GPT-5.6 Fully Available: A Deep Dive into the Sol/Terra/Luna Three-Tier Models and Four-Agent Parallelism

GPT-5.6 Officially Available: The Complete Answer After the Pause

A little over a week ago, OpenAI hit pause on its most powerful model, first submitting it to the U.S. government for review and testing it within a small circle. Today the doors are officially open: the three variants of GPT-5.6 — Sol, Terra, and Luna — are simultaneously available to all users, accessible across ChatGPT, Codex, and the API, with a global rollout completing within 24 hours.

Even more noteworthy is that the company has released the complete evaluation scorecard it had previously held back. It's worth clarifying that all scores in this article come from OpenAI's officially published evaluation results — and you'll only get the full picture if you read them together with the footnotes OpenAI itself has annotated.

A Shift in Naming Conventions: From Generations to Capability Tiers

The naming convention has changed this time too. In the past, numbers indicated which generation a model belonged to; now, the three names Sol, Terra, and Luna represent capability tiers, each able to iterate at its own pace:

• Sol: the flagship, most powerful variant;
• Terra: aimed at high-frequency everyday work, matching the performance of the previous generation 5.5 but at a lower price;
• Luna: focused on speed and cost savings, following a value-oriented approach.

The shift from generational numbering to capability-tier naming reflects a more mature approach to product-line management. This transition has well-established precedents in the semiconductor and consumer electronics industries — Intel's Core i series, Qualcomm's Snapdragon 8/7/6 series, and Apple's M-series chips (M4 Pro/Max/Ultra) all adopt a similar "flagship-mainstream-entry" tiered architecture. The core advantage of this naming system is that it decouples "product iteration" from "generational succession": the flagship tier can respond quickly to technical breakthroughs, while the mainstream tier prioritizes cost-effectiveness, and the two no longer need to be released in sync.

Traditional version-numbering systems (like GPT-4, GPT-4.5) carry an implicit assumption of linear evolution, where each generation must comprehensively surpass the last — placing pressure on both R&D cadence and market expectations. Tier-based naming like Sol/Terra/Luna, by contrast, allows different tiers to evolve independently along their own iteration cycles, and lets users make more precise choices based on actual needs and budgets rather than being forced to chase the latest flagship. For AI models, this shift is particularly important — models of different scales differ significantly in their training, tuning, and deployment cycles, and forcing them under the same generational number would either hold back the fast ones or rush the slow ones. This shift in naming logic also hints at OpenAI's long-term product-line planning: models at different capability tiers will coexist over the long run, rather than being replaced by a single flagship.

The core narrative of this release comes down to one sentence: get more performance for every hundred million dollars. The emphasis is no longer on topping the leaderboard, but on completing the same task with fewer tokens, in less time, and at lower cost.

Interpreting the Evaluation Results: Not Just Scores, but Efficiency Curves

Agent and Coding Capabilities

The history of AI evaluation benchmarks is itself a chronicle of evolving perceptions of capability. The early MMLU (Massive Multitask Language Understanding) test contained multiple-choice questions across 57 subjects including math, history, law, and medicine, requiring models to pick the correct answer from four options, emphasizing breadth of static knowledge. HumanEval, meanwhile, required models to generate Python code that passes unit tests based on function signatures and docstrings, emphasizing the accuracy of single-shot code generation. Both types of benchmark test what a model "knows," but as top models approached perfect scores on these tests, their discriminative power quickly saturated, and they became severely disconnected from real work scenarios — real work almost never involves "one-shot Q&A" and instead requires multi-turn interaction, tool calling, and error recovery.

It was against this backdrop that a third generation of evaluation paradigms emerged. The "Agent Ultimate Exam" covers long-horizon workflows across 55 professional domains and is a comprehensive benchmark designed specifically to evaluate an AI's sustained decision-making ability in real tasks. Unlike traditional single-turn Q&A benchmarks, it requires models to maintain contextual coherence across multi-step tasks, correctly use tool calls, and handle intermediate errors. The Artificial Analysis coding agent index, operated by an independent third party, focuses on the overall quality of code generation, debugging, and multi-turn interaction, and is more aligned with engineering reality than earlier benchmarks like HumanEval. Both benchmarks point to the same trend: the focus of evaluation is shifting from "first-attempt accuracy" to "end-to-end efficiency in completing tasks" — a shift that directly influences the direction of model development, moving training objectives from "getting it right" to "completing real tasks with the fewest resources," and driving major advances in reinforcement learning and tool-use capabilities.

