GPT-5.6 Sol, Terra, and Luna: OpenAI's Subagent Era, Ultra Mode, and the Most Robust Safety Stack Yet
OpenAI launched the GPT-5.6 family on June 26, 2026 — Sol (flagship), Terra (balanced), and Luna (fast/affordable) — with a new ultra mode leveraging coordinated subagents, max reasoning effort, 700,000 GPU hours of automated red-teaming, and Cerebras integration at 750 tokens/second. Sol Ultra achieves 91.9% on Terminal-Bench 2.1, competitive with Mythos Preview on ExploitBench² using 1/3 the tokens. Currently in limited preview for ~20 government-vetted organizations.
GPT-5.6 Sol, Terra, and Luna: OpenAI's Subagent Era, Ultra Mode, and the Most Robust Safety Stack Yet
Executive Summary
On June 26, 2026, OpenAI launched the GPT-5.6 family — a three-model lineup consisting of Sol (flagship), Terra (balanced mid-tier), and Luna (fast and affordable). This represents OpenAI's most significant model release since GPT-5.5, introducing two architectural innovations that could redefine how frontier models operate: max reasoning effort (giving Sol extended time for deep reasoning) and ultra mode (leveraging coordinated subagents to parallelize complex work beyond a single agent's capabilities).
The launch comes with OpenAI's most robust safety stack to date, including over 700,000 A100-equivalent GPU hours dedicated to automated red-teaming for universal jailbreaks, newly added activation classifiers for Sol and Terra that can intervene during generation, and real-time output scanning for high-risk conversations. Under OpenAI's Preparedness Framework, all three models are classified as High capability in both Cybersecurity and Biological/Chemical risk, though none reach the Critical threshold.
On benchmarks, GPT-5.6 Sol Ultra achieves 91.9% on Terminal-Bench 2.1 — the highest publicly recorded score on that benchmark, surpassing Claude Mythos 5 (84.3%) and GPT-5.5 (88.0%). On ExploitBench², Sol is competitive with Mythos Preview using only ~1/3 of the output tokens. On GeneBench-Pro, the GPT-5.6 family shows a clear scaling curve: Luna at 16.5%, Terra at 23.3%, with Sol achieving even higher results.
The models are currently in limited preview for approximately 20 government-vetted organizations through the API and Codex, following a White House request that OpenAI stage the rollout while national-security cybersecurity reviews are completed. OpenAI plans general availability "in the coming weeks" and has stated it does not believe government-gated access should become the long-term default.
1. The GPT-5.6 Family: Three Models, One Architecture
1.1 Model Specifications
| Model | Positioning | Input (per 1M) | Output (per 1M) | API Identifier | Key Strength |
|---|---|---|---|---|---|
| GPT-5.6 Sol | Flagship | $5.00 | $30.00 | gpt-5.6-sol | Deep reasoning, cyber, science, ultra mode |
| GPT-5.6 Terra | Balanced | $2.50 | $15.00 | gpt-5.6-terra | Competitive with GPT-5.5 at 2× cheaper |
| GPT-5.6 Luna | Fast & affordable | $1.00 | $6.00 | gpt-5.6-luna | Strong capability at lowest cost |
Note: The naming system introduced with GPT-5.6 separates generation (the number, 5.6) from capability tier (Sol/Terra/Luna), creating durable tiers that can advance on their own cadence. This is a structural change from the previous GPT-5.4/GPT-5.5 naming.
1.2 The Reasoning Foundation
All three models are reasoning models trained through reinforcement learning, capable of producing long internal chains of thought before responding. They learn to refine their thinking, try different strategies, and recognize mistakes — which also helps them follow safety policies and resist jailbreak attempts.
2. Ultra Mode: The Subagent Architecture
2.1 What Is Ultra Mode?
Ultra mode is GPT-5.6's most significant architectural innovation. It goes beyond the capabilities of a single agent by leveraging coordinated subagents to split complex long-horizon tasks and execute them in parallel. This is not simply parallel API calls — it is a built-in multi-agent orchestration layer within the model itself.
OpenAI's description:
"A new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work."
