OpenAI shipped GPT-5.6 on July 9, 2026 — and it's not one model. It's three. Sol, Terra, and Luna sit at different price points, capability levels, and access tiers. The spread between the cheapest and most expensive is 5x on input tokens and 5x on output tokens. For enterprise teams running AI at scale, picking the wrong one is an expensive mistake made silently, at volume, every single day.
This is the model selection guide enterprise leaders needed the moment GPT-5.6 dropped.
Three Tiers, One Generation
OpenAI structured GPT-5.6 the way enterprise software vendors have long structured their product lines: good, better, best. But unlike a SaaS pricing page where "Professional" unlocks features you may never use, each tier here reflects a genuine capability difference with hard cost implications.
Sol is the flagship. At $5 per million input tokens and $30 per million output tokens, it is OpenAI's most powerful model to date. Sol introduces a new "max" reasoning effort and an "ultra" mode that runs four parallel subagents to tackle complex multi-step work.
Terra is the everyday tier. At $2.50 input and $15 output per million tokens — half Sol's price — Terra delivers performance OpenAI describes as competitive with GPT-5.5, the previous flagship. For the vast majority of enterprise workloads, Terra will be the right call.
Luna is the cost-optimized option. At $1 input and $6 output per million tokens, Luna is five times cheaper than Sol. It is built for high-volume, lower-complexity tasks where throughput matters more than peak capability.
The prompt caching economics changed too. GPT-5.6 supports explicit cache breakpoints and a 30-minute minimum cache window. Cache writes cost 1.25x the uncached input rate. Cache reads continue to receive a 90% discount. For enterprises building retrieval-augmented or context-heavy workflows, this new caching architecture materially changes the math.
What the Benchmarks Actually Say
Enterprise AI decisions should not be made on marketing copy. Here is what OpenAI's published evaluation table shows — including where GPT-5.6 does not lead.
Where Sol wins clearly:
- Artificial Analysis Coding Agent Index v1.1: Sol at max reasoning scores 80, topping Claude Fable 5 at 77.2 and delivering that result in less than half the output tokens and less than half the time.
- Agents' Last Exam (55-field professional workflow eval): Sol at 52.7%, well ahead of Fable 5 at 40.5%.
- OSWorld 2.0 (computer use): Sol at 62.6%, ahead of Claude Opus 4.8 at 54.8% — while using 85% fewer output tokens.
- Terminal-Bench 2.1 (command-line workflows): Sol at 88.8%, rising to 91.9% in ultra mode.
Where the gap is narrower than headlines suggest:
- Broad intelligence (Artificial Analysis Intelligence Index v4.1): Fable 5 leads at 59.9 versus Sol's 58.9. The gap is real but not dramatic.
- Tool use (Toolathlon): Sol at 58%, Fable 5 at 61.7%, Opus 4.8 at 59.9%. GPT-5.6 does not lead here.
Where GPT-5.6 has a genuine deficit:
- SWE-Bench Pro (code review and real-world software engineering tasks): Sol at 64.6%, Claude Fable 5 at 80%, Claude Mythos 5 at 80.3%. That is a 15-point gap on a benchmark enterprise engineering teams should care about. If your primary use case is code review at scale, this matters.
The honest takeaway: GPT-5.6 Sol is the strongest available model for agentic, multi-step workflows and complex reasoning. It is not the strongest model for everything. Teams running code review pipelines at volume should weight the SWE-Bench Pro gap seriously.
Enterprise Use Case Mapping
The three-tier structure was built for exactly this kind of workload differentiation. Here is how the tiers map to real enterprise use cases.
Sol (use sparingly and intentionally):
- Complex legal contract analysis where nuance and reasoning depth directly affect business outcomes
- Vulnerability research and security assessments in cybersecurity workflows
- Long-horizon financial modeling, M&A scenario analysis, or strategic planning synthesis
- Executive briefing generation from unstructured data across multiple sources
- Any autonomous agentic workflow where quality of output has significant downstream consequences
Sol is the right tool when a wrong answer costs more than the model call. Put differently: use Sol where you would put your best analyst. Do not use Sol to summarize meeting notes.
Terra (the enterprise default):
- Software development assistance, code generation, and debugging across engineering teams
- Customer-facing support escalations requiring substantive reasoning
- Marketing content generation and editing at scale
- Internal knowledge base Q&A where accuracy matters but doesn't require frontier-level reasoning
- Competitive analysis, market research summaries, and business intelligence workflows
Terra matches GPT-5.5 performance at half the price. For most of what enterprise teams actually need AI to do daily, Terra is the responsible default until you have data proving a task needs Sol.
Luna (high-volume, structured workloads):
- Document classification and routing at scale
- First-pass extraction from invoices, contracts, or structured forms
- Intent detection in customer service pipelines before human escalation
- Batch processing where throughput and cost-per-unit dominate quality requirements
- Internal search reranking and retrieval assistance
Luna enables workflows that would be cost-prohibitive at Sol or even Terra prices. At $1 per million input tokens, Luna opens AI access to data volumes that never made economic sense before.
The Cost Math Enterprise Leaders Need
Let's make this concrete. Assume an enterprise team runs 2,000 API calls per day, with an average of 12,000 input tokens and 3,000 output tokens per request.
At Sol pricing: That is $5 × 24M tokens/day input = $120/day, plus $30 × 6M output tokens/day = $180/day. Total: $300/day, or ~$109,000/year.
