OpenAI GPT-5.6 Sol: Benchmarks, Specs, Pricing, Safety Evals, and What This Model Really Means
Last updated: June 26, 2026. Information current as of: June 26, 2026. GPT-5.6 Sol is in limited preview, and model availability, pricing, benchmark reporting, safety policies, and product access can change quickly.
OpenAI’s GPT-5.6 Sol preview is not just another “new model is smarter” announcement. The interesting part is the whole package: a three-model GPT-5.6 family, more expensive flagship pricing, a new reasoning tier, agentic ultra mode, stronger coding, biology, and cyber evaluations, a system card that puts the model family in High capability territory for cyber and bio/chemical risk, and a limited rollout that OpenAI says started with trusted partners after previewing plans and capabilities to the U.S. government.
That last detail matters. If GPT-5.6 Sol were merely a better chatbot, release strategy would be product marketing. Instead, release strategy is part of the product. The model is being introduced as a frontier agentic system that can write code, reason over scientific workflows, probe security problems, and operate through tool-heavy workflows. That is exciting. It is also exactly why OpenAI is wrapping the launch in heavier safety infrastructure, access control, monitoring, and government coordination.
This Kingy AI breakdown looks at what OpenAI actually announced, what is proven, what is still first-party and unproven, and what GPT-5.6 Sol means for developers, startups, open-source model teams, and the broader AI market.
What OpenAI Announced
OpenAI announced GPT-5.6 as a family of three models: Sol, Terra, and Luna. Sol is the flagship. Terra is positioned as the balanced lower-cost option. Luna is positioned as the fast and affordable option. OpenAI says Terra is competitive with GPT-5.5 while being 2x cheaper, while Luna offers strong capability at OpenAI’s lowest GPT-5.6 cost.
The preview is intentionally constrained. OpenAI says the models are available through API and Codex to select trusted partners during preview, with broader availability for ChatGPT, Codex, and API coming soon. The company also says GPT-5.6 Sol is planned for Cerebras at up to 750 tokens per second in July, which is a notable signal: OpenAI is not only selling model quality; it is also thinking about low-latency agentic execution.
OpenAI’s public framing is carefully calibrated. The announcement highlights state-of-the-art results on Terminal-Bench 2.1, stronger GeneBench v1 results using fewer tokens than GPT-5.5, and its most capable cybersecurity model yet. But it also points readers to the GPT-5.6 preview system card, where the safety story gets more serious.
The GPT-5.6 Model Family: Sol, Terra, and Luna
The family structure is strategic. Instead of forcing every workload onto a single model, OpenAI is separating capability, cost, and latency into three tiers.
| Model | Positioning | What OpenAI is signaling |
|---|---|---|
| GPT-5.6 Sol | Flagship model | Maximum capability for hard coding, science, cyber, and agentic workflows. |
| GPT-5.6 Terra | Balanced lower-cost model | GPT-5.5-class capability at lower cost, according to OpenAI. |
| GPT-5.6 Luna | Fast and affordable model | High-volume and latency-sensitive use cases where cost matters more than peak reasoning. |
That structure should look familiar to teams building with frontier APIs. The question is no longer “which model is best?” The better question is “which model should route to which task?” Sol is for the expensive hard stuff. Terra is for workloads that need strong reasoning without flagship cost. Luna is for throughput, product UX, and broad automation where the marginal price of every output token matters.
GPT-5.6 Sol Specifications and Access Details
OpenAI has not published every architectural detail, and this article will not pretend otherwise. The public specs that matter for builders are access, pricing, reasoning controls, caching, and deployment posture.
- Access: OpenAI says GPT-5.6 models are in limited preview through API and Codex for select trusted partners, with broader ChatGPT, Codex, and API access coming soon.
- Model family: GPT-5.6 includes Sol, Terra, and Luna.
- Reasoning: Sol introduces a new max reasoning effort, and GPT-5.6 adds ultra mode for complex work.
- Tooling posture: The announcement emphasizes agentic coding, benchmarked terminal work, and multi-step tasks.
- Caching: GPT-5.6 adds explicit cache breakpoints and a 30-minute minimum cache life.