2.2 How It Works
Based on the system card and official documentation, ultra mode operates as follows:
- Task decomposition: The primary agent (Sol) analyzes the user request and identifies sub-tasks that can be parallelized.
- Subagent spawning: Coordinated subagents are activated, each handling a specific aspect of the task.
- Parallel execution: Subagents work concurrently on their assigned sub-tasks.
- Result synthesis: The primary agent aggregates subagent outputs into a coherent final response.
2.3 Performance Impact
The impact of ultra mode is dramatic on complex agentic tasks:
| Configuration | Terminal-Bench 2.1 | Improvement |
|---|---|---|
| GPT-5.6 Sol Ultra | 91.9% | +3.1pp over Sol |
| GPT-5.6 Sol | 88.8% | Baseline |
| Claude Mythos 5 | 84.3% | — |
| GPT-5.5 | 88.0% | — |
| Gemini 3.5 Flash | 76.2% | — |
Analysis: The 3.1-point jump from Sol (88.8%) to Sol Ultra (91.9%) demonstrates that the subagent architecture provides a measurable advantage on tasks requiring planning, iteration, and tool coordination. At 91.9%, Sol Ultra is the highest publicly recorded score on Terminal-Bench 2.1.
2.4 Connection to Prior Work
This subagent architecture is a natural evolution of the agentic coding paradigm documented in Qwen3 7 Max Agent Centric Era Long Horizon Execution 2026 07 01 (Qwen3.7-Max's agent-centric capabilities) and Gemini 3 5 Flash Agentic Frontier Multimodal Reasoning 1m Context 2026 07 03 (Gemini 3.5 Flash's MCP Atlas performance). However, GPT-5.6's approach is distinct: rather than requiring external orchestration (as with Qwen-AgentWorld or MCP-based workflows), the subagent coordination is built into the model itself.
3. Cybersecurity Capabilities: Stronger But Below Critical
3.1 The Cyber Assessment
Under OpenAI's Preparedness Framework, GPT-5.6 Sol, Terra, and Luna are all classified as High capability in Cybersecurity. None reach the Critical threshold.
OpenAI's key findings from cybersecurity testing:
- Sol and Terra can find vulnerabilities and pieces of exploits but were unable to carry out autonomous, end-to-end attacks against hardened targets.
- On ExploitBench², GPT-5.6 Sol is competitive with Mythos Preview using only ~1/3 of the output tokens.
- On ExploitGym (created by UC Berkeley researchers in collaboration with OpenAI and other frontier labs), all three models (Sol, Terra, Luna) demonstrate strong improvements in cyber capabilities as reasoning effort increases.
- In evaluations involving Chromium and Firefox, Sol identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit under the conditions tested.
3.2 The Defender Advantage
OpenAI's assessment emphasizes a critical asymmetry:
"GPT-5.6 Sol is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks."
This gives defenders an opportunity to harden systems before vulnerabilities are exploited — an opportunity that may narrow as offensive capabilities improve.
3.3 Comparison With Fable 5
The cyber capabilities of GPT-5.6 Sol are significant in the context of the Fable 5 suspension documented in Claude Fable 5 Mythos 5 Redeployment Export Control Lifted Safeguards Industry Framework 2026 07 02:
| Aspect | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|
| Cyber classification | High (below Critical) | High (below Critical) |
| End-to-end exploits | Unable against hardened targets | Safeguards block exploit generation |
| Token efficiency | ~1/3 tokens vs Mythos Preview | Not directly comparable |
| Government gating | Yes (~20 orgs, preview) | Yes (was suspended 19 days) |
| Safeguard approach | Multi-layered with activation classifiers | Safety classifier + Opus 4.8 fallback |
Both models demonstrate that frontier cyber capabilities are now a standard feature of flagship models, and both have been subject to government coordination before broad release.
4. The Safety Stack: 700,000 GPU Hours of Red-Teaming
4.1 Layered Safeguards
OpenAI describes the GPT-5.6 safety stack as "more than the sum of its parts":
4.2 Key Safety Technologies
-
Model-level training: Models are trained via reinforcement learning to refuse prohibited cyber assistance, including when users attempt to disguise intent or jailbreak.
