theLLMs

Last checked: 2026-07-07

Scope: Global. GPT-5.6 preview data as of June 26, 2026; benchmark results from SIE Terminal-Bench 2.1 and CTF evaluations; access restrictions subject to White House safety review timeline.

AI draft model: qwen3.6:35b

AI review model: qwen3.6:35b

Key terms:

>_ large language model (LLM) >_ token >_ system prompt >_ open weights >_ JSON mode >_ reasoning model
Hero image for OpenAI Retires Version Numbers for Sol, Terra, and Luna Tier Naming System

OpenAI Retires Version Numbers for Sol, Terra, and Luna Tier Naming System

TL;DR

OpenAI has retired its traditional versioned model numbering in favor of three durable capability tiers—Sol, Terra, and Luna—where the tier name signals capability level and the number now indicates model generation rather than a fixed product line. GPT-5.6 lands at the top of Sol, the flagship tier with the strongest safety stack, while Terra offers roughly GPT-5.5-level performance at half the cost and Luna serves as the entry-level tier at about 20% of Sol’s price. The tier system consolidates OpenAI’s previously separate “o” series and GPT-5.x lines under a single taxonomy, and the model family is currently rolling out to a ~20-partner trusted preview with general availability planned in the coming weeks.

The Announcement

On June 26, 2026, OpenAI unveiled GPT-5.6 — not as a single model, but as a three-tier family under a new naming system that abandons versioned suffixes in favor of durable capability tiers. At the press event, OpenAI described the tier names as labels that “won’t change as models get better.” The number still tracks generation (this is the 5.6 generation), while the tier name signals capability. Sol is the flagship tier, Terra is balanced, and Luna is the entry-level option. The announcement arrived the same week as reporting on U.S. government gating, meaning the models were initially reviewed by government officials before being released to a small group of trusted partners. OpenAI began a limited preview to roughly 20 vetted enterprise and developer partners, with general availability — including ChatGPT, Codex, and API access — planned “in the coming weeks.”

The Sol tier is described as OpenAI’s “newest flagship model for developers and enterprises,” built for frontier reasoning and long-horizon agentic work. Terra is positioned as a “balanced model for everyday work” with performance broadly comparable to GPT-5.5 at half the cost. Luna rounds out the family as the “fast and affordable model” for high-volume, cost-sensitive tasks, with capabilities slightly above GPT-5.4. The move consolidates OpenAI’s previously separate “o” series (o1, o3, o3-pro) for reasoning and the GPT-5.x line for general-purpose models under one unified tier taxonomy.

The Tier Breakdown

The three tiers map to distinct capability and price bands, creating a product stack that gives developers and enterprises meaningful choice at inference time.

Sol is the top-tier model, optimized for complex reasoning, coding, science, and cybersecurity work. It introduces a new “max reasoning effort” mode and the aforementioned Ultra mode. Sol posts the strongest results across benchmarks — including 91.9% on TerminalBench 2.1 when Ultra mode is enabled, compared to 88.8% in plain mode. Sol also scores 96.7% on CTF cybersecurity benchmarks, which OpenAI calls state-of-the-art for long-horizon security tasks including vulnerability research and exploitation.

Terra is the workhorse tier. Independent analyses place Terra’s performance slightly below GPT-5.5 but at half the per-token cost, making it the likely default for most production workloads. OpenAI’s own characterization describes Terra as a “balanced model for everyday work” — it does not have Ultra mode or the highest reasoning budget, but it covers the vast majority of API use cases at a fraction of Sol’s price.

Luna is the economy tier. It outperforms GPT-5.4 and targets speed and cost efficiency over peak reasoning quality. Luna is best suited for high-volume tasks where latency and cost matter more than absolute accuracy — classification, routing, summarization, and other throughput-heavy patterns.

Together, the tiers create a spectrum: developers can route simple tasks to Luna, moderate work to Terra, and frontier-reasoning or security-critical tasks to Sol, potentially reducing overall inference costs by 60–80% compared to a single-model strategy.

