OpenAI announced GPT-5.6 this morning — three models, three price points, and a procurement decision most enterprise teams will get wrong in the next 48 hours. The headlines are all about Sol, the flagship. But if you're buying AI compute for your business, the most important model in this launch isn't Sol. It's Terra.
Here's the complete breakdown of what shipped, what it costs, and how to make the right call for your organization.
What OpenAI Actually Shipped
The GPT-5.6 family has three models — each designed for a different position on the capability-cost curve:
Sol is the flagship. OpenAI describes it as their strongest model yet, with new state-of-the-art performance on coding, biology, and cybersecurity benchmarks. It introduces two new inference modes: "max reasoning effort" (gives Sol more time to reason through complex problems) and "ultra mode" (spawns subagents to parallelize complex work). Price: $5 per million input tokens, $30 per million output tokens.
Terra is the balanced tier. OpenAI's positioning is precise: competitive performance to GPT-5.5 at half the price. Terra targets the everyday enterprise workloads — document processing, customer support automation, code review, knowledge retrieval. Price: $2.50 per million input tokens, $15 per million output tokens.
Luna is the speed/cost tier. Strong capability at the lowest cost OpenAI has offered for a frontier-class model. Price: $1 per million input tokens, $6 per million output tokens.
All three are in limited preview as of today. General availability is expected in the coming weeks. OpenAI coordinated with the U.S. government before this release — a first-of-its-kind pre-release review process that OpenAI says it doesn't intend to make a long-term default.
The Numbers That Drive the Procurement Decision
The critical number in this announcement is not in the benchmark tables. It's in the Terra pricing.
GPT-5.5 — the previous generation balanced model — was priced at $5/$30 per million tokens. Terra delivers competitive performance at $2.50/$15. That is a 50% cost reduction on the most common enterprise AI workload: high-volume, moderate-complexity inference.
For a company running 10 billion tokens per month across document processing, internal search, and customer-facing automation — the delta between GPT-5.5 and Terra is roughly $250,000 per month. On an annualized basis, that is $3 million in operating cost before any infrastructure optimization.
Sol, by contrast, is priced identically to GPT-5.5: $5/$30. You are not paying more for Sol than you paid for the previous flagship. But you are also not paying less — and for most enterprise workloads, the difference in raw capability between Sol and Terra will not show up in production metrics.
Luna sits at $1/$6. For high-volume, low-complexity tasks — classification, extraction, summarization of short documents — Luna is the default choice, not Sol or Terra.
The Three-Tier Model Is Actually a Decision Framework
OpenAI is giving you three tools and implicitly telling you to route workloads by task complexity. Most enterprise teams will miss this and default to Sol everywhere — the same mistake organizations made when they defaulted to GPT-4 for every workload when GPT-3.5 would have handled 80% of their traffic at a fraction of the cost.
Here is a practical routing framework:
Use Luna for: Classification tasks, intent detection, short-document summarization, FAQ answering, simple data extraction from structured inputs. Volume is the primary driver here — if you are running tens of millions of requests per day, Luna's economics are dramatically better.
Use Terra for: Multi-step reasoning on business documents, code review and generation, customer support workflows with context retrieval, report drafting, vendor contract analysis, and most RAG-based applications. This is the 60-70% of enterprise AI workload where GPT-5.5 was your previous choice.
Use Sol for: Complex agentic workflows where errors have high downstream cost, security vulnerability research, scientific or technical analysis requiring deep reasoning, multi-stage tasks where the "ultra mode" subagent capability materially reduces completion time, and any production workflow where you have benchmarked Sol to produce meaningfully better outputs than Terra.
The honest question to ask before defaulting to Sol: can you measure the difference? If your QA metrics cannot detect the performance gap in production, you are paying double for unmeasured improvement.
The Cerebras Partnership: 750 Tokens Per Second
The speed story in this launch is genuinely interesting for a specific enterprise use case.
OpenAI is launching Sol on Cerebras — Cerebras builds purpose-built AI accelerator chips with dramatically different memory architecture than NVIDIA GPUs — at up to 750 tokens per second. To put that in context: a 1,500-word article generates in under two seconds. A 50-page contract analysis completes in under a minute.
This matters for interactive applications where latency is the constraint. Customer-facing chatbots, real-time document review in legal workflows, live code completion in developer tools — these use cases have been constrained by the 30-50 tokens/second ceiling that characterized previous generation APIs. At 750 tokens/second, that constraint disappears for most interactive workloads.
The caveat: Cerebras availability starts with select customers while capacity expands. If your use case is latency-critical and your volume justifies priority access, this is worth a direct conversation with your OpenAI account team.
For the majority of enterprise batch workloads — nightly data processing, asynchronous document enrichment, scheduled report generation — this speed improvement is irrelevant. Your bottleneck is I/O and orchestration, not inference throughput.