- Inference expansion: OpenAI says Sol is planned for Cerebras in July at up to 750 tokens per second.
The big missing piece is independent, third-party benchmark replication. OpenAI’s claims are meaningful because they come from the company operating the model, but they are still first-party claims. Treat them as launch evidence, not settled science.
Release Limitations and Access Caveats
The limited preview is not a boring logistics note. It changes how builders should interpret the launch. During preview, GPT-5.6 is not a fully open public default that every developer can immediately plug into a production system. OpenAI says access is through API and Codex for select trusted partners, with broader access coming later. That means early feedback, benchmarks, and product examples will come from a narrower set of users than a general release.
That has three practical consequences. First, public reports may overrepresent sophisticated teams with better eval harnesses, stronger prompts, and closer support from OpenAI. Second, the model’s behavior in high-volume consumer use is not yet as visible as it would be after broad ChatGPT availability. Third, serious AI companies should avoid building a go-to-market promise around Sol until they know their access tier, rate limits, policy boundaries, data controls, and cost profile.
There is also a safety and compliance angle. A model that is High capability in cyber and biological/chemical risk will not be treated like a generic autocomplete engine. If your product touches security testing, vulnerability discovery, biological analysis, chemistry, or autonomous coding in sensitive environments, your evaluation plan needs to include policy review, logging, abuse monitoring, escalation paths, and fallback behavior. That is how you keep a capable agent from becoming an unmanaged liability.
Pricing Breakdown: Sol vs Terra vs Luna
The launch post lists the GPT-5.6 API pricing clearly:
| Model | Input price | Output price | Use case fit |
|---|---|---|---|
| Sol | $5 per 1M input tokens | $30 per 1M output tokens | Highest-stakes reasoning, coding agents, deep research, security analysis. |
| Terra | $2.50 per 1M input tokens | $15 per 1M output tokens | Balanced production workloads and GPT-5.5-style performance at lower cost. |
| Luna | $1 per 1M input tokens | $6 per 1M output tokens | Fast, affordable default automation and broad product surfaces. |
The caching details are unusually important. OpenAI says GPT-5.6 supports explicit cache breakpoints, has a 30-minute minimum cache life, bills cache writes at 1.25x the uncached input rate, and gives cache reads a 90% cached-input discount. For agentic systems, that can matter as much as the base rate. Long-running coding agents and research agents often revisit the same repo context, instructions, tools, and policy scaffolding. If developers design prompts and context packs around caching, the effective cost can change dramatically.
New Reasoning Modes: Max Reasoning and Ultra Mode
Sol introduces a new max reasoning effort, which is OpenAI’s signal that the flagship is meant for deeper, longer, more compute-intensive problem solving. This is not the mode you blindly use for every support ticket. It is the mode you reach for when the task is hard enough that extra thinking is worth the extra latency and cost.
GPT-5.6 also introduces ultra mode, which OpenAI says uses subagents for complex work. That is a meaningful product-direction signal. Frontier models are moving from one-shot assistants toward orchestrated systems: planner agents, specialist agents, coding agents, browser agents, evaluator agents, and safety layers that check the work. “The model” increasingly means the model plus the agent runtime around it.
For developers, the practical lesson is simple: GPT-5.6 Sol is not just a completion endpoint. It is built for workflows where the model can plan, call tools, inspect outputs, delegate subtasks, and revise. That is powerful, but it also makes evaluation harder. A single prompt result is not enough. You need task-level pass rates, cost per successful run, latency, retry behavior, and failure-mode analysis.
How Builders Should Evaluate GPT-5.6 Sol
The worst way to test Sol is to paste a clever prompt into a chat box, watch one impressive answer, and declare migration complete. Frontier agent models should be evaluated like systems. That means a private task suite, a baseline model, a pass/fail grader, cost tracking, latency tracking, and logs for tool calls, retries, and refusals.
