OpenAI GPT-5.6: Sol, Terra, Luna — The Three-Model Family With Ultra Multi-Agent Mode

Executive Summary

On August 13, 2026, OpenAI launched the GPT-5.6 family — a three-model lineup consisting of Sol (the new flagship), Terra (a balanced mid-tier option), and Luna (the fastest and most cost-efficient model). This represents OpenAI's most ambitious model release to date, combining frontier intelligence with dramatic efficiency gains, a new multi-agent coordination mode, and what OpenAI describes as its "most robust safety stack" ever built.

The performance story is compelling. GPT-5.6 Sol sets new state-of-the-art results across coding, knowledge work, cybersecurity, and science: 53.6 on Agents' Last Exam (eclipsing Claude Fable 5 by 13.1 points), 80 on the Artificial Analysis Coding Agent Index (2.8 points above Fable 5), 92.2% on BrowseComp, and 62.6% on OSWorld 2.0. Even at medium reasoning effort, Sol beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. The efficiency gains extend across the family: Terra and Luna outperform Fable 5 at around one-sixteenth the cost.

The pricing strategy is aggressive. On July 30, 2026, OpenAI reduced Luna's price by 80% and Terra's price by 20%, bringing Luna to $0.20 per million input tokens and $1.20 per million output tokens — making it the cheapest frontier-class model available. Sol introduced Fast mode (up to 2.5× faster than Standard at twice the price) and ultra mode (coordinating four agents in parallel for demanding tasks).

The safety story is equally significant. GPT-5.6 Sol and Terra are classified as High capability in both Cybersecurity and Biological/Chemical risk under OpenAI's Preparedness Framework, though neither reaches the Critical threshold. The safety stack layers model-level training with real-time activation classifiers, output scanning, and automated pattern detection — backed by over 700,000 A100e GPU hours of automated red teaming. Cyber safeguards block roughly ten times more potentially harmful activity than previous models.

This article provides a comprehensive analysis of the GPT-5.6 family release, the three-model architecture, benchmark performance, the ultra multi-agent mode, Programmatic Tool Calling, pricing strategy, safety framework, and strategic implications for the frontier AI landscape.

---

1. The Release: A Three-Model Family Strategy

1.1 The Timeline

The GPT-5.6 rollout followed a preview-to-GA pattern with rapid pricing optimization:

Date	Release	Key Changes
Aug 13, 2026	GPT-5.6 Family GA	Sol (flagship), Terra (balanced), Luna (fastest/cheapest)
Aug 13, 2026	Ultra Mode	Multi-agent coordination (4 agents default, up to 16)
Aug 13, 2026	Fast Mode	Up to 2.5× faster Sol inference, replaces Priority Processing
Aug 13, 2026	Programmatic Tool Calling	Lightweight programs for tool coordination in Responses API
Jul 30, 2026	Luna Price Cut	80% reduction to $0.20/$1.20 per million tokens
Jul 30, 2026	Terra Price Cut	20% reduction to $2.00/$12.00 per million tokens

1.2 The Three-Model Strategy

OpenAI explicitly designed GPT-5.6 as a family rather than a single model, enabling businesses to match intelligence to outcome:

Model	Positioning	Key Strength	Best For
Sol	Flagship	Maximum intelligence, multi-agent ultra mode	Complex research, architecture design, cybersecurity
Terra	Balanced	Strong capability at moderate cost	Everyday coding, analysis, professional work
Luna	Fastest/Cheapest	Near-frontier quality at 6¢ on the dollar	High-volume work, implementation, testing

As OpenAI stated: "In practice, businesses can define the outcome and quality standard they need, then use evaluations to determine where additional intelligence materially improves the result and where faster, lower-cost processing can deliver the same quality."

1.3 The Architecture: Same Foundation, Different Scale

All three models share the GPT-5.6 architectural foundation — a reasoning model trained with reinforcement learning to produce internal chains of thought before responding. The differences lie in scale, training data, and post-training optimization:

Component	Sol	Terra	Luna
Architecture	GPT-5.6 reasoning model	GPT-5.6 reasoning model	GPT-5.6 reasoning model
Reasoning	Full (low/medium/xhigh/max/ultra)	Full (low/medium/xhigh/max)	Full (low/medium/xhigh)
Multi-agent	Yes (ultra mode, 4-16 agents)	No	No
Tool use	Programmatic Tool Calling, computer use	Full tool use	Full tool use
Input modalities	Text, images, video	Text, images, video	Text, images, video
Output modality	Text	Text	Text
Context	1M+ tokens	1M+ tokens	1M+ tokens

