OpenAI just handed enterprise buyers something they've been asking for since ChatGPT launched: a clear, tiered model lineup with predictable per-token pricing and a dedicated productivity agent for business users. On July 9–10, 2026, the company unveiled ChatGPT Work — a new multi-step agent — alongside GPT-5.6, a three-model family named Sol, Terra, and Luna. For enterprise leaders who've been navigating a confusing alphabet soup of model versions, this is the clearest buying signal yet.
Here's what actually changed, why it matters for your organization, and how to make the right call on which tier to deploy.
What Is ChatGPT Work?
ChatGPT Work is OpenAI's answer to the productivity gap between a general chat tool and a full-blown software engineering agent. It's designed for tasks that are too long or too complex for a standard chat session — multi-step research projects, cross-file document analysis, extended data workflows, and scenarios where the model needs to coordinate across multiple tools over a longer time horizon.
Think of it as the bridge between "I asked ChatGPT a question" and "I deployed an AI agent." It's available on web, mobile, and desktop, and it ships inside the updated ChatGPT desktop application alongside the newly integrated Codex environment.
What makes this enterprise-relevant is what's unlocked at higher access tiers. Business and Enterprise accounts get access to all three GPT-5.6 model tiers within ChatGPT Work, including the new "ultra" mode for Pro and Enterprise users — which coordinates multiple subagents across parallel workstreams on the same task.
That last part is significant. If your team has been manually breaking complex tasks into subtasks and routing them to different tools or team members, ultra mode is the first time an off-the-shelf AI product can do that parallelization automatically.
The GPT-5.6 Model Family: Sol, Terra, Luna
OpenAI has moved away from numbered sub-versions toward capability tiers that can evolve independently. The three models introduced with GPT-5.6 are:
Sol — The flagship. State-of-the-art performance across coding, knowledge work, cybersecurity, and scientific analysis. OpenAI positions it as outperforming previous frontier models with fewer tokens at lower estimated cost-per-task. API pricing: $5 per million input tokens / $30 per million output tokens.
Terra — The balanced everyday model. Designed for the bulk of enterprise knowledge work: drafting, summarization, analysis, reporting. API pricing: $2.50 per million input tokens / $15 per million output tokens.
Luna — The fast, affordable option for high-volume, lower-complexity tasks: classification, data extraction, simple Q&A at scale, routing decisions. API pricing: $1 per million input tokens / $6 per million output tokens.
This structure isn't just a product naming exercise. It's a signal that OpenAI is explicitly building for enterprise cost optimization. Most organizations running AI at scale don't need frontier-level reasoning for every request — they need to match the right capability to the right task and manage the cost curve accordingly.
Why Tiered Pricing Changes the Enterprise Calculus
Until recently, enterprise AI procurement looked like a seat-license problem: how many users, which plan, what's the monthly contract? The shift to per-token pricing with explicit capability tiers changes the procurement conversation entirely.
For CTOs and architects: you now have a decision matrix. Sol for complex reasoning tasks that touch compliance, strategic analysis, or production-grade code review. Terra for the bulk of your internal tool stack — HR drafting, finance summarization, legal document review at moderate complexity. Luna for any workflow where speed matters more than depth: routing tickets, classifying customer feedback, extracting structured data from forms.
The right architecture isn't picking one tier and deploying it everywhere. It's building a routing layer that sends requests to the appropriate model based on task complexity. A CTO conversation I had recently framed it well: "We were spending $18K a month on frontier model calls, 60% of which were simple classification tasks that a smaller model handles just as well." Deploying Luna for classification alone would cut that 60% of spend by roughly 80%.
For CFOs: the per-token structure makes AI spend forecastable in a way that per-seat contracts never could. You can model token consumption by department, set spending alerts (which Anthropic recently introduced on their side), and tie AI cost directly to business output — tokens per processed document, cost per customer interaction, ROI per workflow automated.
The catch is that per-token forecasting requires instrumentation. If your engineering team isn't already logging model calls by use case, this is the moment to build that capability. Flying blind on AI spend with per-token pricing will create the same budget surprises that per-seat licenses did — just faster.
The Ultra Mode: Parallel Agents Are Now a Product
The "ultra" feature available to Pro and Enterprise users in ChatGPT Work deserves its own section, because it represents a architectural shift in how AI handles complex work.
Standard AI interactions are sequential: you ask, it responds, you ask again. Ultra mode coordinates multiple subagents concurrently on the same task and synthesizes their output into a unified result. For enterprise use cases, this matters in several scenarios:
Due diligence workflows. A legal team reviewing an acquisition could run parallel agents simultaneously analyzing the target's contracts, financial disclosures, IP documentation, and regulatory filings — then receive a synthesized risk summary in a fraction of the time a sequential process would take.
Multi-source research and intelligence. Strategy teams pulling market analysis from multiple data sources, analyst reports, and internal data can run parallel extraction and get a synthesized briefing rather than manually aggregating five separate outputs.
