OpenAI Launches GPT-6 Sol and Luna, Halving API Prices to Undercut Rivals

OpenAI introduced two new artificial intelligence models on Tuesday, GPT-6 Sol and GPT-6 Luna, reducing application programming interface (API) costs by 50 percent compared to the promotional pricing of its previous generation.

The two models sit below GPT-6 Astra—the company’s most capable model launched earlier this month—and are designed as direct replacements for the GPT-5.6 lineup. OpenAI stated that Sol and Luna were trained using methods similar to Astra, with the goal of bringing the flagship model’s reasoning, coding, and factual reliability to faster, more affordable deployment tiers.

The release marks a deliberate shift in strategy. Rather than competing solely on raw intelligence, OpenAI is heavily targeting the cost-efficiency frontier, aiming to capture enterprise workloads by aggressively undercutting competitors like Anthropic.

New Economics and Pricing Structures

The primary appeal of the Sol and Luna models is their economics, driven by underlying improvements to caching and inference operations.

GPT-6 Sol, engineered for complex reasoning, multi-step task automation, and software engineering, now costs $2 per million input tokens and $10 per million output tokens. This halves the $4 and $20 rates previously charged for GPT-5.6 Sol.

GPT-6 Luna, built as a lightweight model for high-volume consumer applications, summarization, and data extraction, drops to $0.10 per million input tokens and $0.50 per million output tokens. While OpenAI marketed this as a 50 percent reduction across the board, the output cost for Luna actually represents a steeper 58.3 percent cut from its previous $1.20 rate.

Both models share a context window of 1,050,000 tokens and a maximum output ceiling of 128,000 tokens. However, OpenAI has introduced tiered prompt pricing: prompts exceeding 272,000 tokens are billed at double the standard input rate and 1.5 times the output rate. Developers running asynchronous batch jobs will continue to receive a 50 percent discount on standard rates.

Benchmarking Against Anthropic

OpenAI positioned Sol explicitly as a lower-cost alternative to Anthropic’s Claude 5 family, backing the claim with a series of agentic and coding benchmarks.

On AutomationBench—a test evaluating real-world business workflows across 47 different applications—OpenAI reported that GPT-6 Sol at its highest reasoning effort setting outperformed Claude Opus 5. According to the company, Sol achieved this result while costing just 9 percent of what Anthropic’s model costs per task.

In Agents’ Last Exam, an evaluation of complex professional workflows across 55 sub-industries, GPT-6 Sol scored 56.4 percent at maximum effort. This places it above Claude Opus 5’s top score, while operating at a 60 percent lower cost per task.

Computer use capabilities also show tight competition. On the offline set of OSWorld 2.0, Sol at an “extra-high” effort setting scored 60.5 percent, edging out Claude Opus 5 operating at medium effort (60.3 percent). OpenAI estimates this translates to an 80 percent reduction in cost per task for comparable performance.

Factuality Improvements and Regressions

A persistent weakness of large language models has been factuality. OpenAI claims Sol now makes roughly half as many factual errors as its predecessor, based on internal evaluations utilizing real conversations where users had previously flagged mistakes.

Independent testing by AI research firm Artificial Analysis corroborated reductions in hallucination rates. On the firm’s AA-Omniscience benchmark, Sol dropped its hallucination rate from 92 percent to 60 percent, largely by declining to answer questions it did not have information for. Luna reduced its hallucination rate from 93 percent to 77 percent. Despite the improvements, analysts noted that Luna still struggles with factual accuracy without explicit source grounding.

The cost reductions do come with trade-offs. While Sol improved on coding evaluations like Terminal-Bench 4.0 and SWE-Atlas-QnA, Artificial Analysis noted regressions in specific areas for both models. In GDPval-AA v2.1, a benchmark based on economically valuable occupational tasks, both Sol and Luna dropped in performance when running at maximum effort.

Researchers attributed this regression to the models generating shorter deliverables that occasionally omitted required elements from grading rubrics. GPT-6 Luna also regressed slightly on the independent Coding Agent Index, dropping two points compared to its GPT-5.6 predecessor.

For users requiring uncompromising performance without dropped instructions or shortened deliverables, OpenAI continues to recommend GPT-6 Astra.

Availability and Rollout

Both models are deployable immediately via the OpenAI API under the names gpt-6-sol and gpt-6-luna. Because they are API-exclusive models, OpenAI is not releasing the underlying weights for local hosting.

For end consumers, the models are rolling out progressively today. GPT-6 Sol and Luna are now available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers. Free and Go tier users will gain access to Luna exclusively through the desktop application. Neither model is currently active in the standard ChatGPT web interface for free users.

The simultaneous launch of Sol and Luna tightens the margins in the enterprise AI market. By adopting a variable-compute approach—where developers can scale a single model from “low” to “max” reasoning effort—OpenAI is attempting to consolidate varied workloads into a single API framework. As competitors respond, the industry focus is clearly shifting from raw intelligence benchmarks to the practical economics of software integration.