For coding teams, the task suite should include real repository issues: failing tests, migration work, dependency updates, security fixes, documentation changes, and ambiguous bug reports. For security teams, it should separate defensive workflows from prohibited or risky offensive workflows, then measure whether the model stays inside the intended lane. For science teams, especially in bio and chemistry, evaluation should include domain-expert review and strict boundaries around actionable wet-lab or harmful operational detail.
The useful comparison is not “Sol versus everything on a public leaderboard.” The useful comparison is Sol versus your current stack on your private work. If Sol passes more tasks but spends 4x the output tokens, maybe Terra wins. If Luna handles 80% of work at a fraction of the cost, maybe Sol should sit behind a router. If Sol solves the hardest 5% of tasks that no other model can touch, the price may be cheap. The answer depends on completed work, not model aura.
Coding Benchmarks and Terminal-Bench 2.1
OpenAI says GPT-5.6 Sol sets a new state of the art on Terminal-Bench 2.1. Terminal-style benchmarks matter because they test closer to real developer work than generic code snippets: navigating files, running commands, interpreting failures, changing code, and completing tasks inside an environment.
The claim is impressive, but the right reading is cautious. If OpenAI does not publish exact machine-readable scores or if some results live in charts rather than structured tables, do not invent numbers. The safer takeaway is that OpenAI is explicitly competing on long-horizon coding-agent performance, not just chat code generation. That puts GPT-5.6 Sol directly in the arena of Codex, Claude Code-style workflows, open-weight coding agents, and repo-scale automation tools.
Biology Benchmarks and GeneBench v1
OpenAI says GPT-5.6 Sol improves on GeneBench v1 while using fewer tokens than GPT-5.5. That is an important combination: stronger performance and better efficiency. In science workflows, fewer tokens can mean lower cost, lower latency, and less wasted reasoning. But biology is also one of the domains where capability gains trigger more serious safety review.
The GPT-5.6 system card classifies the model family as High capability in Biological/Chemical risk. That does not mean the models are being publicly released as unconstrained biology tools. It means the evaluated capability level is high enough that safety controls, monitoring, and use restrictions matter. For readers outside AI safety, the distinction is important: the risk classification is not an accusation that the model is malicious. It is a deployment classification about what a sufficiently capable user might do with it.
Cybersecurity Benchmarks: ExploitBench and ExploitGym
Cyber is where GPT-5.6 Sol becomes especially interesting. OpenAI says Sol is its most capable cybersecurity model yet. The announcement says that on ExploitBench, Sol is competitive with Anthropic Mythos Preview while using roughly one third of the output tokens. It also says Sol, Terra, and Luna improve on ExploitGym as reasoning effort increases.
ExploitGym is relevant because it tests exploit-reasoning style capability rather than harmless textbook security trivia. That is exactly the class of evaluation that matters for frontier-model deployment. Stronger cyber reasoning can help defenders: vulnerability triage, patch validation, secure code review, incident response, and red-team simulation. The same capability can also help attackers if access, monitoring, and policy controls are weak.
OpenAI’s system-card line is carefully drawn: Sol and Terra can find vulnerabilities and exploit primitives, but in Chromium and Firefox testing they did not autonomously produce full-chain exploits. OpenAI also says Sol does not cross the Cyber Critical threshold. That is the key safety boundary in the public framing. High capability is serious. Critical capability would be a different release conversation.
Preparedness Framework: High vs Critical Risk
OpenAI’s Preparedness Framework is the policy structure behind these classifications. The important idea is that models are evaluated against risk categories and thresholds before deployment. High capability does not automatically mean no release. Critical capability is the line where deployment becomes much more constrained.
For GPT-5.6, the system card says the models are High capability in Cybersecurity and Biological/Chemical risk, but not High in AI Self-Improvement. That combination is revealing. OpenAI is saying the models have advanced domain capability in cyber and bio/chemical work, while not meeting the public threshold for dangerous self-improvement capability.
That should shape how readers interpret the launch. The story is not “OpenAI says everything is safe, carry on.” The story is “OpenAI says this model family is highly capable in sensitive domains, has not crossed specific Critical thresholds, and is being released through a controlled preview with additional mitigations.”
