OpenAI GPT-5.6 Model Costs and Performance Benchmarks

OpenAI has launched the GPT-5.6 model family, with the flagship "Sol" variant achieving an intelligence score of 59 on the Artificial Analysis Intelligence Index—just one point behind the industry-leading Claude Fable 5—while operating at approximately one-third of the cost [1]. This release introduces the first cache-write pricing structure at OpenAI, signaling a strategic shift in how the company manages the cost-to-serve for high-reasoning tasks [1].

Efficiency and performance benchmarks

The GPT-5.6 lineup, which includes Sol, Terra, and Luna, establishes a new Pareto frontier for intelligence versus cost per task [1]. GPT-5.6 Sol (max) leads the Artificial Analysis Coding Agent Index with a score of 80, outperforming competitors in evaluations such as DeepSWE and Terminal-Bench v2 [1]. Notably, the model achieves this while maintaining a per-task cost roughly 40% lower than Claude Fable 5 and 10% lower than Claude Opus 4.8 in agentic coding environments [1].

While Sol leads in coding and presentation quality—recording the highest "Presentation Elo" of any model in the AA-Briefcase benchmark—it faces trade-offs [1]. The model shows a slight increase in hallucination rates compared to its predecessor, GPT-5.5, despite a marginal improvement in overall accuracy [1]. Furthermore, while Sol is the most efficient in the new family, the Terra and Luna variants offer lower cost-per-task options, with Luna costing approximately 80% less than Sol [1].

New pricing and tokenomics

OpenAI has implemented a cache-write pricing model for the first time, charging a 1.25x premium on input tokens when they are committed to memory [1]. This structure aligns with industry trends seen at Anthropic, where cache writes are billed to reflect the memory resources occupied by tokens regardless of reuse frequency [1].

The pricing tiers for the new models are as follows:

What to watch

• Reasoning Effort Scaling: Monitor how the "max reasoning effort" levels impact real-world project costs, as OpenAI now offers a range of effort levels that shift the Pareto frontier for users [1].
• Cache-Write Adoption: Observe whether the 1.25x cache-write premium influences developer behavior regarding long-context memory usage compared to the 90% discount retained for cache reads [1].
• Competitive Gap: Watch for updates to the Claude Fable 5 rubric scores, as the current 56% to 42% lead in the AA-Briefcase rubric remains the primary differentiator between the two top-tier models [1].

The introduction of GPT-5.6 suggests a pivot toward optimizing for "intelligence at scale," where OpenAI is prioritizing cost-efficiency to capture market share in agentic coding and complex knowledge work. Whether these models can sustain their lead in coding evaluations while managing the observed increase in hallucination rates remains the key technical hurdle for the platform.