What is the difference between GPT-5.6 Sol, Terra, and Luna?

GPT-5.6 Sol is the flagship reasoning model — highest capability across coding, research, cybersecurity, and complex analysis. GPT-5.6 Terra is the cost-efficient middle tier — strong general-purpose performance at roughly one-third of Sol's price. GPT-5.6 Luna is the speed and volume tier — optimized for low latency, high concurrency, and consumer-scale traffic. All three are now GA on Amazon Bedrock as of July 2026. Choose Sol for hard reasoning, Terra for most enterprise use, Luna for latency-sensitive or high-volume workloads.

How much does each model cost on Amazon Bedrock?

Pricing tracks OpenAI's direct API with a small Bedrock delivery premium. GPT-5.6 Sol runs approximately $10/$40 per million input/output tokens. GPT-5.6 Terra runs approximately $3/$12 per million tokens. GPT-5.6 Luna runs approximately $0.60/$2.40 per million tokens. On-demand pricing on Bedrock also supports batch (50% discount) and provisioned throughput (reserved capacity) for high-volume workloads. Actual pricing verified July 2026.

Why use OpenAI models on Amazon Bedrock instead of OpenAI directly?

Three main reasons. First, unified AWS billing and IAM — same account, same VPC, same logging as your other AWS services. Second, cross-model routing on Bedrock — you can call Claude Sonnet 5, Nova, Llama derivatives, Mistral, and now GPT-5.6 through one API. Third, data residency and compliance — Bedrock keeps data within your AWS region and never trains on your traffic. The tradeoff: Sora video and ChatGPT Work agent features stay OpenAI-direct only.

When did OpenAI models arrive on Amazon Bedrock?

The deal was announced in April 2026 when OpenAI ended its Microsoft cloud exclusivity. GPT-5.5 arrived on Bedrock in May 2026. GPT-5.6 Sol, Terra, and Luna reached General Availability on Amazon Bedrock in July 2026, following the July 9, 2026 GPT-5.6 Sol launch. Amazon committed up to $50 billion in cloud infrastructure to OpenAI as part of the partnership. A Stateful Runtime Environment for OpenAI models on Bedrock is slated for H2 2026.

Quick Answer

GPT-5.6 Sol vs Terra vs Luna on Amazon Bedrock: Which to Pick (2026)

Published: July 15, 2026

GPT-5.6 Sol vs Terra vs Luna on Amazon Bedrock: Which to Pick (July 2026)

Amazon Bedrock now offers all three GPT-5.6 tiers — Sol (flagship), Terra (cost-efficient), and Luna (speed) — as of the July 2026 general-availability announcement. For AWS customers who were waiting for OpenAI models on Bedrock, the picker is simple in the abstract and tricky in practice. Here is the full decision guide, with pricing, benchmarks, and workload fit for each tier.

Last verified: July 15, 2026

The Three Tiers at a Glance

	GPT-5.6 Sol	GPT-5.6 Terra	GPT-5.6 Luna
Position	Flagship reasoning	Cost-efficient mid-tier	Speed / volume
Best for	Hard reasoning, coding, cybersecurity, deep research	80% of enterprise use cases	Latency-sensitive, consumer-scale
SWE-bench Verified	~73%	~62%	~48%
MMLU-Pro	~88%	~79%	~66%
Context window	400K in / 128K out	200K in / 64K out	128K in / 32K out
Latency (first token)	~1.2-2.5s	~500-900ms	~150-350ms
Bedrock input/output cost (per MTok)	~$10 / $40	~$3 / $12	~$0.60 / $2.40
Reasoning modes	Full chain-of-thought, agentic loops	Chain-of-thought optional	Fast-only
Multimodal	Text + image + audio in; text out	Text + image in; text out	Text in; text out

Numbers verified against AWS Bedrock July 2026 GA announcement, OpenAI’s own GPT-5.6 documentation, and Reuters / Bloomberg pricing coverage.

When to Pick GPT-5.6 Sol

Use Sol when the workload is limited by reasoning quality, not throughput:

Complex coding agents running long autonomous loops (Cursor + Sol Auto, Codex Sol)
Cybersecurity vulnerability analysis — including submitting to GOLD EAGLE-style workflows
Legal, finance, and medical research where a wrong answer is expensive
Multi-step scientific analysis with tool use
Anything where the human reviewer’s time costs more than the token spend

Sol pricing (~$10/$40 per MTok on Bedrock) is roughly 15× more expensive than Luna and 3× more expensive than Terra. For a workload that needs one right answer, that math is trivial. For batch consumer workloads, it is prohibitive.

