GPT-5.6 Sol vs Terra vs Luna: OpenAI's New Naming Resets Coding Cost Tiers

A New Naming System, Three New Price Points

On June 27, 2026, OpenAI previewed the GPT-5.6 family on its research blog under the title "Previewing GPT-5.6 Sol: a next-generation model." The headline change is not just a new model — it is a new naming system. From this generation onward, OpenAI separates the generation number (5.6) from the capability tier (Sol, Terra, Luna), and each tier can advance independently. The Pro/Mini/Nano vocabulary is retired for the 5.6 line.

The three launch SKUs and their official per-million-token pricing:

• GPT-5.6 Sol — flagship. $5 input / $30 output. Same headline price as GPT-5.5 but with stronger agentic coding, biology, and cybersecurity benchmarks.
• GPT-5.6 Terra — balanced everyday tier. $2.50 input / $15 output. OpenAI claims "competitive performance with GPT-5.5 at 2x cheaper."
• GPT-5.6 Luna — fast and affordable. $1 input / $6 output. Lowest cost in the 5.6 family.

Access is in limited preview through the API and Codex to "trusted partners and organizations." OpenAI says broad availability via ChatGPT, Codex, and the public API is coming "in the coming weeks." The phased rollout is a response to a US government request related to cyber capabilities, not a paid-tier decision.

What the Naming Change Actually Signals

OpenAI's old Pro/Mini/Nano taxonomy mapped to size — Pro was the biggest, Nano was the smallest. The new Sun-system names map to role instead of size: Sol is the flagship, Terra is the everyday workhorse, Luna is the fast/cheap choice. The shift matters because each role can iterate at its own pace. You could see Terra get smarter in a 5.6.x point release without Sol changing, or Luna get faster without affecting Sol or Terra.

For cost planning, this changes one important thing: the gap between tiers is now stable as a multiplier rather than a function of model size. Sol → Terra is a 2x cost drop. Terra → Luna is roughly another 2.5x drop. That predictability lets teams budget by tier without re-doing the math every release.

The "Terra Is GPT-5.5 at Half Price" Claim

The most quoted line from the GPT-5.6 announcement is OpenAI's own framing: "Terra has competitive performance to GPT-5.5 while being 2x cheaper." If that holds in production — and the production data is not public yet — Terra is the obvious upgrade path for anyone currently running GPT-5.5 ($5/$30) on coding workloads. Same target output quality, half the bill.

The caveat: "competitive" does not mean "identical." OpenAI's own released benchmarks emphasize Sol, not Terra. On agentic coding tasks where the marginal token earns its keep, Terra likely matches 5.5 closely; on long-horizon reasoning chains where 5.5 was already weak, Terra may match or fall behind. The recommended migration pattern is to A/B test Terra against your current 5.5 prompts on a real subset of your workload before flipping the default.

Cost-Per-Coding-Task Math: Sol vs Terra vs Luna

To make the tier choice concrete, here is a rough cost model for a typical "AI coding agent fixes one bug" interaction — 25K input tokens (codebase context + history) and 5K output tokens (patch + explanation):

• Sol: 25K × $5/M + 5K × $30/M ≈ $0.275 per fix
• Terra: 25K × $2.50/M + 5K × $15/M ≈ $0.1375 per fix
• Luna: 25K × $1/M + 5K × $6/M ≈ $0.055 per fix

At 200 bug-fix interactions per developer per month — roughly one focused coding session — that's $55 (Sol), $27.50 (Terra), or $11 (Luna) per developer per month, before prompt caching. The 30-minute minimum cache life on 5.6 models can drop those by another 60-80% on repeated-context workflows.

When to Pick Each Tier

A practical rule of thumb, given what OpenAI has said publicly so far:

• Sol for end-to-end agent runs with long reasoning, cyber/security work, novel architecture decisions, and the new "ultra mode" subagent coordination. Anything where a single wrong turn costs hours of retry tokens.
• Terra as the new default for normal coding work — bug fixes, refactors, test generation, code review. Should replace GPT-5.5 as the everyday model for most teams.
• Luna for high-volume, lower-stakes operations — commit message generation, lint fixes, doc string writing, batch transformations across many files.

How Sol/Terra/Luna Stack Against the Field

Looking sideways at the competing flagship/mid/budget triads as of late June 2026:

• Sol ($5/$30) vs Claude Opus 4.8 ($5/$25) — same input price, Claude $5 cheaper on output. Gemini 3.1 Pro ($2/$12) is meaningfully cheaper but on different benchmarks.
• Terra ($2.50/$15) vs Claude Sonnet 4.6 ($3/$15) — slightly cheaper input, same output. Vs Gemini 3.5 Flash ($1.50/$9) — Gemini wins on price.
• Luna ($1/$6) vs DeepSeek V4 Pro ($0.435/$0.87) — DeepSeek is materially cheaper for open-weight workloads. Vs Gemini 3 Flash ($0.5/$3) — also cheaper than Luna.

In short: OpenAI is competitive but not dominant on price at any single tier. The value proposition of the 5.6 family is the combination of three tiers under one consistent prompt-caching contract, plus the new ultra/multi-agent reasoning features that ladder up across the tiers.

Bottom Line

The GPT-5.6 family is less about raw capability gains than about repricing the OpenAI ladder. Sol holds the line at $5/$30, Terra delivers most of 5.5 at half the price, and Luna gives developers an unambiguous "cheap fast OpenAI" option for the first time at $1/$6. Until GA, the practical move is to plan migrations: Terra likely replaces 5.5 as your default, Luna replaces 5.4-mini or 4o-mini for high-volume tasks, and Sol stays reserved for the agent runs that need the new reasoning ceiling. We've already added all three to our pricing dataset and the cost estimator.