GPT-5.6 Trio Launches: Luna, Terra, and Sol Built for Long-Running Agentic Tasks

OpenAI Launches the GPT-5.6 Family: Luna, Terra, and Sol

OpenAI's latest flagship model, GPT-5.6, is now fully available, launching with three distinct versions at once: Luna, Terra, and Sol, corresponding to capability and compute tiers ranging from small to large. Unlike previous release strategies centered on a single flagship, this tiered approach lets developers flexibly choose based on task complexity and cost budget, continuing OpenAI's product philosophy of making "intelligence more accessible and affordable."

All three models share the same base specifications: a knowledge cutoff date of February 16, 2026, support for a million-token (1M) context window, and a maximum output length of up to 128,000 tokens. Whichever tier you choose, developers can handle extra-long documents, large codebases, or complex multi-turn conversations. The differences lie mainly in reasoning capability and cost.

Pricing Strategy: Why "Cost Per Token" No Longer Tells the Whole Story

The three models are priced per million input/output tokens: Luna at $1/$6, Terra at $2.50/$15, and Sol at $5/$30. For comparison, Anthropic's Claude Opus series is $5/$25, while Claude Fable 5 runs as high as $10/$50.

However, as veteran AI observer Simon Willison points out, looking at "price per million tokens" alone is now barely informative. Understanding this judgment requires first grasping the nature of Reasoning Tokens: these are the tokens consumed when modern large language models perform internal "chain-of-thought" reasoning before generating a final answer. Take OpenAI's o-series and Anthropic's Claude extended thinking mode as examples—the model first generates thousands or even tens of thousands of "thinking step" tokens. These tokens are typically invisible to the user but are still counted for billing. This means a model with a seemingly cheap price tag could rack up a bill far higher than expected if a given task triggers extensive reasoning. Different models can consume vastly different numbers of reasoning tokens to complete the same task—a lower-priced model that needs several times more reasoning tokens to reach an answer may actually cost more in the end. The truly meaningful dimension for comparison has shifted from "unit price" to "total cost of completing a specific task."

The Core Selling Point: Performance on Long-Running Agentic Tasks

OpenAI's most significant benchmark claim this time centers on "long-running agentic tasks," reflecting how the industry's focus is shifting from "single question, single answer" toward "autonomously completing multi-step, complex workflows."

To appreciate the significance of this shift, one must clarify the fundamental difference between agentic tasks and the traditional Q&A paradigm. Traditional LLM interaction is "single-turn Q&A": the user enters a question, the model returns an answer, and the entire process completes in milliseconds to seconds. Agentic tasks, by contrast, require the model to autonomously plan, break down goals, iteratively call tools (search, code execution, file read/write, API calls, etc.), and dynamically adjust its strategy based on the results of each step—a process that may last minutes or even hours. In this mode, the model needs capabilities like long-term memory management, error recovery, and resource scheduling, placing far higher demands on context window length and instruction-following stability than single-turn Q&A. This is precisely why GPT-5.6 emphasizes its million-token context and long-running capabilities.

Officials cited an evaluation called Agents' Last Exam—an assessment framework covering 55 specialized domains and simulating long-running professional workflows. The data shows:

GPT-5.6 Sol achieved a new high score of 53.6, beating Claude Fable 5 (using adaptive reasoning) by 13.1 points
Under "medium reasoning effort," Sol still led Fable 5 by 11.4 points at roughly a quarter of the estimated cost
The smaller Terra and Luna could also surpass Fable 5 at roughly 1/16 the cost

OpenAI emphasizes that GPT-5.6 was trained with the goal of "squeezing more useful work output from every token." If these figures hold up under independent verification, the real significance isn't a higher benchmark score, but an overall downward shift in the price-performance curve—enabling agentic tasks that previously required a top-tier model to be handled by cheaper small and mid-size models.

A Notable Exception: The SWE-Bench Pro Coding Benchmark

Not every benchmark favors GPT-5.6. On the widely-watched coding benchmark SWE-Bench Pro, Claude Fable 5 led decisively with a score of 80%, well ahead of GPT-5.6 Sol's 64.6%.

Understanding this result requires knowing how SWE-Bench is designed. This benchmark, proposed by a Princeton University research team in 2023, is built around extracting resolved issues from real GitHub repositories and requiring the model to automatically generate code patches that pass the corresponding unit tests. Its Pro version further filters for more complex, cross-file real engineering problems, and is regarded as one of the gold standards for measuring a model's "real software engineering ability." However, the benchmark's limitation is that test cases depend on the quality of the repository's original test suite—if the original tests are incomplete or ambiguous, the evaluation results become distorted.

