GPT-5.6 Sol, Terra, Luna pricing decoded — and why Sol Ultra should worry you

GPT-5.6 is OpenAI's three-tier frontier model family — Sol, Terra, and Luna — currently in limited preview as of July 2026, priced from $1 to $5 per million input tokens depending on capability tier. It replaces the single-SKU model selection developers used with GPT-5.5: Sol targets frontier agentic reasoning, Terra is the new production workhorse, and Luna handles high-volume cost-sensitive workloads — and one tier has no published price at all.

Key Facts

1. Sol: $5 input / $30 output per million tokens — flagship tier for frontier reasoning, complex coding, and long-horizon agentic work [1][2]
2. Terra: $2.50 input / $15 output per million tokens — balanced everyday model with GPT-5.5-competitive performance at 2× lower cost [2][6]
3. Luna: $1 input / $6 output per million tokens — fastest and most affordable tier for high-throughput, cost-sensitive use cases [2][6]
4. Cache-write billing change: Cache writes are now billed at 1.25× the model's uncached input rate — previously free. Cache reads keep the existing 90% discount [2]
5. Sol Ultra: Exists as an API parameter (ultra mode, activating subagent orchestration). No published price [1]
6. Competitive context: Claude Fable 5 runs at $10/$50 per million tokens (input/output) — 2× Sol's input cost [4]. Gemini 3.1 Pro is priced in a comparable range to Terra on input tokens — verify current figures at Google Cloud's pricing documentation before committing to cross-provider cost models.
7. Access status: GPT-5.6 is in limited preview, available only to approved API and Codex partners. No public waitlist, no ChatGPT access, no GA date announced [2][6]

The Tier Selection Framework

OpenAI is being unusually explicit about routing with this release. The preview documentation describes Sol as designed for "frontier reasoning and long-horizon agentic-workflows," Terra as "a balanced everyday model with GPT-5.5-competitive performance at 2× lower cost," and Luna as "the fastest, most affordable member of the family" [1][6]. That is a routing signal encoded in the product definition — not just marketing.

The practical rule: default to Terra, escalate to Sol only for tasks where an incorrect answer has measurable downstream cost, and use Luna where you can cheaply verify the output. Do not route anything to Sol Ultra yet.

The Cache-Write Billing Change: What It Actually Costs You

This is the most underreported change in GPT-5.6. Cache writes are now billed at 1.25× the model's uncached input rate. Previously, writing to the prompt-caching layer was free — you only paid on reads, which still receive a 90% discount [2].

Concrete example: an application using Terra ($2.50/MTok input) with a 10,000-token system prompt cached on every cold start:

• Before GPT-5.6: $0 for the cache write, then $0.025 per cached read (10k tokens × $2.50/MTok × 10% of uncached rate)
• After GPT-5.6: $0.03125 per cache write (10k × $2.50/MTok × 1.25 / 1,000,000)

If your application cold-starts 1,000 times per day, that is $31.25/day in new cache-write costs — roughly $940/month at Terra rates, and $2,500/month at Sol rates. For teams who built their GPT-5.5 cost models assuming cache writes were free, this is a material budget change.

The mitigation: extend cache lifetime. GPT-5.6 introduces a 30-minute minimum cache life and explicit cache breakpoints [2], so a single write can be amortised across more reads before expiry. Audit your cold-start frequency and session length before migrating.

Competitive Routing: Where GPT-5.6 Fits Against Claude Fable 5 and Gemini 3.1 Pro

For teams already on OpenAI's API, Terra is the competitive answer to Gemini 3.1 Pro — similar price band, no provider switch required. For teams choosing between Sol and Claude Fable 5, capability tradeoffs will matter more than the 2× price gap once independent benchmarks appear. Until then, Sol is the lower-cost bet at comparable capability claims.

The Sol Ultra Problem

Sol's preview announcement describes an ultra mode that uses subagents to accelerate complex, multi-step work [1]. It surfaces as an API parameter in the Sol spec. It has no published price.

This is a planning risk for any team sizing infrastructure costs now. If your pipeline will eventually require Sol Ultra for long-horizon agentic-workflows, you are designing around a cost anchor that OpenAI has not published. OpenAI is reserving the right to set the price for the model tier that sits above the leaderboard — and teams who commit infrastructure design to Sol Ultra before that price lands have no floor to budget against.

The conservative design posture: scope your budgets to Sol ($5/$30) and treat Sol Ultra as out-of-scope until pricing is announced. Do not architect workflows that require it for correctness unless you can absorb an unknown multiplier.

Try It: Route a Task to the Right GPT-5.6 Tier

What Changes If You Are Already Using GPT-5.5

Three checks before migrating:

1. Cache-write billing: Audit workflows that cache frequently — the 1.25× write rate is a new cost GPT-5.5 budgets did not include.
2. Access gating: Confirm preview access before blocking roadmap work on a migration that has not been granted.
3. Sol Ultra unknowns: Hold off on cost commitments for top-tier agentic workloads until OpenAI publishes Sol Ultra pricing.

Knowledge Check

The Bottom Line

GPT-5.6's three-tier structure makes model routing explicit for the first time at OpenAI's frontier tier. Terra is the default move for most production workloads — similar price to Gemini 3.1 Pro, no provider switch required. Sol is justified where an incorrect answer has measurable downstream cost. Luna is the right call for high-volume, verifiable tasks.

The cache-write billing change is the sleeper cost: audit cold-start cache patterns before migrating. Treat Sol Ultra as a planning unknown until pricing appears.

Want a structured framework for picking the right frontier model? The Koenig AI Academy's picking-a-frontier-model-2026-q2 course covers the full decision matrix, including cost modelling for agentic pipelines.