Pricing Structure

The API pricing for GPT-5.6 Sol, Terra, and Luna reflects a wide capability-cost spread:

TierInput ($/1M tokens)Output ($/1M tokens)
Sol$5.00$30.00
Terra$2.50$15.00
Luna$1.00$6.00

Terra’s pricing is exactly half of Sol’s across the board. Luna sits at about 20% of Sol’s cost for both input and output tokens. This spread is structurally significant: it means that most production traffic — which typically consists of routine tasks like summarization, extraction, or conversational turns — can be routed to Luna or Terra without materially impacting quality, while reserving Sol for the small fraction of prompts that genuinely require frontier reasoning capability.

For context, the previous GPT-5.5 tier was priced significantly closer to Sol. Luna at $1/$6 represents OpenAI’s first entry-level API tier and marks a deliberate move toward cost segmentation that competitors have yet to match.

Ultra Mode and Reasoning Architecture

GPT-5.6 introduces two significant reasoning-related features exclusive to the Sol tier: max reasoning effort and Ultra mode. Max reasoning effort gives Sol more time to reason deeply on difficult tasks, analogous to OpenAI’s previous “extended thinking” but integrated directly into the API rather than gated behind a toggle.

Ultra mode goes further. Rather than relying on a single model pass, Ultra mode farms subtasks out to internal subagents and recombines the results — effectively turning Sol into a multi-agent system for complex work. This is a fundamental shift from OpenAI’s single-model execution paradigm. In practice, Ultra mode is what lifts Sol’s TerminalBench 2.1 score from 88.8% (plain) to 91.9%, representing a 3.1-point gain on a benchmark where the margin between competing frontier models is often measured in fractions of a point.

Neither max reasoning effort nor Ultra mode is available on Terra or Luna. These tiers use standard single-pass inference. This architectural split means that the tier distinction is not purely about model size or parameters — it reflects different execution strategies, with Sol offering an agentic execution pattern that the cheaper tiers do not support.

Safety and Government Gatekeeping

GPT-5.6’s safety profile is notable on two fronts: its internal classification and its external access controls.

On the internal side, OpenAI assigned GPT-5.6 Sol a “High” rating (rather than “Critical”) on its own exploit-composition scale. In OpenAI’s framework, a “High” rating means the model can identify bugs, draft exploit scaffolds, and perform vulnerability research in codebases like Chromium and Firefox — but in testing, it could not chain a full multi-stage attack. This is the first time OpenAI has published this classification for a GPT-5.x generation model, and the “High” (vs. “Critical”) designation suggests the company’s safety evals improved meaningfully relative to GPT-5.5’s assessment.

On the external side, U.S. government review has gated the initial rollout. OpenAI confirmed that GPT-5.6 was first reviewed by U.S. government officials before any public or partner access, and the initial trusted-preview program of ~20 partners was chosen with government guidance. TechCrunch and VentureBeat both reported that the U.S. request for pre-release review was part of a broader executive order framework for frontier models, and that OpenAI stated publicly it “doesn’t think restrictions should be the norm” — a clear signal that the gating is a one-time event for this launch, not a precedent for future releases.

The safety system card for GPT-5.6 preview is published at the OpenAI Deployment Safety Hub and includes the full list of evaluations performed prior to launch.

What Comes Next

Several open questions will define the trajectory of the Sol/Terra/Luna system:

GA rollout timeline. OpenAI has said general availability is coming “in the coming weeks” as of June 26, but it remains unclear which tiers will be available at launch. Sol, with its government gating, may arrive later than Terra and Luna. Users should watch for the ChatGPT and Codex launch announcements to determine which tier(s) open first.

Competitor response. How Anthropic’s Claude and Google’s Gemini families respond to the tier naming convention will be a telling signal. If competitors adopt a similar capability-tier approach, the Sol/Terra/Luna system could become an industry standard for model labeling. If not, it may remain an OpenAI-specific experiment.