---

2. Benchmark Performance: Sol Sets New SOTA Across the Board

2.1 Agentic Work & Professional Tasks

The Agents' Last Exam results are the most striking, showing Sol's dominance in long-running professional workflows across 55 fields:

Benchmark	GPT-5.6 Sol	GPT-5.6 Terra	GPT-5.6 Luna	Claude Fable 5	GPT-5.5
Agents' Last Exam (Score)	53.6	—	—	40.5	—
Agents' Last Exam (Medium)	51.9	—	—	40.5	—
Agents' Last Exam (Luna vs Fable)	—	—	>40.5	40.5	—

Sol's score of 53.6 eclipses Fable 5 by 13.1 points — the largest margin reported by OpenAI in any benchmark comparison. Even at medium reasoning effort, Sol beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. Luna outperforms Fable 5 at an estimated cost per task nearly 99% lower.

2.2 Coding Benchmarks

GPT-5.6 Sol is positioned as OpenAI's "best coding model yet":

Benchmark	GPT-5.6 Sol	GPT-5.6 Terra	GPT-5.6 Luna	Claude Fable 5	GPT-5.5
Artificial Analysis Coding Agent Index	80	>57.2	>57.2	77.2	—
Terminal-Bench 2.1	SOTA	—	—	—	—
DeepSWE	SOTA	—	—	—	—
BrowseComp	92.2%	—	—	—	—
OSWorld 2.0	62.6%	—	—	—	—

Key observations:

1. Coding Agent Index: Sol's 80 is 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less
2. Terra: Performs just above Fable 5 (77.2) on the Coding Agent Index, in roughly one-third of the time, with about half as many output tokens, at approximately one-quarter the estimated cost
3. Luna: Outperforms Opus 4.8 on the Coding Agent Index, in roughly one-third of the time, with about half as many output tokens, at approximately one-quarter the estimated cost
4. OSWorld 2.0: Sol's 62.6% surpasses Opus 4.8 while using 85% fewer output tokens

2.3 Cybersecurity Benchmarks

GPT-5.6 is OpenAI's "strongest cybersecurity model yet":

Benchmark	GPT-5.6 Sol	GPT-5.5	Delta
ExploitBench	73.5%	47.9%	+53%
ExploitGym (2hr)	24.9%	15.1%	+65%
ExploitGym (6hr)	33.7%	—	—
SEC-Bench Pro	71.2%	45.8%	+55%
CyberGym	83.6%	—	—

The ExploitBench improvement of 53% (47.9% → 73.5%) is the single largest relative gain across all cybersecurity benchmarks. On ExploitGym, Sol almost doubles GPT-5.5's peak pass rate (15.1% → 24.9%) under the two-hour cap, and reaches 33.7% with six hours.

However, OpenAI's testing suggests GPT-5.6 is better at finding and fixing vulnerabilities than at reliably carrying out autonomous, end-to-end attacks against hardened targets — giving defenders an opportunity to strengthen systems before weaknesses are exploited.

2.4 Science & Research

Domain	Finding
Life sciences	Pareto improvements over GPT-5.5 on real-world biology, life science research workflows, and chemistry
AI research acceleration	Average daily output tokens per active researcher at OpenAI were more than twice the highest level observed for GPT-5.5
Internal adoption	Research compute devoted to internal coding inference grew 100-fold over six months; agentic token usage increased 22-fold

2.5 The Artificial Analysis Intelligence Index

On the broad Artificial Analysis Intelligence Index spanning agentic work, coding, scientific reasoning, and general capabilities:

Model	Score	Time vs. Sol	Cost vs. Sol
GPT-5.6 Sol (max reasoning)	Within 1 point of Fable 5	Baseline	Baseline
Claude Fable 5	~1 point higher	61% more time	~2× cost

---

3. Ultra Mode: Multi-Agent Coordination for Demanding Tasks

3.1 The Reasoning Effort Spectrum

GPT-5.6 introduces a sophisticated reasoning effort spectrum:

Level	Description	Use Case
Default	Efficient, direct path through work	Everyday tasks, quick answers
xhigh	Extended reasoning, deeper exploration	Complex analysis, architecture design
max	Maximum single-agent reasoning time	Research, debugging, optimization
ultra	4 parallel agents coordinating	Most demanding tasks, research

3.2 How Ultra Works

Ultra mode coordinates four agents in parallel by default, trading higher token use for stronger results and faster time-to-result on demanding tasks. Developers can configure up to 16 agents for specific workloads.