-
Activation classifiers (new): Sol and Terra are served with newly added activation classifiers focused on sensitive domains. These watch the model's internal activations during generation and can intervene to stop unsafe answers before they reach the user.
-
Real-time output scanning: For higher-risk cases, generation may be paused while a larger reasoning model reviews the conversation context. If the output is assessed as disallowed, it is withheld before reaching the user.
-
Account-level review: Flagged activity triggers review across relevant conversations and risk signals, distinguishing persistent malicious behavior from legitimate dual-use security work.
-
Automated red-teaming: Over 700,000 A100-equivalent GPU hours were dedicated to automatically finding universal jailbreaks — attacks that work across many prompts or contexts, not just one narrow setting.
4.3 Deployment Simulation
OpenAI now uses deployment simulation to forecast safety performance before release:
- Past ChatGPT production conversations with GPT-5.5 are resampled with GPT-5.6 Sol.
- The resampled completions are automatically labeled for disallowed content.
- Results predict that GPT-5.6 Sol will have approximately the same amount of disallowed content violations as GPT-5.5 during deployment.
- The only significant changes: sexual disallowed content increased by 40% (from 0.05% to 0.07%), and disallowed mental health responses reduced by ~40% (from 0.03% to 0.02%).
Analysis: The deployment simulation approach — validated in OpenAI's recent research — represents a maturation of pre-release safety evaluation. Rather than relying solely on static benchmarks, OpenAI now simulates how the model will behave on real production traffic.
5. Benchmark Performance: Across the Family
5.1 Coding: Terminal-Bench 2.1
| Model | Terminal-Bench 2.1 | Notes |
|---|---|---|
| GPT-5.6 Sol Ultra | 91.9% | Highest recorded score |
| GPT-5.6 Sol | 88.8% | +0.8pp over GPT-5.5 |
| GPT-5.5 | 88.0% | Previous flagship |
| Claude Mythos 5 | 84.3% | — |
| Claude Fable 5 | 83.4% | — |
| Gemini 3.5 Flash | 76.2% | — |
5.2 Cybersecurity: ExploitBench²
| Model | ExploitBench² | Token Efficiency |
|---|---|---|
| GPT-5.6 Sol | Competitive with Mythos Preview | ~1/3 output tokens vs Mythos |
| Mythos Preview | Baseline | Baseline |
Analysis: Sol matching Mythos Preview's performance while using only one-third the output tokens represents a significant efficiency gain for cybersecurity workloads.
5.3 Biology: GeneBench-Pro
GeneBench-Pro evaluates multistage statistical reasoning in genomics and quantitative biology:
| Model | GeneBench-Pro Pass Rate |
|---|---|
| GPT-5.6 Sol | >23.3% (highest in family) |
| GPT-5.6 Terra | 23.3% |
| GPT-5.6 Luna | 16.5% |
| GPT-5.5 | 12.0% |
| GPT-5.4 | 8.9% |
| GPT-5.2 | 4.9% |
Analysis: The clear scaling curve across the GPT-5.6 family (16.5% → 23.3% → higher for Sol) and across generations (4.9% → 12.0% → 16.5%+) demonstrates consistent improvement in scientific reasoning under quantitative uncertainty.
5.4 Prompt Injection Robustness
| Model | Connectors | Search & Function-Calling |
|---|---|---|
| GPT-5.6 Sol | 1.000 | 0.910 |
| GPT-5.6 Terra | 1.000 | 0.946 |
| GPT-5.6 Luna | 0.999 | 0.897 |
| GPT-5.5 | 1.000 | — |
| GPT-5.4 Thinking | 0.998 | 0.697 |
Analysis: The GPT-5.6 family shows substantial improvement over GPT-5.4 Thinking on search and function-calling prompt injection (0.697 → 0.897–0.946), a critical capability for agent systems that use tools and web search.
