OpenAI announced GPT-5.6 on July 9, 2026, introducing a three-tier model family — Sol, Terra, and Luna — each designed for a distinct performance-cost profile. Unlike previous OpenAI releases that launched a single flagship model, the GPT-5.6 family reflects the company's strategy of disaggregating its capabilities across specialised models rather than optimising a single general-purpose system.

The Three Tiers

Sol is the flagship performance tier. OpenAI positions Sol as achieving state-of-the-art results on Terminal-Bench 2.1 — the most demanding agentic coding benchmark — surpassing both GPT-5 and Claude Opus 4's published scores on that specific evaluation. Sol is primarily targeting enterprise and developer use cases requiring maximum capability: complex multi-step agentic tasks, advanced reasoning over large codebases, and long-horizon research workflows. It is available through OpenAI's API at a premium tier pricing and through the ChatGPT Pro subscription ($200/month).

Terra is the mid-tier model and is likely to be the most widely deployed of the three. OpenAI states that Terra delivers performance comparable to GPT-5.5 — last year's mid-tier release — at approximately half the API cost. For the vast majority of practical use cases (customer service automation, document analysis, code assistance for standard projects, content generation), Terra represents the better price-performance choice over Sol. Terra is also the model powering ChatGPT's standard Plus subscription tier.

Luna is the speed-optimised tier — the fastest model OpenAI has released and the cheapest to deploy via API. Luna is positioned for high-volume, latency-sensitive applications: real-time voice assistants, inline code completion, interactive product features where response time matters more than maximum capability. Luna's benchmark scores are lower than Terra and Sol on reasoning-intensive tasks, but its speed advantage makes it the appropriate choice for applications where users expect sub-second responses.

Benchmark Context

OpenAI's Terminal-Bench 2.1 top score for Sol should be read with appropriate context. Benchmark performance in agentic coding is highly sensitive to evaluation methodology, and independent replication of company-reported results has historically shown smaller gaps between competing frontier models than company benchmarks suggest. The practical performance difference between GPT-5.6 Sol, Claude Opus 4, and Gemini 2.0 Ultra on real enterprise tasks is likely to be smaller than the benchmark gap implies.

Where GPT-5.6 Sol's edge appears more durable based on independent early testing is in long-context tasks specifically — processing and reasoning over very long documents or codebases — where OpenAI has invested significant architectural work.

The Government Review

Notably, GPT-5.6 went through a pre-deployment review period initiated by the White House — approximately 45 days — before public release. This review, requested under executive authority over national security-adjacent AI capabilities, is the first time a major US AI release has undergone formal pre-deployment government review. OpenAI completed the review and received clearance to proceed with full deployment. The precedent matters: it establishes a de facto process by which the US government can request review periods for future frontier model releases, an informal governance mechanism that may expand over time.

Which Tier Is Right for You

For developers building production applications: Terra is almost certainly the right choice for most workloads, with Luna reserved for latency-sensitive features and Sol reserved for the small subset of tasks that genuinely require maximum reasoning capability and can justify the cost premium.

For ChatGPT users: Plus subscribers automatically get Terra-level capability. Pro subscribers ($200/month) get access to Sol for demanding tasks and Luna for fast conversational use. Free users access a limited version of Luna.

For enterprises evaluating frontier AI: benchmark scores matter less than total cost of ownership and integration reliability. Terra's GPT-5.5-level quality at half the price makes it the practical frontrunner for most enterprise deployment decisions in the second half of 2026.

Pricing and Access Structure

OpenAI has structured GPT-5.6 access across four distinct tiers, each mapping to different use cases and price points.

Free tier users access a rate-limited version of Luna — fast responses, basic capability, with usage caps that reset daily. This tier is designed for casual users and as a product introduction for potential paying subscribers.

ChatGPT Plus ($20/month) provides access to Terra as the primary model, with Luna available for quick interactions. Plus subscribers can perform most professional tasks — writing, analysis, coding, research — without encountering capability gaps that require a higher tier.

ChatGPT Pro ($200/month) unlocks Sol for complex tasks, with unlimited Terra and Luna access. Pro is positioned explicitly for power users: developers, researchers, financial analysts, and legal professionals whose work demands maximum reasoning capability and who generate enough value from AI assistance to justify the premium.

API access is priced per token across all three models, with Sol priced approximately 3x Terra and 8x Luna. OpenAI has also introduced "Flex" API pricing — a lower rate with variable latency guarantees — for batch workloads that are not latency-sensitive, which is relevant for document processing, data analysis pipelines, and overnight automation tasks.

GPT-Live: The Real-Time Differentiator

Alongside GPT-5.6, OpenAI released GPT-Live — a real-time voice and video interaction mode that builds on the Advanced Voice Mode released in 2025 but adds persistent visual context. GPT-Live can maintain awareness of a shared screen or camera feed throughout a conversation, enabling use cases like real-time coding assistance (the model sees the IDE and terminal), live document review (the model follows along as a document is scrolled), and interactive tutoring where the model can see the student's work.

GPT-Live is available in GPT-5.6 Terra and Sol. It represents OpenAI's most direct competitive response to Google's Gemini Live feature set and Apple's upcoming on-device multimodal capabilities in AFM 3. The real-time visual context capability closes a significant gap that had made GPT-5 feel behind compared to Google's Gemini product in consumer demos.

Terminal-Bench 2.1: What the Benchmark Actually Measures

OpenAI's claim that Sol achieves SOTA on Terminal-Bench 2.1 warrants a brief explanation of what that benchmark actually tests. Terminal-Bench 2.1 is a set of 500 software engineering tasks presented in a realistic terminal environment: the model is given a codebase, a bug report or feature request, and must independently navigate the file system, run tests, edit code, and verify that the task is complete — all without human intervention.

The benchmark is specifically designed to measure autonomous coding capability — the ability to operate as an engineering agent, not merely to answer coding questions. It is distinct from HumanEval or SWE-Bench, which are more commonly cited in mainstream AI coverage. Sol's SOTA claim on Terminal-Bench 2.1 is meaningful for enterprise customers evaluating AI coding agents, but it is a narrower claim than "best model for all tasks."

Independent researchers have noted that Sol's advantage on Terminal-Bench correlates with its extended context window (reportedly 512K tokens for Sol, versus 256K for Terra) and improved tool-use reliability — the model makes fewer errors when calling external tools like search engines, code interpreters, and file system commands. These architectural improvements translate directly into better performance on long-horizon agentic tasks, which is the use case Terminal-Bench is designed to evaluate.