Why the Cyber Results Matter
Cyber capability is not a niche benchmark flex. It sits at the center of frontier AI deployment because it combines agency, tool use, code understanding, adversarial reasoning, and real-world consequences. A model that can reason through vulnerabilities can be a defender’s multiplier. It can also compress the skill gap for offensive work.
That is why the ExploitBench and ExploitGym framing matters even without public exact scores in every row. OpenAI is telling developers that Sol has stronger cyber-relevant reasoning. It is also telling policymakers and enterprise buyers that this capability is being paired with monitoring, refusals, account review, and enforcement. The model is more useful because it is more capable. The release is more constrained for the same reason.
Safety Stack: Refusals, Classifiers, Monitoring, and Enforcement
The announcement and system card describe a layered safety posture rather than a single filter. The stack includes model refusals, real-time classifiers, account-level review, monitoring, and enforcement. That matters because frontier misuse does not show up as one obvious bad prompt. It can show up as a sequence of apparently technical steps spread across sessions, tools, accounts, and intermediate artifacts.
The stronger the model, the less credible it is to rely only on static refusal behavior. A serious safety system needs to look at intent, context, workflows, users, and repeated behavior. OpenAI’s safety posture reflects that. The model layer is one control. The deployment layer is another. Account review and enforcement are also part of the system.
Automated Red-Teaming and the 700,000 A100-Equivalent GPU Hours Claim
OpenAI says it used automated red-teaming at substantial scale, including a claim of more than 700,000 A100-equivalent GPU hours. Read that as a signal about evaluation industrialization. Frontier labs are no longer only testing models with manual prompt lists and small expert review panels. They are using models and compute to search for failures, stress capabilities, and probe dangerous edges.
The caveat: more red-teaming compute does not prove completeness. It proves effort and scale. Automated red-teaming can find many problems, but it still depends on the quality of the search process, scenario design, measurement, and follow-up mitigation. The right conclusion is neither “the model is unsafe” nor “the model is solved.” The right conclusion is that evaluation itself has become a frontier workload.
Government Involvement and Limited Preview
The government angle is arguably as important as the model. OpenAI says it previewed plans and capabilities to the U.S. government and, at the government’s request, started with a limited preview for trusted partners. That is a sharp signal about where frontier AI is going.
For ordinary software, a government-preview note would feel unusual. For a frontier model with High cyber and bio/chemical capability classifications, it is becoming less surprising. Model releases are starting to look more like controlled deployment events, where access policy, national-security concerns, safety thresholds, and product launch timelines are intertwined.
What GPT-5.6 Sol Means for Developers
For developers, GPT-5.6 Sol looks like a premium model for hard agentic workloads. The obvious use cases are coding agents, repo migration, debugging, deep research, security review, biological research assistance under appropriate controls, and complex multi-step planning.
The cost profile means you should not treat Sol as the default for every task. Use routing. Send easy work to Luna or smaller models. Send medium work to Terra. Reserve Sol for tasks where success rate, reliability, and deep reasoning are worth the output-token price. The real metric is not dollars per million tokens. It is successful completed tasks per dollar.
Caching also becomes a design primitive. If your agent repeatedly uses the same repo map, coding standards, tool instructions, and system context, explicit cache breakpoints and a 30-minute cache minimum can materially change economics. The teams that win with GPT-5.6 will not just prompt harder. They will engineer context, caching, evals, and model routing.
What It Means for AI Companies and Startups
For AI startups, GPT-5.6 Sol is both opportunity and pressure. The opportunity is obvious: better long-horizon agents, stronger coding, more capable security workflows, and potentially faster high-end inference through Cerebras. The pressure is that the frontier is becoming more controlled, more expensive at the top end, and more policy-mediated.
If your product depends on the absolute best model, your product may also depend on access rules you do not control. Limited previews, trusted-partner gates, policy enforcement, account review, and future government coordination are not side details. They are operational dependencies. The smartest startups will build with frontier models while maintaining model-routing flexibility, fallback models, and workload-specific evals.
What It Means for Open Source Models
Open-source and open-weight models are not dead because Sol is stronger. If anything, controlled frontier releases make open models more strategically important. Open-weight models give teams portability, fallback options, private deployment paths, and more control over model behavior. Closed frontier models give teams peak capability, polished products, and managed infrastructure.