At Terra pricing: Half of Sol. $60/day input + $90/day output = $150/day, or ~$54,750/year.
At Luna pricing: $1 × 24M = $24/day input, $6 × 6M = $36/day output. $60/day, or ~$21,900/year.
The difference between using Sol versus Luna for the same 2,000 daily requests: $87,100/year from a single workflow. Multiply that across ten workflows running across different teams, and the model selection decision becomes a material line item in the AI budget conversation.
The correct approach is tiered deployment: identify your five to ten highest-value workflows and put Sol on those. Move everything else to Terra. Run bulk processing on Luna. Most enterprise teams that audit their AI stack honestly will find they can shift 60 to 70% of volume to Terra or Luna without any meaningful quality degradation.
Access Tiers by Surface
The access model varies depending on whether you are using ChatGPT, the ChatGPT Work and Codex surface, or the API directly.
ChatGPT (Plus, Pro, Business, Enterprise): Sol is available at medium and higher reasoning efforts. Pro and Enterprise subscribers can additionally access Sol Pro mode.
ChatGPT Work and Codex: Free and Go users access Terra by default. Paid users select among all three tiers and set effort level per model. The "max" reasoning effort is available to all paid users with GPT-5.6 access.
API: All three tiers — Sol, Terra, and Luna — are available. Programmatic Tool Calling (model-written JavaScript running in an isolated V8 runtime, no network access) and a multi-agent beta both live in the Responses API. This is the path for enterprises building custom agentic workflows where cost and control matter most.
For enterprises on the OpenAI Business or Enterprise plan, admin-level controls allow IT leaders to restrict which tier employees can access from ChatGPT. That is the governance lever CIOs have been waiting for: stop the entire org from defaulting to Sol when a document summary is all that's needed.
The Agentic Workload Warning
Ultra mode deserves special attention — and special caution. Sol's ultra mode runs four subagents in parallel to tackle complex tasks, pushing Terminal-Bench 2.1 performance from 88.8% to 91.9%. That is genuinely impressive.
The cost implication of running four agents in parallel is four agents worth of token consumption, potentially chained across multiple reasoning steps. A single ultra-mode workflow can easily consume 10x the tokens of a standard single-turn query.
Gartner's analysis of agentic AI costs found that chained multi-agent workflows burn five to thirty times more tokens than a single chatbot query. GPT-5.6's ultra mode falls squarely into that range — by design, and for good reason on complex tasks. But enterprises that enable ultra mode without clear governance will face budget surprises.
Before deploying Sol in ultra mode at scale: define the trigger conditions, set token consumption limits per workflow, and route only the subset of tasks that genuinely require multi-agent reasoning. A $30 output rate applied across unconstrained agentic loops is how enterprise AI budgets blow up without anyone noticing until the monthly bill arrives.
What This Means for the CFO
The three-tier structure changes the AI budget conversation. When there was effectively one model, cost control meant restricting access. Now there is a cost-optimized path for every category of work, which means the right CFO question is no longer "how do we spend less" but "are we paying the right price for each workload?"
The framework for finance leaders: Every AI use case should be tagged to a tier. That tagging should be validated periodically, because workload patterns evolve as teams get more comfortable with AI tooling. A task that seemed high-complexity six months ago may be a strong Luna candidate today.
Chargeback models become more nuanced in this world. Business units should see not just AI costs but tier-level breakdowns. Engineering spending disproportionately at Sol prices for tasks that qualify as Terra is a governance issue, not just a budget issue.
OpenAI's updated credit usage analytics and spend controls in ChatGPT Enterprise give administrators the visibility to make this governance possible. The tools exist. Using them is a choice.
What This Means for the CTO and CIO
Three decisions need to be made before GPT-5.6 goes into production:
1. Default tier policy. What tier does the organization default to when no specific guidance is given? The answer for most organizations should be Terra, not Sol. Sol access should require justification tied to a specific high-value workflow.
2. Ultra mode governance. Who can trigger ultra mode, for which workflows, and with what token caps? This should be written policy, not an assumption.
3. API versus ChatGPT access. For teams building custom workflows, the API's three-tier access with Programmatic Tool Calling enables cost optimization that the ChatGPT interface does not. Technical teams should be routed to API deployments where tier selection can be controlled programmatically rather than left to individual preference.
The benchmarks confirm one more thing for technical leaders: where SWE-Bench Pro performance matters to your engineering workflow — and 15 points is a significant gap — Claude Fable 5 or Mythos 5 remains the stronger choice for automated code review. GPT-5.6 Sol wins on agentic and reasoning tasks. The models are not interchangeable across all enterprise use cases, and the winning strategy is rarely betting everything on a single vendor.
The Bottom Line
GPT-5.6 is a genuine capability step forward. Sol leads on the task types that matter most for complex enterprise workflows: agentic reasoning, long-horizon professional tasks, computer use, and complex coding. Terra delivers most of that capability at half the price. Luna makes AI economics viable for high-volume bulk workloads.
The enterprise teams that get this right will build a tiered deployment model in the next 90 days. The ones that do not will absorb Sol pricing across workloads that Terra handles just as well — quietly, at volume, month after month.
The 5x cost difference between Sol and Luna is not a footnote. It is the enterprise AI strategy question of 2026.
Sources: OpenAI GPT-5.6 Preview announcement (July 9, 2026), OpenAI published evaluation tables, MarkTechPost GPT-5.6 analysis, Gartner agentic AI token cost analysis.
