GPT-5.6 Sol vs Terra vs Luna: Which Model Tier Should You Use?

On June 26, 2026, OpenAI announced GPT-5.6 — but instead of shipping one model, they shipped three: Sol (flagship), Terra (balanced), and Luna (fast and cheap). This marks a structural shift from the GPT-5/5.5 approach of "one model, choose a reasoning effort level" to "three models, choose a tier."

For teams already routing across providers, the question is straightforward: which tier goes where in your stack? This post breaks down the pricing, benchmarks, capabilities, and practical routing decisions across all three GPT-5.6 tiers — and how they compare to competitors you might already be using.

OpenAI-compatible means a provider exposes a chat-completions endpoint whose request and response shape matches the OpenAI API contract closely enough that an unmodified OpenAI SDK call works against it after swapping three values: API key, base URL, and model name. The minimum surface in practice is POST /v1/chat/completions with messages, model, and an OpenAI-shaped streaming response.

Sources: OpenAI GPT-5.6 Announcement, retrieved 2026-06-30; OpenAI Help Center — GPT-5.6 Preview, retrieved 2026-06-30; Anthropic Pricing, retrieved 2026-06-30; GPT-5.6 Developer Guide (LushBinary), retrieved 2026-06-30; VentureBeat Coverage, retrieved 2026-06-30.

TL;DR — Sol vs Terra vs Luna at a Glance

Prices from OpenAI, retrieved June 30, 2026. GPT-5.6 is currently in limited preview — availability is staged and may vary by account and region.

What Changed from GPT-5.5 to the 5.6 Family

GPT-5.5 shipped as a single model (plus a Pro variant for deeper reasoning). You controlled capability through reasoning effort settings.

GPT-5.6 replaces that with a named tier system: the generation number (5.6) tells you how new the model is, while the tier name (Sol, Terra, Luna) tells you where it sits on the capability/cost curve. OpenAI says these tier names are durable — future generations will use the same Sol/Terra/Luna naming, with each tier advancing on its own cadence.

The practical upshot: instead of tweaking reasoning_effort: "high" on one model, you now pick gpt-5.6-sol for hard problems and gpt-5.6-luna for high-volume work. This is more like the Anthropic model family (Opus/Sonnet/Haiku) than the old OpenAI "one model, many modes" approach.

Pricing comparison: GPT-5.5 vs GPT-5.6 tiers

Prices from OpenAI and TheRouter model catalog.

Sol at $5/$30 is roughly 40% cheaper than GPT-5.5 on input and 38% cheaper on output — while outperforming it on benchmarks. Terra matches GPT-5.5's capability at about half the cost. Luna opens up a price tier OpenAI didn't have before.

New prompt caching model

GPT-5.6 also introduces improved prompt caching:

• Cache writes billed at 1.25× the uncached input rate (vs. free writes on GPT-5.5)
• Cache reads continue to receive the 90% discount
• 30-minute minimum cache life (up from the variable TTL on GPT-5.5)
• Explicit cache breakpoints for deterministic caching

For long system prompts or few-shot examples, caching shifts the economics significantly. A system prompt that hits cache on every request costs 10% of its uncached price — making even Sol competitive with Luna on the input side for repeated prompts.

For details on caching strategies across providers, see our Prompt Caching Cost Optimization Guide.

Sol: The Frontier Flagship

Sol is for tasks where correctness matters more than cost. It leads on TerminalBench 2.1 (88.8%) and delivers OpenAI's strongest results on ExploitBench and GeneBench v1.

Sol Ultra is a compute-intensive high-effort mode that pushes Sol further, hitting 91.9% on TerminalBench 2.1. Treat Ultra as the mode for the hardest multi-step problems — it spends more compute per request and therefore costs more in practice (token consumption increases, though the per-token rate stays the same).

When to pick Sol

• Long-horizon agentic coding: autonomous multi-file edits, architecture decisions, complex debugging
• Cybersecurity workflows: vulnerability research, exploit analysis, patch development
• Biology/scientific research: genomics analyses, protein structure reasoning
• Any task where a wrong answer is expensive enough to justify the 5× cost over Luna

How Sol compares to competitors

TerminalBench 2.1 scores from OpenAI announcement, retrieved June 30, 2026. Scores are vendor-reported; independent verification pending.

Sol outperforms Claude Opus 4.8 on TerminalBench 2.1 by nearly 10 points at comparable pricing. Against GPT-5.5, Sol is both cheaper and more capable — a rare combination.