The gap that matters is task-specific. If Sol dramatically improves agentic coding or cyber workflows, open-weight teams will need stronger long-horizon evals, better tool-use harnesses, and lower total cost. But open models can still win when buyers need sovereignty, inspectability, fine-tuning, local deployment, or lower risk of access disruption.
Kingy AI Take
GPT-5.6 Sol is not just another chatbot upgrade. It is a frontier agentic model with stronger long-horizon coding, biology, and cyber capability, paired with heavier safety infrastructure and a more controlled rollout.
The story is not only “new model is smarter.” The story is that frontier models are becoming powerful enough that release strategy, government coordination, and access control are now part of the product. Sol may be a major leap for coding, bio, and cyber workflows, but the limited preview is not a footnote. It is the signal.
What Feels Proven
- GPT-5.6 is a three-model family with Sol, Terra, and Luna.
- Sol is the flagship, Terra is the balanced lower-cost model, and Luna is the fast affordable model.
- OpenAI has published clear API pricing for all three models.
- OpenAI is emphasizing agentic coding, Terminal-Bench 2.1, GeneBench v1, ExploitBench, and ExploitGym.
- The GPT-5.6 system card classifies the family as High in Cybersecurity and Biological/Chemical risk, but not High in AI Self-Improvement.
- The rollout is intentionally controlled through a trusted-partner preview.
What Feels Unproven
- Independent benchmark replication is still needed.
- Real-world developer economics will depend on tool-loop length, cache design, retries, and success rates.
- Ultra mode’s practical value will depend on how reliably subagents coordinate, verify work, and avoid compounding mistakes.
- The broader availability timeline is still “coming soon,” not a firm public date.
- Enterprise buyers still need to see policy terms, auditability, data controls, and operational guarantees.
Final Verdict
GPT-5.6 Sol looks like one of OpenAI’s most consequential model previews because it combines three trends at once: stronger agentic capability, more explicit sensitive-domain risk management, and a release process that treats access control as a first-class feature.
For builders, the playbook is not to blindly switch everything to Sol. Test it where the work is genuinely hard. Measure completed tasks, not vibes. Use Terra and Luna for cost-sensitive routing. Design around caching. Watch safety-policy boundaries carefully. And assume that frontier AI deployment will keep becoming more like critical infrastructure: more capable, more useful, more expensive, and more governed.
Related Kingy AI Reading
- AI News
- AI Launch Tracker
- AI agents
- AI coding tools
- Open-weight model launches
- OpenAI tool profile
- Codex tool profile
FAQ
What is GPT-5.6 Sol?
GPT-5.6 Sol is OpenAI’s flagship model in the GPT-5.6 family. OpenAI positions it as the highest-capability option for difficult coding, science, cyber, and agentic workflows.
What are GPT-5.6 Terra and Luna?
Terra is the balanced lower-cost GPT-5.6 model, while Luna is the faster and more affordable model. OpenAI says Terra is competitive with GPT-5.5 while being 2x cheaper, and Luna offers strong capability at its lowest GPT-5.6 cost.
How much does GPT-5.6 Sol cost?
OpenAI lists Sol at $5 per 1M input tokens and $30 per 1M output tokens. Terra is $2.50 input and $15 output per 1M tokens. Luna is $1 input and $6 output per 1M tokens.
Is GPT-5.6 Sol available in ChatGPT?
During preview, OpenAI says GPT-5.6 models are available through API and Codex to select trusted partners. Broader availability for ChatGPT, Codex, and API is coming soon, according to OpenAI.
Did GPT-5.6 Sol cross OpenAI’s Cyber Critical threshold?
No. OpenAI says Sol is its most capable cybersecurity model yet and is High capability in cyber, but the system card says it does not cross the Cyber Critical threshold.
Why does government involvement matter?
OpenAI says it previewed plans and capabilities to the U.S. government and began with a limited trusted-partner preview at the government’s request. That shows frontier model release strategy is increasingly tied to safety, policy, and access control.