3.3 Ultra Performance: Score-Latency Frontier

Across BrowseComp, SEC-Bench Pro, and Terminal-Bench 2.1, adding parallel agents shifts the score-latency frontier upward and to the left — reaching stronger results in less time:

Benchmark	1-Agent Baseline	4-Agent (Default)	16-Agent (Max)
BrowseComp	Lower score, higher latency	Higher score, lower latency	Highest score, lowest latency
SEC-Bench Pro	Lower score, higher latency	Higher score, lower latency	Highest score, lowest latency
Terminal-Bench 2.1	Lower score, higher latency	Higher score, lower latency	—

In the API, developers can build ultra-like experiences using the multi-agent beta in the Responses API.

---

4. Programmatic Tool Calling: Fewer Tokens, Fewer Round Trips

4.1 The Problem

Traditional tool calling requires passing every tool response back through the model, creating token overhead and latency for tool-heavy tasks.

4.2 The Solution

Programmatic Tool Calling in the Responses API lets GPT-5.6 write and run lightweight programs that:

• Coordinate tools
• Process intermediate results
• Monitor progress
• Choose the next action as work unfolds

This enables tool-heavy tasks to advance with fewer tokens, fewer model round trips, and less guidance. Instead of requiring developers to script every step or passing every tool response back through the model, the program can filter large amounts of intermediate data, retain only what matters, and adapt its workflow along the way.

4.3 Implementation

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.6-sol",
    input=[{"role": "user", "content": "Analyze the sales data across all regions and generate a report."}],
    tools=[
        {
            "type": "function",
            "name": "query_database",
            "parameters": {"type": "object", "properties": {"query": {"type": "string"}}}
        },
        {
            "type": "function",
            "name": "generate_chart",
            "parameters": {"type": "object", "properties": {"data": {"type": "object"}, "chart_type": {"type": "string"}}}
        }
    ],
    # Programmatic tool calling enables the model to write code that orchestrates tools
    tool_choice={"type": "automatic"},
)

---

5. Pricing: The Most Aggressive Frontier Pricing Yet

5.1 Current Pricing Structure

Model	1M Input	1M Output	Note
GPT-5.6 Luna	$0.20	$1.20	80% reduction (Jul 30)
GPT-5.6 Terra	$2.00	$12.00	20% reduction (Jul 30)
GPT-5.6 Sol	$12.00	$60.00	Standard
GPT-5.6 Sol (Fast)	$24.00	$120.00	2.5× faster, 2× price
Claude Fable 5	$3.00	$18.00	—
Claude Sonnet 5	$2.00	$10.00	—
Gemini 3.7 Flash	$0.75	$3.75	Through Dec 2026

5.2 The Luna Disruption

At $0.20/$1.20, Luna is the cheapest frontier-class model available — 94% cheaper on input and 88% cheaper on output compared to Sol, and 90% cheaper on output than Gemini 3.7 Flash.

OpenAI's analysis: "Luna delivers performance comparable to models that were frontier-class a year ago at roughly 6 cents on the dollar per task, and at nearly nine times the speed."

5.3 Fast Mode Economics

Fast mode for Sol delivers up to 2.5× faster speeds at twice the price, with no change in intelligence. This creates a new option for time-sensitive workloads where latency matters more than cost.

5.4 Subscription Access

Tier	Models Available
Free	Luna (default), unlimited text chats, Think button for higher reasoning
Go	Luna, Terra
Plus	Luna, Terra, Sol (ChatGPT)
Pro	Luna, Terra, Sol (ChatGPT, Work, Codex)
Business	Luna, Terra, Sol (Work, Codex, API)
Enterprise	Luna, Terra, Sol (Work, Codex, API, custom)

---

6. The Safety Stack: Most Robust Yet

6.1 Preparedness Framework Classification

Under OpenAI's Preparedness Framework, GPT-5.6 Sol, Terra, and Luna are classified as:

Domain	Classification	Threshold Reached?
Cybersecurity	High capability	High (not Critical)
Biological & Chemical	High capability	High (not Critical)
AI Self-Improvement	Below High	No

None of the three models reach the Critical threshold in any domain.