Real workloads Sol makes sense for:

SWE-bench-style code repair over medium-to-large codebases
Legal document analysis with strict citation accuracy requirements
Financial deep-research pipelines (equity, credit)
Cybersecurity code review (see GOLD EAGLE launch, July 14, 2026)
Agentic workflows requiring 5-20 tool calls with correct planning

When to Pick GPT-5.6 Terra

Terra is the “default” tier for most enterprise workloads. It is where roughly 60-70% of AWS Bedrock GPT-5.6 traffic lands based on typical enterprise usage patterns:

Customer support agents — support ticket triage, response generation, escalation logic
Enterprise RAG applications — retrieval + generation on internal docs
Marketing content generation at scale
Data extraction from unstructured docs (contracts, invoices, forms)
Internal analytics summaries
Standard code assist for developers (autocomplete, docstrings, basic refactors)

Terra’s ~~$3/$12 per MTok pricing is close to Anthropic Claude Sonnet 5’s Bedrock pricing (~~$3/$15). For enterprise buyers, the choice between Terra and Sonnet 5 usually comes down to model style, tool-use behavior, and existing prompt libraries rather than price.

Terra is not a good fit for:

Cybersecurity analysis (use Sol)
Speed-sensitive consumer chat (use Luna)
Multimodal video or audio out (neither Terra nor any GPT-5.6 tier produces video; use OpenAI direct for Sora)

When to Pick GPT-5.6 Luna

Luna is optimized for speed and unit cost at consumer scale:

Consumer chatbot backends where P95 first-token latency under 300ms is a product requirement
Autocomplete-style suggestions in IDEs, docs, and productivity tools
High-volume classification (spam, toxicity, intent classification)
Content moderation queues processing millions of items per day
Voice-adjacent workflows where end-to-end latency budget is tight
Streaming search summaries and chat-with-your-data experiences

Luna’s ~$0.60/$2.40 per MTok pricing is competitive with Anthropic Claude Haiku, Gemini 3.5 Flash, and Amazon’s own Nova Pro. If you are doing millions of low-complexity requests per day, Luna is where the economics work.

Luna trades off:

Meaningfully worse on complex reasoning (SWE-bench ~48% vs. Sol ~73%)
No agentic multi-step planning capability equivalent to Sol
Text-only output (no image or audio generation)
Shorter effective context (~128K in / 32K out)

A Decision Tree

Is your workload primarily long-form reasoning, code, or high-stakes analysis? → Sol
Is your workload latency-sensitive (sub-300ms first token) or high-volume (>1M requests/day)? → Luna
Otherwise → Terra (this is where 60-70% of enterprise workloads end up)

Real-World Routing Pattern

Sophisticated Bedrock users route based on request complexity:

Incoming request →
  Complexity classifier (Luna itself, ~10ms) →
    If simple / high-volume → Luna
    If moderate → Terra
    If complex reasoning / agentic → Sol
  → Return

This is the “AI router pattern” popularized in the coding-tools space (Cursor Auto, Continue.dev, RouteLLM) applied at the Bedrock-invocation layer. A typical enterprise sees 70% Luna, 25% Terra, 5% Sol traffic when routing intelligently, at roughly 25% of the cost of running everything on Sol.

GPT-5.6 on Bedrock vs. OpenAI Direct

You would use Bedrock when:

You are already an AWS shop and want unified IAM + VPC + billing
Data residency + no-training-on-your-data guarantees matter
You want to A/B test GPT-5.6 vs. Claude Sonnet 5 vs. Nova through one API
Batch pricing (50% discount) matters
Provisioned throughput / reserved capacity fits your steady-state

You would go OpenAI direct when:

You need Sora video generation (Bedrock does not host)
You need ChatGPT Work agent features (Bedrock does not host)
You need the newest features on day zero (OpenAI direct always ships first)
You want function-calling behaviors that match OpenAI’s official reference exactly

What Is Missing from Bedrock

As of July 15, 2026, Bedrock hosts GPT-5.6 Sol, Terra, and Luna text generation but does not yet host:

Sora / Sora 2 Pro video generation
ChatGPT Work agent
GPT-5.6 realtime / voice
Codex Sol cloud coding agent (though the base Sol model can be used with your own agent scaffold)

AWS said in the July 2026 GA announcement that a Stateful Runtime Environment for OpenAI models is slated for H2 2026, which is expected to unlock agent-style workloads.

The Frame

For most AWS shops, the answer is: use Terra by default, route to Sol for hard reasoning, route to Luna for volume. That gets you 90% of the value at 30-40% of the cost of running everything on Sol.

If you are picking between GPT-5.6 (any tier) and Claude Sonnet 5 on Bedrock, the answer is workload-dependent — Claude Sonnet 5 leads on SWE-bench and long-horizon coding agents; GPT-5.6 Sol leads on general reasoning and multimodal. Terra and Sonnet 5 are close enough that most enterprises test both on their actual workload.

Sources

AWS: OpenAI GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock — July 2026
About Amazon: Bedrock adds OpenAI models — 2026
TechRadar: AWS and OpenAI confirm cloud partnership — April 2026
OpenAI: GPT-5.6 release notes — July 2026