On the aforementioned "Agent Ultimate Exam," Sol scored 53.6, setting a new record — 13.1 points higher than its benchmark rival Fable 5. Even with the reasoning tier dialed down to medium, it still led by 11.4 points, at an estimated cost of only about 1.4 points. More interestingly, the cheaper Terra and Luna also surpassed Fable 5, at a cost of about 1.16 points.

On the Artificial Analysis coding agent index, Sol at its highest reasoning tier scored 80, setting a new record, 2.8 points higher than Fable 5. The real highlight is in the second half of that sentence: it used less than half the output tokens, took less than half the time, and cost about a third less. Within the same family, Terra was slightly higher than Fable 5, while Luna surpassed Opus 4.8.

Knowledge Work and Desktop Operation

BrowseComp is a web-browsing capability benchmark developed in-house by OpenAI. Its core challenge is that it requires models to retrieve and synthesize information through multiple rounds of real web interaction rather than relying on pretraining memory, making it a more effective measure of a model's real-time information-processing ability. Unlike traditional RAG (retrieval-augmented generation) tests where the retrieval corpus is prepared in advance, BrowseComp requires models to autonomously decide search strategies, judge the credibility of retrieval results, and integrate information across pages — much closer to a real information-analysis workflow.

OSWorld 2.0, on the other hand, is a computer-operation benchmark that simulates users executing cross-application tasks in a real operating-system interface (such as organizing files in a file manager and then pasting them into an email), involving screenshot understanding, GUI element localization, and multi-step execution. This benchmark places especially stringent demands on multimodal capabilities: the model must simultaneously understand the visual semantics of interface screenshots, infer the current task progress, and decide the next mouse or keyboard action. Achieving this capability chain depends on multimodal fusion training — the model no longer translates images into text descriptions before reasoning, but instead lets the visual encoder and language decoder share a single representation space, directly performing cross-modal association within a unified feature space, so that the spatial relationships and semantic meaning of interface elements can be understood simultaneously. This is the technical foundation for the "see and act" capability that OSWorld-type tasks require, and the most direct product of the multimodal fusion training of the past couple of years.

On the web-browsing benchmark BrowseComp, Sol scored 92.2%; on the computer-operation simulation OSWorld 2.0, it scored 62.6% — both current state-of-the-art results. On OSWorld in particular, it not only surpassed Opus 4.8 but reduced output tokens by 85%. This figure is especially worth noting — it means the model has learned to more directly "understand the interface and act," rather than compensating for gaps in visual perception with extensive verbal reasoning, reflecting a substantive leap in multimodal understanding and action-planning capabilities.

These numbers were never just about scores, but about the same results with fewer tokens, less time, and lower cost. This efficiency curve is more worthy of ongoing attention than being first on a leaderboard.

Ultra Mode Unveiled: Four Agents Working in Parallel by Default

Many people previously treated Ultra as something mystical, but this time the company has finally spelled it out. The model spends more time reasoning, identifying paths, self-checking, and correcting than before; Ultra goes a step further — it coordinates four agents working in parallel by default, trading more tokens for stronger results and shorter completion times.

Ultra mode coordinating four agents in parallel corresponds academically to a combination of "ensemble inference" and "multi-path search." A single model's chain-of-thought reasoning suffers from path dependence: once an early step goes astray, subsequent reasoning struggles to self-correct — a phenomenon especially evident in math problem-solving, where an incorrect intermediate variable is unquestioningly inherited by later steps. A multi-agent parallel architecture allows multiple solution paths to be explored simultaneously, with the final optimal result chosen through voting, scoring, or the judgment of another model.

This essentially expands "reasoning width" from serial to parallel, trading higher compute for a lower error rate and shorter wall-clock time. Different agents can simultaneously explore different solution paths, or each take on a different role such as planning, execution, verification, and correction — this division of roles corresponds to the "separation of concerns" principle in software engineering, manifesting in AI systems as assigning metacognition (how should I do this) and execution (the specific operations) to different dedicated modules, avoiding frequent switching by a single model between the two cognitive loads. DeepMind's AlphaCode 2 and Meta's multi-agent code-generation research have both validated the effectiveness of this approach: rather than having a single model do longer chain-of-thought reasoning, it's better to use parallelization to distribute the cognitive load and shorten wall-clock time. The cost is a multiplicative increase in token consumption, and the engineering challenges lie in how to design task-decomposition granularity to maximize parallel gains, and how to avoid the problem of "the worst answer contaminating the best answer" during the result-aggregation stage.