6. Pricing and Economic Analysis
6.1 Cost Comparison
| Model | Input (per 1M) | Output (per 1M) | 100K in + 50K out |
|---|---|---|---|
| GPT-5.6 Luna | $1.00 | $6.00 | $0.40 |
| GPT-5.6 Terra | $2.50 | $15.00 | $1.00 |
| GPT-5.6 Sol | $5.00 | $30.00 | $2.00 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $0.60 |
| Qwen3.7-Max (50% off) | $1.25 | $3.75 | $0.31 |
| Claude Fable 5 | $10.00 | $50.00 | $3.50 |
Analysis: Luna at $1/$6 is OpenAI's cheapest frontier model to date, undercutting Gemini 3.5 Flash ($1.50/$9) and competitive with Qwen3.7-Max's promotional pricing. Terra at $2.50/$15 is positioned as "competitive with GPT-5.5 while being 2× cheaper." Sol at $5/$30 is premium but 67% cheaper than Fable 5 ($10/$50).
6.2 Prompt Caching Improvements
GPT-5.6 introduces more predictable prompt caching:
- Explicit cache breakpoints: Developers can control where caching occurs.
- 30-minute minimum cache life: Cached content persists for at least 30 minutes.
- Cache writes: Billed at 1.25× the model's uncached input rate.
- Cache reads: Continue to receive the 90% cached-input discount.
Analysis: The explicit cache breakpoints and guaranteed minimum cache life address two common complaints about previous caching implementations: unpredictability and short cache lifetimes. For agentic workloads with repeated context (e.g., codebase analysis, document review), this could significantly reduce costs.
7. Cerebras Integration: 750 Tokens Per Second
7.1 The Speed Breakthrough
OpenAI is launching GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second in July 2026. This represents a dramatic increase in throughput for frontier inference.
| Deployment | Estimated Throughput | Use Case |
|---|---|---|
| Cerebras (GPT-5.6 Sol) | 750 tokens/s | Real-time agentic workflows |
| Standard API | ~50-100 tokens/s (estimated) | General purpose |
| Priority processing | Higher than standard | Time-sensitive work |
Analysis: At 750 tokens/second, GPT-5.6 Sol on Cerebras can generate a 10,000-token response in approximately 13 seconds. For agent workloads that fire many sequential calls (each requiring reasoning, tool use, and iteration), this kind of throughput changes what feels interactive versus what feels like a batch job.
7.2 Strategic Implications
The Cerebras partnership signals that OpenAI is diversifying its inference infrastructure beyond its own hardware. This has implications for:
- Capacity scaling: Access to Cerebras' wafer-scale engines provides additional compute for high-demand periods.
- Latency reduction: 750 tokens/second enables near-real-time interaction even for complex reasoning tasks.
- Cost optimization: Specialized hardware may reduce per-token inference costs at scale.
8. Government Gating and the Path to General Availability
8.1 Current Access
As of July 6, 2026, GPT-5.6 is available only to approximately 20 government-vetted organizations through the API and Codex. This follows a White House request that OpenAI stage the rollout while national-security cybersecurity capability reviews are completed.
8.2 OpenAI's Position
OpenAI has been clear about its opposition to permanent government gating:
"We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks."
8.3 The Regulatory Context
The GPT-5.6 preview launch occurs in the shadow of:
- The Fable 5 suspension (Claude Fable 5 Mythos 5 Redeployment Export Control Lifted Safeguards Industry Framework 2026 07 02) — a 19-day disruption that demonstrated government power to halt model deployment.
- The August 1 Executive Order deadline — NSA, Treasury, and CISA have 60 days to build a voluntary pre-release framework for covered frontier models.
- The shared jailbreak framework — co-developed by Anthropic, Amazon, Microsoft, and Google to establish consistent standards for assessing model bypass severity.
OpenAI's approach — voluntary pre-release sharing with the government in exchange for staged rather than blocked release — represents a middle path between the Fable 5 incident (where the government acted unilaterally) and unrestricted release.
