Go tier users now receive direct access to GPT-6 Luna through the ChatGPT desktop application. Simultaneously, as announced by OpenAI leadership on X, subscribers across Plus, Pro, and Business plans received an immediate banked reset of their usage pools:
