OpenAI Halves Prices for GPT-6 Sol and Luna

OpenAI launches GPT-6 Sol and Luna with prices cut in half compared to the 5.6 models, along with enhanced caching capabilities. However, independent experts estimate that the advancements in intelligence are modest, with some gains in specific areas and regressions elsewhere. Access begins in ChatGPT Work and Codex, while the API exposes gpt-6-sol and gpt-6-luna.

Scores Close to GPT-5.6 and Internal Discrepancies Between Sol and Luna

Artificial Analysis estimates that the intelligence scores of GPT-6 Sol and Luna remain largely at the level of GPT-5.6, with improvements in some tests and regressions in others. On its Intelligence Index, GPT-6 Sol increases from 47 to 48 points, while Luna remains at 37. On the coding agents index, Sol improves by 2 points while Luna declines by 2 points. The organization reports regressions on two knowledge work benchmarks: on GDPval-AA v2.1, which covers 44 professional domains, Sol loses about 100 Elo points and Luna about 75, with these underperformances primarily attributed to lower presentation quality and incomplete results. On DeepSWE, Luna at maximum effort matches Sol's performance at "xhigh" while costing 78% less, and pushing Sol to maximum effort only places it 2.2 points above Luna.

Prices Cut in Half and 90% Discount on Cached Prompts

The prices for GPT-6 Sol and Luna are reduced by 50% compared to the 5.6 generation. GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens; GPT-6 Luna costs $0.10 for input and $0.50 for output. OpenAI attributes this reduction to improvements in caching and inference and claims to directly pass on these savings, with prices in the same range as some cheaper open-weight models. On the operational side, OpenAI announces a 90% discount on cached input tokens, enhanced prompt caching, a new dashboard and diagnostic tool, and the ability to adjust reasoning effort and tool availability without invalidating the cache. The Terra model, previously the cheapest in the lineup, is discontinued.

In-House Benchmarks: Lower Costs Compared to Claude, Nuanced Results

OpenAI positions its models on value for money against Anthropic's Claude range. On OSWorld 2.0, the company claims that GPT-6 delivers results similar to Claude Opus 5 at about 80% lower cost, while noting that Astra remains ahead. On AutomationBench, which utilizes 47 tools, OpenAI states that GPT-6 Sol at maximum effort outperforms Claude Opus 5 at the same effort level and estimates Sol's cost per task at 9% of Opus's. OpenAI also indicates that Luna improves by 5.4 percentage points compared to its predecessor while costing 58% less. For coding, FrontierCode 1.1 credits GPT-6 Sol with 49.3% at maximum effort for $2.14 per task, compared to Claude Fable 5.1 at 50.3% for $12.83 and Claude Opus 5 at 53.4% for $4.31 at average effort. On DeepSWE v1.1, GPT-6 Sol achieves 68.8% at maximum effort, just 1.1 points shy of Fable 5's best score of 69.9% ("xhigh"), while Fable at "max" is at 69.7% for $21.63, Opus 5 at 73.7% for $11.84, and GPT-5.6 Sol at 72.7% for $6.46. OpenAI presents Sol's goal as approaching the best scores at a lower cost, citing 66.6% at "xhigh" for $1.00 per task, a score that Luna matches at maximum effort for $0.22, with the company claiming comparability to Opus 5 and Fable 5 at average effort, with costs 93% and 96% lower respectively.

Displayed Use Cases and Gradual Availability in Products

According to OpenAI, Sol is suitable for complex recurring tasks, such as feature development, debugging, or data analysis. Luna, on the other hand, is indicated for performing well-defined and low-cost operations at scale, such as document synthesis, information extraction, or producing brief responses. At launch, both models are accessible via ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers; Luna is also available for free users and Go through the desktop application. Neither model is initially available in the regular chat. On the API side, the references are gpt-6-sol and gpt-6-luna, and deployment in ChatGPT is gradual.

Third-Party Measures Previously Contested, Caution on Conclusions

According to Artificial Analysis, the cost per task has been halved for GPT-6 Sol and Luna compared to the previous generation, but this source had previously faced criticism for assigning a too-low score to Astra due to outdated benchmarks. Two updates to the benchmarks were subsequently published, after which Astra regained the top position. These past events suggest caution in interpreting the gap between cost reductions and limited score advancements.