For Anthropic's latest model lineup and how it compares, see our Anthropic Claude API Guide.

Terra: The Balanced Default

Terra is what most teams should start with. OpenAI positions it as matching GPT-5.5 capability at roughly half the cost, which makes it the natural successor for general application traffic.

TerminalBench 2.1: 84.3%

That puts Terra dead even with Claude Fable 5 (also 84.3%) and ahead of GPT-5.5 (83.4%). At $2.50/$15.00 per million tokens, Terra undercuts GPT-5.5 by more than 3×.

When to pick Terra

• General chat and assistant workloads — the quality matches or beats GPT-5.5 at half the cost
• Tool-using agent flows — function calling, structured output, multi-turn conversations
• Content generation — drafting, summarization, translation
• Prototyping — develop against Terra, escalate to Sol only for production tasks that need it

The production default

If you're migrating from GPT-5.5 and want the simplest path, change model: "gpt-5.5" to model: "gpt-5.6-terra" and expect similar or better quality at lower cost. That's the pitch — and the TerminalBench numbers back it up.

from openai import OpenAI

client = OpenAI(api_key="sk-...")

response = client.chat.completions.create(
    model="gpt-5.6-terra",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain the CAP theorem with a real-world example."}
    ]
)
print(response.choices[0].message.content)

Luna: Fast, Cheap, and Surprisingly Capable

Luna is the tier for high-volume work where unit economics dominate. At $1/$6 per million tokens, it's the cheapest model OpenAI has ever offered at this capability level.

TerminalBench 2.1: 82.5%

That's above Claude Opus 4.8 (78.9%) — a frontier model from two months ago. Luna is positioned against lightweight models like Claude Haiku 3.5, Gemini Flash, and GPT-5 Mini, but its benchmark scores suggest it punches above its weight class.

When to pick Luna

• Classification and tagging — sentiment analysis, content moderation, intent detection
• Data extraction — pulling structured data from unstructured text
• Request routing — using a cheap model to classify incoming requests before escalating to a more expensive one
• High-volume summarization — processing large document sets where per-request cost matters
• First-pass filtering — Luna handles the easy majority so Sol only sees the hard tail

response = client.chat.completions.create(
    model="gpt-5.6-luna",
    messages=[
        {"role": "user", "content": "Classify this support ticket: 'My invoice shows the wrong amount for April.'"}
    ],
    response_format={"type": "json_object"}
)

Routing Decision Matrix

The three-tier system maps naturally to a cascading routing strategy:

Cascade pattern in practice

A common production setup processes requests through a confidence-based cascade:

1. Luna handles the easy majority (70–80% of traffic). Classification, extraction, and short replies.
2. Terra catches the middle tier (15–25%). Multi-turn conversations, tool-calling flows, and tasks that need more reasoning.
3. Sol handles the hard tail (5–10%). Flagged by a complexity signal, a confidence check, or an explicit user request for maximum quality.

This keeps your blended cost close to Luna while preserving Sol-grade quality where it counts. The per-request difference between Luna ($1/$6) and Sol ($5/$30) is 5× — so routing 80% of traffic to Luna instead of Sol cuts your bill by ~65%.

For more on building fallback chains across providers, see our LLM API Fallback Routing Guide.

Migration from GPT-5/5.5

If you're on GPT-5 or GPT-5.5, migrating to GPT-5.6 is mostly a model string swap:

Note: GPT-5 models are on OpenAI's deprecation schedule. Plan your migration now — see our GPT-5 Deprecation Migration Guide for timelines and gotchas.

API compatibility

The API surface is the same as GPT-5.5: Chat Completions with function calling, structured outputs, streaming, and reasoning support. The model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna.

One notable change: prompt caching now charges for cache writes (1.25× input rate). If your workload relies heavily on free cache writes from GPT-5.5, factor this into your cost projections.

GPT-5.6 vs Competitors

How does Sol compare to Claude Opus 4.8?

Sol ($5/$30) and Opus 4.8 ($5/$25) are priced similarly, but Sol leads on TerminalBench 2.1 by ~10 points (88.8% vs 78.9%). Opus 4.8's strengths lie in extended thinking for nuanced analysis and its 200K context window (expandable to 1M with premium context). For pure agentic coding benchmarks, Sol currently leads.

How does Terra compare to Claude Sonnet 4.5?

Terra at $2.50/$15.00 targets the same "balanced everyday model" slot as Claude Sonnet 4.5 ($3/$15). Terra matches GPT-5.5 on benchmarks at half the cost. If you're currently on Sonnet 4.5 and cost-sensitive, Terra is worth testing.