Ultra Mode and What It Means for Agentic Workflows
Sol's "ultra mode" is the most architecturally interesting piece of this launch. OpenAI describes it as going "beyond the capabilities of a single agent by leveraging subagents to accelerate complex work."
In practice, this means Sol can decompose a complex task, spin up parallel sub-processes to handle components simultaneously, and synthesize results — all within a single API call. The analogy is moving from a single senior analyst to a team of analysts with a coordinator.
For enterprise teams that have been building agentic workflows — multi-step processes that chain together search, analysis, writing, and decision-making — ultra mode potentially collapses what previously required custom orchestration code into a single model call.
The business case is straightforward: if your current agentic pipeline takes 15 minutes and requires five sequential API calls, ultra mode may handle the same task in four minutes with one call. The cost per task goes up (more compute per call), but the latency and engineering overhead drop significantly.
This is still early. Ultra mode is in preview and production reliability data does not yet exist at scale. Do not redesign your architecture around it before you have benchmarked it against your specific workflows.
Cache Pricing Change: Read the Fine Print
This launch includes a pricing model change that will catch enterprise teams off guard.
For GPT-5.6 and all future models, cache writes are now billed at 1.25x the model's uncached input rate. Cache reads continue to receive the 90% discount on cached input.
Under the previous pricing model, cache writes were effectively free. You loaded context into cache, then paid the discounted rate on reads. Now there is a write cost.
The impact depends on your caching patterns. If you maintain persistent system prompts or large context windows that you refresh infrequently and read many times, the math still strongly favors caching. If you are doing single-turn queries with no repeated context, caching was not helping you anyway.
The threshold where caching remains economically rational: if you read the same cached content more than roughly eight times per write, you still come out ahead. Below that threshold, you are paying a premium to cache content you barely reuse.
Audit your current caching implementation before migrating to GPT-5.6. Teams running aggressive caching strategies may see a bill increase in the transition period until they optimize their cache hit ratios.
The Government Coordination Piece: What It Signals
OpenAI's pre-release coordination with the U.S. government before today's launch is worth noting for enterprise teams in regulated industries.
OpenAI previewed GPT-5.6's capabilities to government stakeholders before public release. The company explicitly says this was a one-time arrangement tied to the current policy environment around the cyber Executive Order framework — not a permanent precedent.
For CISOs and compliance teams: this signals that AI model releases are increasingly being treated as having national security implications, particularly models with advanced cybersecurity capabilities. OpenAI notes that Sol is their "most capable model yet for cybersecurity" — it can identify vulnerabilities and exploitation primitives, though it does not autonomously produce functional full-chain exploits under tested conditions.
For organizations in defense, critical infrastructure, financial services, and healthcare: the policy direction of travel is toward more structured AI model governance, not less. Building internal AI governance frameworks now — before regulatory requirements force the issue — positions you ahead of the curve rather than behind it.
What Enterprise Leaders Should Do Right Now
For CIOs and CTOs: Add GPT-5.6 to your model evaluation queue. The priority workload to test against Terra is your current highest-volume GPT-5.5 deployment. If Terra matches GPT-5.5 performance on your specific task distribution, you have a straightforward cost reduction on the table. Do not wait for GA to start benchmarking — preview access is available to API customers.
For CFOs: The cost reduction story on Terra is real and measurable. Request a 30-day cost comparison from your AI engineering team that models current spend at GPT-5.5 rates against projected Terra rates on the same workload volume. This is not a speculative future saving — it is a pricing delta you can calculate today with existing usage data.
For AI/ML Engineering Teams: The cache pricing change is the immediate technical action item. Review your current cache write-to-read ratios before migrating to GPT-5.6. Audit ultra mode availability for any agentic pipelines where latency is a primary constraint. Benchmark Luna against your classification and extraction workloads — the $1/$6 pricing changes the economics of use cases you may have previously rejected as too expensive.
For Security Teams: Review OpenAI's GPT-5.6 Preview system card, which documents the layered safeguard architecture in detail. Sol's cybersecurity capabilities are dual-use — they benefit your defensive security posture but also require appropriate access controls if you are deploying Sol in environments accessible to external users.
The Bottom Line
GPT-5.6 is a genuine step change in OpenAI's lineup — Sol introduces meaningful new capabilities, Terra delivers near-identical performance to the previous generation flagship at half the cost, and Luna opens up use cases that were previously uneconomical at scale.
The procurement mistake to avoid: defaulting to Sol because it is the flagship. The right answer for most enterprise workloads is Terra, with Luna handling the high-volume, low-complexity tail. Sol earns its place in your architecture for specific workloads where its advanced reasoning and ultra mode capabilities produce measurable, production-validated improvements over Terra.
Limited preview is live today. General availability is expected within weeks. The teams that benchmark their workloads against all three tiers during preview will be positioned to optimize their model spend from day one of GA — rather than spending the next six months auditing a decision made in haste when headlines said "new flagship."
OpenAI pricing and capability details sourced from the official GPT-5.6 preview announcement and GPT-5.6 Preview system card.