Official charts show that after introducing parallel agents, the frontier of score versus latency moves toward the upper left — higher scores and shorter times. Developers who want to build this experience themselves can use the multi-agent capabilities still in beta within the Responses API — they need to design task-decomposition strategies and result-aggregation logic themselves, which is one of the core challenges in deploying AI engineering today. OpenAI making this capability a product default rather than an experimental feature marks the transition of "multi-agent collaboration" from research prototype to engineering standard, moving it from concept to a default product capability.

Cybersecurity: The Most Significant Improvement, and the One Requiring the Most Caution

Cybersecurity is the area with the largest improvement in this release, and also the part that requires the most careful framing.

• ExploitBench Gen 2: 73.5%, up from just 47.9% in the previous generation;
• ExploitGym (two-hour limit): peak pass rate rose from 15.1% to 24.9%, nearly doubling; relaxing to six hours reaches 33.7%;
• CyberSecBench Pro: 71.2%, up from 45.8% in the previous generation.

ExploitBench and ExploitGym are security benchmarks designed specifically for vulnerability-exploitation capabilities. The former tests whether a model can find and exploit real CVE (Common Vulnerabilities and Exposures) vulnerabilities in a controlled environment, while the latter introduces time limits to simulate real penetration-testing scenarios — the two-hour limit corresponds to the typical time window for a single problem in high-intensity CTF (Capture The Flag) competitions. The CVE system is maintained by MITRE Corporation and is the unified naming standard for cybersecurity vulnerabilities worldwide; being able to reproduce a CVE exploitation path in a controlled environment means the model has mastered the complete reasoning chain from vulnerability principle to attack execution — a capability of significant value to both defenders (vulnerability remediation, impact assessment) and attackers (actual exploitation), which is precisely why such evaluations require particular attention to context. Here we must add the footnote OpenAI itself annotated: these cybersecurity evaluations were run under reduced-protection conditions — meaning that under normal product conditions, the model's capabilities would be actively suppressed by the safety layer, and the evaluation scores reflect the model's capability ceiling rather than its everyday usable level.

"Decoupled capability-and-protection testing" reveals a fundamental tension in AI safety research: to accurately assess a model's true capability ceiling, the protection layer must be temporarily removed; but once the results of testing with the protection layer removed are made public, they can themselves become a reference benchmark for attackers, helping them judge which attack paths are worth investing in. This dilemma also exists in high-risk fields like nuclear weapons and biotechnology, where the usual solution is "restricted disclosure" — revealing detailed methodology only to specific security researchers. OpenAI's choice to explicitly annotate the "tested under reduced-protection conditions" footnote in its official report is a compromise transparency practice: disclosing enough information for academic scrutiny while guiding readers through the footnote to correctly understand what the numbers mean. However, footnotes are easily overlooked as information spreads widely, which is precisely the core value of the statement "you only get the full picture if you read them together with the official footnotes." Similarly, the cost and latency figures given on the official site are simulated estimates, not actual bills, and the real differences may be substantial. Furthermore, the biology benchmark GenieBench evaluation did not include comparison models, so it cannot simply be described as "comprehensively leading."

The Price of Greater Capability

This time OpenAI provided a quantitative answer. The potentially harmful activity intercepted by the new model reaches 10 times that of previous models. Before the official release, they invested around 700,000 A100-equivalent GPU hours in black-box automated red-teaming, specifically seeking jailbreak methods that generalize across scenarios. The protection mechanism also no longer relies solely on classifiers assigning labels, but introduces a monitor with reasoning capabilities that reads the full conversation before judging whether risk is present.

Traditional content-safety mechanisms rely on classifiers: mapping input text to predefined harmful categories — fast but with fixed rules, easily bypassed by carefully crafted wording — for example, wrapping true intent in a "hypothetical scenario" or "academic discussion." The newly introduced reasoning monitor is essentially an independent language model that, after reading the full conversation context, makes a semantic-level risk judgment, able to understand implicit intent and cross-sentence connections in multi-turn conversations. This approach is closer to how human reviewers work, with stronger recognition of complex jailbreak attacks (such as role-play wrapping, step-by-step inducement, and nested fictional frameworks).

From an architectural standpoint, the reasoning monitor is equivalent to adding a "metacognitive layer" to the main model's output pipeline: the main model generates content, while the monitor reviews whether the content complies with safety policies. This "generate-review" two-layer architecture echoes the idea of Constitutional AI — the latter, proposed by Anthropic, improves safety by having the model self-critique and revise its outputs according to a set of preset principles. The core difference between the two is that the reasoning monitor is an independent external reviewer, whereas Constitutional AI is an internalized constraint of the model itself. This structural difference gives the reasoning monitor an important advantage: it is inherently immune to "self-deception"-type jailbreak attacks — the kind of manipulation where an attacker induces the main model to believe that harmful output complies with safety principles. The cost is higher compute — each response requires running an additional complete reasoning step — and there's a risk of misjudging sensitive but legitimate content, such as identifying a security researcher's vulnerability-analysis report as attack instructions. It is precisely this misjudgment risk that directly leads to the "collateral damage" problem and the model-switching option described below.

