Version 2

OpenAI launched GPT-5.6 Sol, Terra, and Luna publicly on July 9, 2026, then on July 30 cut Luna's price 80% to $0.20 per million input tokens and Terra's price 20%, while introducing Fast mode for Sol at 2.5x speed [7]. Sol is positioned on efficiency-adjusted benchmark performance, claiming to outperform Claude Fable 5 at one-quarter the cost [3], and OpenAI separately claims Sol autonomously optimized its own inference stack, cutting serving costs 20% [6][7]. The ARC-AGI-3 benchmark dispute — OpenAI argues a harness artifact depressed Sol's initial 7.8% score, critics say OpenAI's preferred settings violate competition rules — has spread from Hacker News to broader tech press [10][12].

Luna at $0.20 per million input tokens is now one-fifth the price of Anthropic's cheapest current model (Claude Haiku 4.5 at $1 per million input tokens) [8], which materially changes the cost calculus for high-volume deployments well before any independent validation of Sol's benchmark claims. The ARC-AGI-3 dispute is a proxy for a broader question: who controls evaluation methodology when model developers disagree with benchmark organizers about what constitutes a valid test.

OpenAI's GPT-5.6 model family — Sol (frontier reasoning), Terra, and Luna — launched publicly on July 9, 2026, after a June 26 preview that disclosed capability details and an unusual release mechanism [1][2]. The preview acknowledged that Sol underwent a government-coordinated phased rollout before public availability, which OpenAI framed as a short-term concession: 'We don't believe this kind of government access process should become the long-term default' [1]. Sol can identify exploitation primitives in Chromium and Firefox but did not autonomously produce full-chain exploits under tested conditions; OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red teaming [1].

The launch announcement centered on efficiency-adjusted benchmark performance against Anthropic's models [3]. OpenAI claims Sol scores 53.6 on Agents' Last Exam, 13.1 points above Claude Fable 5 at roughly one-quarter the cost; 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Fable 5 at less than half the output tokens; and 62.6% on OSWorld 2.0, surpassing Claude Opus 4.8 while using 85% fewer output tokens [3]. GPT-5.6 also includes an 'ultra' mode that coordinates four parallel agents for demanding tasks. Independent practitioner comparisons have begun appearing, with mixed results: some reviewers find GPT-5.6 wins on benchmarks while Fable 5 better understands stated user intent [4][5].

On July 29, OpenAI published a technical post claiming Sol autonomously rewrote production GPU kernels in Triton and Gluon, cutting end-to-end serving costs by 20%, and that improvements to speculative decoding partly designed by Sol increased token-generation efficiency by more than 15% [6]. On July 30, OpenAI extended these claims in a pricing post that announced Luna's price drop 80% to $0.20 per million input tokens, Terra's drop 20%, and Fast mode for Sol — up to 2.5x faster at twice the standard price [7]. At $0.20 per million input tokens, Luna is now one-fifth the price of Anthropic's Claude Haiku 4.5. Developer Simon Willison responded immediately, writing that the drop 'completely changes the landscape with respect to lower priced models,' switching his demo site from Gemini to Luna, and updating his llm CLI tool to default to GPT-5.6 Luna from GPT-4o mini [8][9].

A July 29 post separately addressed GPT-5.6 Sol's initial 7.8% ARC-AGI-3 score [10]. OpenAI reported that the benchmark harness discarded private reasoning between actions and applied rolling truncation; enabling retained reasoning and replacing truncation with compaction tripled the score while cutting output tokens sixfold. OpenAI argues third-party evaluators should use production-matched API settings. Critics and ARC Prize rules reportedly prohibit human-tuned harnesses, making OpenAI's preferred settings potentially non-compliant [11]. The dispute has been covered by The Decoder and multiple other outlets [12][13]. A separate group ('schema') claimed roughly 99% on ARC-AGI-3 public using its own custom harness, a result that may itself conflict with competition rules [14][15][16][17].