9. Comparison With the July 2026 Frontier
9.1 The Complete Landscape
| Lab | Model | Agentic | Coding | Cyber | Multimodal | Context | Price (per 1M) | Status |
|---|---|---|---|---|---|---|---|---|
| OpenAI | GPT-5.6 Sol Ultra | 91.9% TB2.1 | SOTA TB2.1 | High (below Critical) | No | ~200K | $5/$30 | Limited preview |
| OpenAI | GPT-5.6 Terra | — | Competitive w/ GPT-5.5 | High | No | ~200K | $2.50/$15 | Limited preview |
| OpenAI | GPT-5.6 Luna | — | Strong | High | No | ~200K | $1/$6 | Limited preview |
| Anthropic | Claude Sonnet 5 | 80.4% TB2.1 | 63.2% SWE-Bench | Safeguards enabled | No | 1M | $2/$10 (intro) | Available |
| Anthropic | Fable 5 | 83.4% TB2.1 | ~88% (est.) | High (below Critical) | No | 200K | $10/$50 | Available |
| Gemini 3.5 Flash | 76.2% TB2.1 | 55.1% SWE-Bench | High | Yes | 1M | $1.50/$9 | Available | |
| Alibaba | Qwen3.7-Max | ~85/100 | 91.1/100 coding | — | No | 1M | $1.25/$3.75 | Available |
9.2 The Subagent Differentiator
GPT-5.6's ultra mode with coordinated subagents is unique in the current landscape. While Qwen3.7-Max and Gemini 3.5 Flash excel at agentic workflows through tool use and MCP integration, they operate as single agents that call external tools. GPT-5.6 Sol Ultra introduces internal multi-agent orchestration — the model itself spawns and coordinates subagents to parallelize complex work.
This is a fundamental architectural difference that could prove significant for tasks requiring:
- Parallel research: Multiple subagents researching different aspects simultaneously.
- Concurrent code generation: Different subagents writing different modules in parallel.
- Distributed testing: Subagents running tests across different environments concurrently.
10. Key Takeaways
-
GPT-5.6 Sol Ultra sets a new agentic coding record. At 91.9% on Terminal-Bench 2.1, it surpasses all previously published scores, including Claude Mythos 5 (84.3%) and GPT-5.5 (88.0%).
-
The subagent architecture is a genuine innovation. Ultra mode's coordinated subagents represent a shift from single-agent tool use to internal multi-agent orchestration — a capability no other frontier model currently offers.
-
The safety stack is the most comprehensive yet. 700,000 GPU hours of automated red-teaming, activation classifiers, real-time output scanning, and deployment simulation represent a significant investment in pre-release safety.
-
Cyber capabilities are strong but below Critical. Sol can find vulnerabilities and exploit primitives but cannot autonomously execute end-to-end attacks against hardened targets — and does so with 3× better token efficiency than Mythos Preview.
-
The three-model lineup covers the full price spectrum. Luna at $1/$6, Terra at $2.50/$15, and Sol at $5/$30 provide options from cost-efficient to flagship, with Terra positioned as "competitive with GPT-5.5 at 2× cheaper."
-
Government gating is temporary but significant. The ~20 organization preview follows a White House request, and OpenAI plans general availability "in the coming weeks." The Fable 5 incident makes this staging politically necessary.
-
The Cerebras integration changes the latency equation. At 750 tokens/second, GPT-5.6 Sol on Cerebras enables near-real-time interaction for complex reasoning tasks — a game-changer for agentic workflows.
-
The naming system is a structural improvement. Separating generation (5.6) from capability tier (Sol/Terra/Luna) creates durable categories that can advance independently, making it easier for users to understand the model landscape.
11. References & Resources
Official Sources
- OpenAI: Previewing GPT-5.6 Sol — Official announcement with capabilities, safeguards, and pricing
- OpenAI: GPT-5.6 Preview System Card — Full system card with safety evaluations, benchmark results, and threat modeling
- OpenAI Help Center: A Preview of GPT-5.6 Sol, Terra, and Luna — Eligibility, availability, and pricing details
- OpenAI: Introducing GeneBench-Pro — GeneBench-Pro benchmark details and results
- OpenAI: GeneBench-Pro Paper (PDF) — Technical paper on GeneBench-Pro evaluation methodology
- OpenAI: Deployment Simulation Research — Research on pre-release deployment simulation
- OpenAI: Preparedness Framework — Risk classification framework
- OpenAI: Approach to Age Prediction — Age-appropriate content protections
- ExploitGym Paper (arXiv) — UC Berkeley benchmark for cybersecurity evaluation
Related Research in This Journal
- Claude Sonnet 5 Agentic Mid Tier Model 2026 07 03 — Claude Sonnet 5's agentic capabilities and pricing
- Gemini 3 5 Flash Agentic Frontier Multimodal Reasoning 1m Context 2026 07 03 — Gemini 3.5 Flash's agentic performance and multimodal capabilities
- Claude Fable 5 Mythos 5 Redeployment Export Control Lifted Safeguards Industry Framework 2026 07 02 — Fable 5 suspension, export controls, and the regulatory landscape
- Qwen3 7 Max Agent Centric Era Long Horizon Execution 2026 07 01 — Qwen3.7-Max's agent-centric capabilities
- Deepseek V4 Dspark Speculative Decoding Open Source Efficiency Breakthrough 2026 06 30 — DeepSeek V4 and open-source efficiency
12. Future Directions
What to Watch
-
General availability timeline: OpenAI's "coming weeks" estimate could mean late July or early August. The Fable 5 precedent suggests that if no major jailbreaks are found during the preview, broad release will follow quickly.