The cost is "collateral damage." Therefore, an option has been retained in ChatGPT and Codex allowing users to switch to a slightly less capable model and re-run.

The most memorable statement from OpenAI is this: over-blocking is itself a security risk. It blocks defenders, preventing them from testing systems and releasing patches; meanwhile, malicious actors can still use other models, including increasingly powerful open-source ones.

Users seeking higher defensive capabilities can apply for a program called Trusted Access, intended for legitimate work such as vulnerability triage, malware analysis, and patch verification. The Trusted Access program's introduction of hardware security keys signals that access control for high-capability AI tools is drawing on financial-grade security architecture. Hardware keys under the FIDO2/WebAuthn standard (such as YubiKey and Google Titan Key) provide "unclonable proof of physical presence": the private key never leaves the hardware device, and even if a user's computer is completely compromised, an attacker cannot impersonate the user without the physical device. This forms a progressive hierarchy with traditional "knowledge factor" authentication (passwords) and "possession factor" authentication (SMS verification codes), achieving a qualitative leap in security.

For AI safety governance, the logic behind this design is to bind capability-access permissions to traceable real identities, enabling attribution and accountability in cases of abuse. This tiered-authorization architecture already has mature practices in finance and healthcare; OpenAI introducing it into AI access control foreshadows that the business model for high-risk AI capabilities may evolve further in the future: from "pay to use" to a multi-threshold system of "payment + real-name verification + trusted devices + usage auditing," analogous to the current compliance requirements for financial derivatives trading or the purchase of hazardous chemicals. But there's one threshold: individual members must enable an advanced account with a hardware key before September 1, or they will revert to default permissions.

Who Can Use It, and at What Price

In ChatGPT, Plus, Pro, Business, and Enterprise users can use Sol and higher-quality versions; free and standard users can use Terra; paying users can freely choose among the three variants and set their own reasoning tier. Max is available to all users who can use GPT-5.6 — just enable it in settings; Ultra is restricted to higher subscription tiers.

API pricing is calculated per million tokens:

• Sol: $5 input / $30 output;
• Terra: $2.5 input / $15 output;
• Luna: $1 input / $6 output (the lowest price).

The caching mechanism is also more predictable this time: it supports explicit cache breakpoints with a minimum retention of 30 minutes; writing to the cache is billed at 1.25 times the uncached input price, while reading from the cache still enjoys a 90% discount.

To understand the value of explicit cache breakpoints, you first need to understand how KV caching works in the Transformer architecture: during each forward pass, the attention layer needs to compute three sets of vectors — Query, Key, and Value — for each position in the sequence. The Key and Value computation results correspond directly to the input content, and for system prompts and historical content that remain unchanged across multiple conversation turns, this portion of KV state can be fully reused. This principle is similar to a CPU's L1/L2 cache — keeping frequently accessed data in fast storage to avoid recomputing from scratch each time — thereby reducing read costs to 10% of the original price.

Explicit cache breakpoints (rather than the system automatically determining cache boundaries) grant developers finer-grained control: they can set different cache boundaries for unchanging business rules versus frequently updated user context, maximizing cache-hit rates for specific business scenarios. For enterprise applications where system prompts exceed several thousand tokens, this mechanism can compress the actual bill by 40% to 70%, making it a cost-optimization measure not to be ignored in high-frequency call scenarios. Billing cache writes at 1.25 times is a reasonable cost allocation — establishing the cache for the first time requires additional storage and management overhead, similar to the cost of building a database index; the 90% discount enjoyed on each subsequent read quickly offsets this initial premium when call volume is large enough.

Understanding the Three Main Threads of This Release

The real message conveyed by this release is not "another number one," but three main threads:

First, all scores come from OpenAI's official evaluations, and you only get the full picture by reading them together with the official footnotes;

Second, what's worth repeatedly scrutinizing is that efficiency curve of performance versus cost — the same task, with fewer tokens, less time, and lower spending;

Third, the stronger the capability, the greater the cost of openness — interception volume up 10x, identity-tiered authorization, and hardware keys exchanged for higher permissions.

Have you started using GPT-5.6 yet? How's it working out? Feel free to share your experience in the comments.

Key Takeaways