Benchmark data. With only ~20 partners in the preview, independent benchmark data for Sol, Terra, and Luna is limited. As general availability opens, third-party evaluations on TerminalBench, LiveBench, and other leaderboards will clarify whether the claimed tier differentiation holds up under independent scrutiny.

Naming durability. OpenAI’s press release explicitly frames Sol, Terra, and Luna as “durable capability tiers that won’t change as models get better.” If the next generation (GPT-6) retains these tier names while introducing new internal architectures, the naming system will have proven its durability. If OpenAI rebrands the tiers for the next generation, it was a one-off experiment — and the GPT-5.6 generation becomes a footnote in OpenAI’s naming history.

The End of Suffixes

OpenAI’s Sol/Terra/Luna system marks a decisive break from the versioned-suffix era that defined the company’s identity for years. Where GPT-4, GPT-4o, and GPT-5.5 each carried a transient label that would be superseded by the next release, the durable tier names are designed to persist across generations — Sol remains Sol whether it runs on a 5.6-era architecture or something entirely unrecognizable five years from now. If the experiment holds, developers and enterprises will begin to think of model capability in terms of tier rather than version, and the Sol/Terra/Luna taxonomy could anchor the broader industry’s approach to model labeling.

The pricing and architectural split between tiers reinforces this shift. With a 30:1 input-output spread between Sol and Luna, OpenAI is not just selling different model sizes — it is selling different execution strategies. Sol’s multi-agent Ultra mode and max-reasoning budget represent an agentic execution paradigm that the cheaper tiers deliberately do not support. The result is a product stack where cost segmentation and capability segmentation are aligned: routine work routes to Luna, moderate tasks go to Terra, and frontier reasoning is reserved for Sol, with the potential to cut overall inference costs by 60–80% relative to a single-model approach.

Government gatekeeping adds an unprecedented dimension to the rollout. U.S. officials reviewed the model before any partner access, and the initial ~20-partner preview was selected under an executive-order framework that OpenAI has signaled is a one-time event rather than a new precedent. Combined with the “High” safety classification — the first time OpenAI has published such an assessment for a GPT-5.x model — the launch reveals a company navigating both product strategy and regulatory expectations simultaneously.

What remains to be seen is whether Sol/Terra/Luna proves durable or becomes a footnote. Independent benchmark data will test the claimed tier differentiation. Competitor responses from Anthropic and Google will determine whether the naming convention spreads beyond OpenAI. And if the next generation retains these names, the system’s durability will have been proven; if it does not, the experiment will be remembered as a bold but temporary rebrand.

Methodology

  • Data checked: 2026-07-06
  • Sources consulted: OpenAI press release, OpenAI Deployment Safety Hub system card, Simon Willison’s weblog, TechCrunch, VentureBeat, TerminalBench 2.1 results (Towards AI), independent benchmark aggregators (explainx.ai, novakvm.com, macgpu.com, kingy.ai)
  • Assumptions: The three-tier capability model described by OpenAI is as stated in their press release and contemporaneous reporting. Pricing figures are as published by OpenAI on June 26, 2026. Government gating details are based on TechCrunch and VentureBeat reporting.
  • Limitations: Independent benchmark data is limited to ~20 partners in the preview period. Third-party evaluations on TerminalBench, LiveBench, and other leaderboards have not yet been published for the full GA release. Competitor response to the tier naming convention is forward-looking and has not yet occurred.
  • Jurisdiction: Global. No jurisdiction-specific regulatory requirements are covered beyond the U.S. executive order framework reported by TechCrunch and VentureBeat.

Source list

Trust Stack

  • Written by: qwen3.6:35b
  • Reviewed by: qwen3.6:35b (human review: pending)
  • Affiliation: None declared
  • Corrections: Contact us to report errors

Change log

  • 2026-07-07: Article assembled from intro, sections, and conclusion components. H1 title added. Source list completed with URLs for all entries.