Complex code reviews. Engineering teams using Codex can run concurrent analysis on security vulnerabilities, performance bottlenecks, and documentation gaps in the same codebase, with a unified review output.
The API also gets multi-agent capabilities in this release. Developers can now run concurrent subagents via the Responses API and receive a synthesized result in a single request — a pattern that was previously possible only with custom orchestration frameworks.
The Codex Integration: What It Means for Engineering Teams
One significant change buried in the announcement is the merging of Codex into the main ChatGPT desktop application. The existing Codex app is now the new ChatGPT desktop app — users who want to keep the Codex experience can keep the Codex icon and preferences, but they're running on the same unified codebase.
New capabilities added to Codex with this release:
- Direct Markdown and code editing in-app with inline annotations
- GitHub pull request review in the sidebar with reviewer feedback alongside the diff
- Multi-repository project support
For CIOs managing software engineering productivity, this signals something important: OpenAI is consolidating its enterprise surface area into a single product that covers both knowledge-work AI (ChatGPT Work) and code-work AI (Codex) from one interface, with shared plugin management and account controls. This reduces the "tool sprawl" problem that's been emerging as engineering teams adopted Codex separately from the ChatGPT tools other departments were using.
Whether your organization centralizes on one vendor or maintains a multi-vendor AI strategy, this consolidation is worth tracking. The fewer integration points your IT team has to manage, the lower the administrative overhead of AI deployment at scale.
Enterprise Access: What's Available at Each Tier
To remove ambiguity on access tiers:
Free and Go users: Access to GPT-5.6 Terra in ChatGPT Work and Codex.
Plus users: Sol, Terra, and Luna in ChatGPT Work and Codex. "Max" effort available.
Business users: Same as Plus. Sol, Terra, and Luna all accessible in ChatGPT Work and Codex.
Enterprise users: Sol, Terra, and Luna in ChatGPT Work and Codex. "Ultra" mode available in ChatGPT Work (parallel multi-agent). GPT-5.6 Sol Pro available for the highest-quality results on complex tasks.
GPT-5.4 is being retired on July 23. GPT-5.5 models remain available during transition.
The Decision Framework for Enterprise Leaders
If you're evaluating this announcement for your organization, here's how I'd structure the decision:
For organizations not yet on a formal AI contract: This is the clearest moment to evaluate OpenAI's Business or Enterprise tier with defined intent. The three-tier model structure gives you a cost optimization lever that wasn't available a year ago. Start with Terra as your default and route complex analytical tasks to Sol. Reserve Luna for high-volume, lower-stakes automation.
For organizations already on enterprise contracts: Review your current model usage against the new tier structure. If you're defaulting all calls to the frontier model, you likely have cost optimization opportunities available immediately. Build a simple routing heuristic: task complexity determines model tier.
For CTOs evaluating multi-agent architecture: The Responses API multi-agent capability now available with GPT-5.6 is worth a proof-of-concept run on your most complex parallel workflows. The question isn't whether parallel agents are faster — they are — but whether your data access controls and compliance posture support autonomous multi-agent execution on sensitive enterprise data.
For CFOs: Request a per-token cost projection from your engineering team before the next budget cycle. The shift from seat licensing to consumption pricing requires different forecasting tools. If your team can't give you token consumption by department and use case, that's the infrastructure gap to close first.
What This Means for the Broader AI Landscape
OpenAI's move to a named capability-tier model (Sol/Terra/Luna instead of numbered sub-versions) mirrors what Anthropic has done with the Haiku/Sonnet/Opus progression and what Google has done with the Gemini tiering. The message across all three major enterprise AI vendors is now consistent: pick the capability tier that fits your use case, not the absolute best model for every task.
The strategic winner in enterprise AI procurement over the next 18 months will not be the vendor with the single best frontier model. It will be the vendor whose tiered lineup, pricing structure, and enterprise governance controls let large organizations deploy AI cost-effectively at scale across multiple departments and use cases.
ChatGPT Work and the GPT-5.6 family represent OpenAI's clearest positioning yet in that race. The question for every enterprise leader now is whether your internal AI strategy is sophisticated enough to use that tiering intentionally — or whether you'll default to the frontier model everywhere and pay the cost.
The Bottom Line
OpenAI shipped two important things today: a productivity agent built for complex multi-step enterprise work, and a model lineup with clear capability tiers and predictable per-token pricing. Both are meaningful for different parts of your organization.
For technical leaders, the multi-agent API and Codex integration give you architectural building blocks that were previously custom work. For business leaders, the tiered pricing structure creates the cost predictability that enterprise AI budgets have been missing. For everyone, the clarity of Sol, Terra, and Luna makes the model selection conversation considerably less murky than it was yesterday.
The organizations that move quickly to build routing logic — matching task complexity to model tier — will find meaningful cost advantages over those that default to frontier models for everything. That's the real productivity play buried inside today's announcement.
Sources: OpenAI GPT-5.6 announcement, 9to5Mac ChatGPT Work coverage
Follow me on LinkedIn or X for more enterprise AI analysis.