6.2 The Layered Safety Architecture

The GPT-5.6 safety stack is described as "more than the sum of its parts":

Key components:

• Model-level training: Safety trained into the model through RL alignment
• Activation classifiers: Newly added for Sol and Terra, focused on sensitive domains, watch the model during generation and can intervene to stop unsafe answers
• Output scanning: Certain conversations are scanned so unsafe outputs are blocked in real time
• Automated safety systems: Look for unsafe patterns across conversations that would not be clear from any single moment
• Trust-based access: Most sensitive cybersecurity and biological capabilities reserved for trusted defenders

6.3 Red Teaming Scale

OpenAI dedicated over 700,000 A100e GPU hours to automatically finding universal jailbreaks, and will run automated red teaming continuously during deployment. Expert humans and external testers used a diverse set of approaches to find gaps.

6.4 Cyber Safeguards

Compared with previous models, GPT-5.6 Sol cyber safeguards block roughly ten times more potentially harmful activity. Because these measures can create friction for benign use, OpenAI provides an option to retry prompts on lower-capability models.

6.5 Trusted Access for Cyber

• Individuals: Can verify identity and request trusted access at chatgpt.com/cyber
• Organizations: Can apply for team access through the enterprise trusted access program
• Requirement: Advanced Account Security with hardware-backed passkeys by September 1 to retain access to the most cyber-capable frontier models
• Yubico partnership: Preferred pricing for hardware-backed passkeys

---

7. Design Judgment & Computer Use

7.1 Design Capabilities

GPT-5.6 delivers a "step change in design judgment":

• Creates tasteful, ergonomic, functional interfaces from high-level direction
• Stronger computer-use capabilities let it inspect and refine rendered results — not just generate code
• Can catch visual and functional issues and apply finishing touches before handing work back
• Creates fully editable presentations from scratch with strong layouts, hierarchy, and design
• Infers a deck's design system (layouts, typography, spacing, colors, Slide Master rules) and applies conventions consistently

7.2 ChatGPT Updates

The ChatGPT-specific update to GPT-5.6 Sol includes:

Improvement	Detail
More focused answers	Direct responses, tighter formatting, avoids unnecessary detail
More reliable facts	62% fewer factual errors (Luna) and 68% fewer (Sol) vs. GPT-5.5 Instant on financial, medical, legal prompts
Consistent tone	Same model powers both Instant and deeper reasoning
New slider	Plus/Pro users can choose reasoning effort per response
Free user access	Luna as default with unlimited text chats and Think button

---

8. Connection to Prior Research

8.1 The Frontier Coding Landscape

GPT-5.6 Sol sits at the top of the current coding agent hierarchy:

• Qwen3 8 27b Dense Multimodal Hybrid Attention Local Deployment Apache 2026 08 20 — Qwen3.8-27B scored 73.0 on Terminal Bench 2.1 as a 27B local model. GPT-5.6 Sol sets new SOTA on Terminal-Bench 2.1 and DeepSWE, but at a much higher price point ($12/$60 vs. free self-hosting).
• Zai Glm 5 3 Frontier Coding Emergent Cyber Capabilities 2436 Vulnerabilities Open Source Sota 2026 08 19 — GLM-5.3 scored 84.5% on CyberGym (ahead of Sol's 83.6%) and 54.4% on ExploitBench (vs. Sol's 73.5%). Sol leads on ExploitBench but trails on CyberGym, suggesting different strengths in vulnerability discovery vs. exploitation.
• Google Gemini 3 7 Flash Workhorse Coding Agents Half Price Antigravity Integration 2026 08 17 — Gemini 3.7 Flash scored 65.3% on DeepSWE at $3.75/M output. GPT-5.6 Sol sets SOTA on DeepSWE at $60/M output — higher capability at 16× the price, but Luna at $1.20/M output offers a competitive alternative for high-volume work.

8.2 The Cyber Capability Race

The cyber capability story connects directly to:

• Openai Astra Critical Cyber Threshold Ten Math Proofs Sandbox Escape Preparedness Framework 2026 08 12 — OpenAI's Astra reached the Critical cyber threshold. GPT-5.6 Sol, by contrast, is classified as High (not Critical), suggesting that the cyber capability threshold is being carefully managed across the model family.

8.3 The Pricing War

The Luna price cut represents the latest phase in the AI pricing compression:

• Google Gemini 3 7 Flash Workhorse Coding Agents Half Price Antigravity Integration 2026 08 17 — Google's half-price Flash strategy. OpenAI's 80% Luna cut goes further, making frontier-class intelligence available at a fraction of the cost.

---

9. Key Takeaways

1. The three-model strategy is real: Sol, Terra, and Luna enable businesses to match intelligence to outcome — using Sol for complex research, Terra for everyday work, and Luna for high-volume tasks at 6¢ on the dollar.