How does Luna compare to Gemini 2.5 Flash?

Luna ($1/$6) competes with Gemini 2.5 Flash ($0.15/$0.60 for thinking, free for non-thinking under limits). Gemini Flash is cheaper per token, but Luna posts higher TerminalBench scores (82.5% vs Gemini 3.1 Pro Preview at 70.7%). The trade-off is cost vs. capability — Luna is more expensive but measurably stronger on agentic tasks.

Access and Availability

GPT-5.6 launched as a limited preview on June 26, 2026. The US government requested that OpenAI restrict the initial rollout, and OpenAI complied — while publicly stating this should not become the norm for frontier model releases.

What this means in practice:

• Access is staged by account, region, and product surface
• ChatGPT and Codex users may see the models before broad API availability
• Do not assume production-grade SLAs or stable rate limits yet
• OpenAI says they plan to make all three tiers "generally available in the coming weeks"

Build behind an abstraction. If your access to GPT-5.6 changes during the preview window, you want to fall back to GPT-5.5 or another provider without code changes. For fallback patterns, see our Fallback Routing Guide.

TheRouter Integration

Caveat: GPT-5.6 Sol, Terra, and Luna are not yet in TheRouter's model catalog. The models were announced June 26 and are in limited preview. We expect to add them once they reach general API availability and we can verify routing behavior.

When GPT-5.6 models are added, the routing pattern looks like any other OpenAI-compatible model:

from openai import OpenAI

# Route through TheRouter
client = OpenAI(
    api_key="your-therouter-key",
    base_url="https://api.therouter.ai/v1"
)

response = client.chat.completions.create(
    model="openai/gpt-5.6-terra",  # expected model ID
    messages=[{"role": "user", "content": "Hello"}]
)

For existing GPT-5.5 routing through TheRouter, see our model catalog.

Worked Cost Examples

To make the tier decision concrete, here's what a typical daily workload costs across all three tiers:

Scenario: 2 million tokens/day, 70% input / 30% output

Cascaded routing: 80% Luna / 15% Terra / 5% Sol

That's a 71% saving vs. routing everything through Sol and only a 43% increase over pure Luna — while getting Sol-quality answers on the hardest 5% of requests.

Related Reading

• OpenAI GPT-5 Deprecation Migration Guide
• Best Reasoning Model APIs 2026
• OpenAI Prompt Caching Cost Optimization Guide
• LLM API Fallback Routing Guide
• OpenAI to TheRouter Migration Guide
• Anthropic Claude API Complete Guide

FAQ

Which GPT-5.6 tier should I use by default?

Terra. It matches GPT-5.5's capability at roughly half the cost, making it the natural default for most application traffic. Only escalate to Sol for tasks that genuinely need frontier-grade reasoning — long-horizon coding agents, security analysis, or research workflows.

Is Sol Ultra a separate model?

No. Sol Ultra is a high-effort mode of Sol, not a separate model with its own pricing. It uses the same per-token rates as Sol but consumes more tokens per request due to deeper reasoning. Think of it as Sol with the compute dial turned to maximum.

Can I use GPT-5.6 through TheRouter right now?

Not yet. GPT-5.6 Sol, Terra, and Luna are in limited preview as of June 2026. We expect to add them to TheRouter's model catalog once they reach general API availability. In the meantime, you can route to GPT-5.5 and GPT-5.5 Pro through TheRouter.

How does GPT-5.6 pricing compare to Claude?

Sol ($5/$30) is comparable to Claude Opus 4.8 ($5/$25) on input but slightly higher on output. Terra ($2.50/$15) undercuts Claude Sonnet 4.5 ($3/$15) on input. Luna ($1/$6) is more expensive than Claude Haiku 3.5 ($0.80/$4) but benchmarks higher on agentic tasks.

Should I migrate from GPT-5.5 now or wait?

Wait until GPT-5.6 exits limited preview and reaches general availability. Once it does, migrating from GPT-5.5 to Terra is a straightforward model string swap with expected cost savings of ~50%. There is no urgency — GPT-5.5 is not on the immediate deprecation schedule.

Does prompt caching work differently on GPT-5.6?

Yes. GPT-5.6 introduces cache write billing at 1.25× the input rate (GPT-5.5 cache writes were free). Cache reads still get a 90% discount. The 30-minute minimum cache life and explicit cache breakpoints make caching more predictable — but you now pay a small premium to populate the cache.