Intriguingly, just one day before the GPT-5.6 launch, OpenAI "happened" to publish an article specifically auditing SWE-Bench Pro, pointing out quality issues with the benchmark:

Based on these results, we estimate that roughly 30% of SWE-bench Pro tasks are themselves "broken," and we recommend that model developers scrutinize the relevant results carefully.

The evaluation set may indeed contain noise, and OpenAI's critique has a technical basis. But proactively challenging a benchmark on the eve of releasing a model that lags on it inevitably invites interpretation as a preemptive PR move. For developers, the rational attitude is: don't put blind faith in any single benchmark, especially data self-reported by model vendors, and instead validate against your own actual workloads.

Real-World Experience: Competitive, but Not Necessarily a Clean Sweep

Simon Willison, who had early access to GPT-5.6 Sol, offered a relatively restrained firsthand assessment: Sol is "indeed very capable," but on the complex coding tasks he uses daily, "so far it hasn't struck me as stronger than Fable."

This observation echoes the SWE-Bench Pro scores and reminds us: the "long-task agentic advantage" emphasized in official marketing won't necessarily translate into a perceptible improvement in every specific scenario (especially hardcore coding tasks). Model selection ultimately comes back to the fundamental question of "what exactly are you going to use it for."

New API Capabilities Not to Be Overlooked

For developers, the API capabilities updated alongside GPT-5.6 are often the parts with the most practical value. The official model usage guide hides several intriguing new features:

Programmatic Tool Calling

To understand the value of this feature, it helps to first understand its relationship with the existing MCP protocol. MCP (Model Context Protocol) is a standardized protocol proposed and open-sourced by Anthropic in late 2024, aiming to unify the communication interface between AI models and external tools and data sources—akin to a "USB-C standard for AI." MCP's calling model is primarily "single declarative calls": the model chooses which tool to call, waits for the result, then decides the next step.

GPT-5.6's programmatic tool calling goes a step further: the feature allows the model to "write and run JavaScript to orchestrate tool calls"—the model is no longer limited to "calling one tool at a time," but can use code to write loops, conditional branches, and parallel-call logic, orchestrating multiple tool calls into a complete script. This promises to bridge the gap between MCP and a full terminal session, letting the model combine CLI tool chains to complete tasks like a human engineer—essentially an upgrade from "ordering off a menu" to "cooking your own meal."

Native Multi-agent Support

The model can now "spin up subagents for parallel, focused work." The "primary agent + subagent" pattern that previously required developers to build themselves at the application layer is now built directly into the core API, greatly lowering the barrier to building complex agentic systems.

Prompt Cache Breakpoints

OpenAI has introduced an explicit prompt caching mechanism similar to Claude's, and its technical principles are worth exploring. The core of prompt caching is reusing already-computed KV caches (Key-Value Cache). During LLM inference, every token in the input sequence must pass through the Attention mechanism to generate corresponding key-value vectors that are stored in the cache. If every request contains the same long prefix (such as a system prompt, tool definitions, or a long document), repeatedly computing these parts is both time-consuming and compute-intensive. Prompt caching allows the API to retain these computation results server-side; when a subsequent request hits the cache, the cached portion is typically billed at a discount to 10%–25% of the original price. Anthropic was the first to introduce explicit cache breakpoint markers, giving developers precise control over what content to cache; OpenAI now follows suit, allowing developers to manually specify cache breakpoints rather than relying entirely on the API's automatic detection. For teams that heavily use long system prompts or RAG (Retrieval-Augmented Generation) scenarios, this can bring significant cost savings. Additionally, image requests now include a detail: original option, which forgoes any downscaling of the image before processing, preserving full image information for vision-sensitive tasks.

Conclusion: The Battleground Shifts from "Smarter" to "More Cost-Effective"

The core narrative of the GPT-5.6 family is not an absolute blowout on any single dimension, but efficiency—completing the same or even more agentic work at lower cost. The three-tier product line, built-in multi-agent support, and programmatic tool calling all point in the same direction: making AI that autonomously completes long tasks cheaper and easier to deploy.

Of course, its lag on SWE-Bench Pro, the benchmark critique on the eve of launch, and the early testers' assessment of "no obvious superiority" all remind us to stay level-headed. In an era of increasingly frequent model iteration, the most reliable basis for judgment remains a single one: take your own real tasks and run them yourself.