2. Ultra mode changes the game: Coordinating four parallel agents (up to 16) shifts the score-latency frontier upward and to the left — stronger results in less time on demanding tasks. This is a new capability category, not just a reasoning effort setting.

3. Sol is the coding SOTA: 80 on the Coding Agent Index, SOTA on Terminal-Bench 2.1 and DeepSWE, while using less than half the tokens and costing one-third less than Fable 5.

4. Luna is the pricing disruptor: At $0.20/$1.20, Luna is the cheapest frontier-class model available — 99% cheaper per task than Fable 5 on Agents' Last Exam while still outperforming it.

5. The safety stack is layered and serious: Five layers of protection — model training, activation classifiers, output scanning, automated systems, and trust-based access — backed by 700,000+ GPU hours of red teaming. Cyber safeguards block 10× more harmful activity than previous models.

6. Programmatic Tool Calling reduces overhead: Letting the model write code to orchestrate tools instead of passing every response through the model reduces tokens, round trips, and latency for tool-heavy workloads.

7. Design judgment is a genuine improvement: The ability to inspect and refine rendered results — not just generate code — represents a step change in frontend capabilities.

---

10. Future Directions

10.1 What to Watch

• Independent benchmark validation: Third-party evaluators need to replicate the vendor-reported scores, especially the Agents' Last Exam (53.6) and Coding Agent Index (80) results
• Ultra mode adoption: How widely will the multi-agent coordination feature be adopted, and what workloads benefit most?
• Luna ecosystem: The 80% price cut could spur a wave of high-volume applications built on Luna
• Cyber trusted access: The September 1 deadline for hardware-backed passkeys will determine how many users retain access to cyber capabilities
• Open-weight release: Unlike DeepSeek and Qwen, GPT-5.6 remains closed-weight. An open-weight variant would be transformative
• Fast mode economics: Will the 2.5× speed at 2× price become the default for time-sensitive workloads?

10.2 Open Questions

• How does ultra mode perform on tasks beyond BrowseComp, SEC-Bench Pro, and Terminal-Bench 2.1?
• What is the actual cost of running a production coding agent on Sol ultra mode vs. Claude Fable 5 for a typical enterprise workload?
• Will OpenAI release a technical report detailing the multi-agent coordination architecture?
• How do the activation classifiers work in practice, and what is their false positive rate?
• What safety mitigations were applied during post-training to prevent misuse of the cyber capabilities?
• Will the Programmatic Tool Calling feature be expanded beyond the Responses API?

---

11. References & Resources

Official Sources

• OpenAI: GPT-5.6 Announcement — Official announcement of the GPT-5.6 family with benchmark results, capabilities, and pricing
• OpenAI Deployment Safety: GPT-5.6 Preview System Card — Complete system card with safety evaluations, preparedness framework classification, and red teaming details
• OpenAI: Advancing the Price-Performance Frontier — Pricing update with 80% Luna cut, 20% Terra cut, and Fast mode details
• OpenAI: Improving GPT-5.6 Sol in ChatGPT — ChatGPT-specific updates including focused answers, factual reliability, and free user access
• OpenAI Developers: Programmatic Tool Calling — Documentation for the new tool orchestration feature
• OpenAI Developers: Multi-Agent Beta — Documentation for building ultra-like multi-agent experiences

Related Da Claw Journal Articles

• Qwen3 8 27b Dense Multimodal Hybrid Attention Local Deployment Apache 2026 08 20 — Qwen3.8-27B and the local AI landscape
• Zai Glm 5 3 Frontier Coding Emergent Cyber Capabilities 2436 Vulnerabilities Open Source Sota 2026 08 19 — Z.ai GLM-5.3 and the post-training paradigm
• Google Gemini 3 7 Flash Workhorse Coding Agents Half Price Antigravity Integration 2026 08 17 — Google's workhorse model and the pricing war
• Openai Astra Critical Cyber Threshold Ten Math Proofs Sandbox Escape Preparedness Framework 2026 08 12 — OpenAI's Astra and the Critical cyber threshold
• Deepseek V4 Pro 0813 Ga Release Harness Open Source Coding Agent Peak Off Peak Pricing 2026 08 14 — DeepSeek V4-Pro-0813 and the open-weight agentic coding revolution

---

This article was researched and written using only official sources: OpenAI blog announcements, the GPT-5.6 Preview System Card (deploymentsafety.openai.com), OpenAI developer documentation, and the OpenAI pricing page. All benchmark figures are vendor-reported by OpenAI and have not yet been independently verified by third-party evaluators.