-
Ultra mode maturity: The subagent architecture is new and will be refined based on production feedback. Expect improvements in task decomposition, subagent coordination, and result synthesis.
-
Cerebras scaling: The 750 tokens/second throughput is initially limited to select customers. Broader availability on Cerebras hardware will depend on capacity expansion.
-
Government framework evolution: The August 1 Executive Order deadline will shape how future model releases are handled. OpenAI's voluntary pre-release sharing with GPT-5.6 may become the template.
-
Terra and Luna benchmarks: The system card focuses on Sol. Detailed benchmarks for Terra and Luna will be published at general availability, which will determine their competitive positioning.
-
ChatGPT integration: GPT-5.6 models will eventually be available in ChatGPT, Codex, and the API. The ChatGPT experience with ultra mode and subagents will be a key differentiator.
-
Enterprise safety controls: OpenAI is working with enterprise customers on "privacy-preserving detection, customer-operated safety controls, and access calibrated to the risk of a customer, user, or workload."
-
Automated red-teaming continuity: The 700,000 GPU hours of pre-release red-teaming will continue during deployment, creating a continuous improvement loop for safety.
The Bigger Picture
GPT-5.6 represents OpenAI's most ambitious release to date in terms of both capability and safety. The subagent architecture in ultra mode could be as significant as the introduction of tool use or reasoning models — it changes the fundamental unit of computation from a single agent to a coordinated team.
The safety investment — 700,000 GPU hours of automated red-teaming, activation classifiers, deployment simulation — sets a new bar for pre-release evaluation. Whether this level of investment becomes the norm for frontier model releases remains to be seen.
The government gating, while temporary, establishes a precedent: the most capable models will be reviewed by national security agencies before broad release. OpenAI's opposition to this becoming permanent reflects a tension between safety and access that will define the industry for years to come.
For organizations building with frontier AI, GPT-5.6 offers a clear upgrade path: Sol for the hardest problems (especially with ultra mode), Terra for everyday work at half the cost of GPT-5.5, and Luna for cost-sensitive workloads. The Cerebras integration adds a speed dimension that could make real-time agentic workflows practical for the first time.
The subagent era has begun, and OpenAI is leading the charge.
Article written July 6, 2026. Sources verified against OpenAI's official blog post, system card, help center documentation, and research papers. All benchmark numbers sourced directly from OpenAI's official evaluations and system card. GeneBench-Pro data from OpenAI's official paper.
🔗 Referenced by
- 📅July 9: GPT-5.6 Public Launch — Sol, Terra, Luna Go Global with Ultra Mode and the Most Robust Cyber Safeguards Yet2026-07-09T00:00:00.000Z
- 🔬GPT-5.6 Public Launch: Sol, Terra, Luna Go Global with Ultra Mode, 750 TPS on Cerebras, and the Most Robust Cyber Safeguards Yet2026-07-09T00:00:00.000Z
- 🔬Meta's Muse Ecosystem: Muse Image Launch, Superintelligence Labs, and the Watermelon Model2026-07-08T00:00:00.000Z
- 🔬Claude Science: Anthropic's AI Workbench for Drug Discovery and Biomedical Research2026-07-07T00:00:00.000Z
- 📅July 6: GPT-5.6 Subagent Era & AI Weekly Roundup2026-07-06T00:00:00.000Z