Back to all posts

GPT-5.6 Explained: Sol vs Terra vs Luna, Pricing, and How It Compares to Claude Sonnet 5

What you will learn in this article:

What OpenAI actually shipped on July 9, and why there are three models instead of one.

How the new naming system works, and why "Sol", "Terra", and "Luna" are tiers, not versions.

What each model costs and which real workloads each tier is designed for.

How the family lines up against Claude Sonnet 5, which launched just ten days earlier.

What the unusual government-preview rollout signals about where frontier releases are heading.

Introduction

On July 9, 2026, OpenAI publicly launched the GPT-5.6 series: three models named Sol, Terra, and Luna, available in both the ChatGPT apps and the API. It's the biggest OpenAI release since GPT-5.5, and it lands in a crowded ten-day window. Anthropic shipped Claude Sonnet 5 at the end of June, so every team running LLMs in production is now re-evaluating the same question: which model, at which price, for which job?

This post walks through what's actually new, what each tier costs, what the day-one benchmark claims say (and how much weight to put on them), and how the family compares to Claude Sonnet 5 for practical work.

The New Naming System: Generations and Tiers Are Now Separate

The change most likely to outlast this release isn't a benchmark score. It's the naming scheme. With GPT-5.6, OpenAI split model identity into two axes:

The number (5.6) identifies the model generation.
The name (Sol, Terra, Luna) identifies a durable capability tier that can advance on its own cadence.

So "Sol" is intended to keep meaning "the flagship tier" across future generations, the way "Opus / Sonnet / Haiku" work for Anthropic's Claude family or "Pro / Flash" work for Google's Gemini. Each tier can now be updated independently, without the awkward version arithmetic that produced names like "GPT-4o-mini" and "o4-mini-high" in earlier eras. If you've ever had to explain to a stakeholder why "mini" was newer than the full-size model it outperformed, this change is for you.

Sol vs Terra vs Luna: What Each Tier Is For

Model	Position	Designed for	API price (per 1M tokens, input / output)
GPT-5.6 Sol	Flagship	The hardest problems: complex coding, security research, long agentic tasks	$5 / $30
GPT-5.6 Terra	Balanced	High-volume business work: customer support, internal tools, document analysis	$2.50 / $15
GPT-5.6 Luna	Fast and affordable	Everyday tasks: summarization, drafting, routine automation	$1 / $6

The pattern will look familiar if you've read our Claude coverage: a flagship for depth, a mid-tier for volume, and a small model for speed. The interesting part is the pricing pressure in the middle. Early reporting describes Terra as delivering performance competitive with GPT-5.5 at roughly half the cost, and the mid-tier is where most production token volume actually lives.

In ChatGPT, reporting on the rollout indicates paid plans (Plus and above) get Sol, free users get Terra, and Pro/Enterprise plans also get a Sol Pro variant plus a new "ultra" setting that coordinates multiple agents (four by default) on a single task. Exact plan mappings may shift during the rollout, so check OpenAI's help center for the current matrix before relying on any specific tier being available.

Day-One Benchmark Claims, and How to Read Them

Two headline numbers are circulating from launch coverage:

OpenAI reports that Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark for agentic terminal/coding work.
On the Artificial Analysis Coding Agent Index, launch reporting places Sol at 80.0 (2.8 points above Claude Fable 5), while using less than half the output tokens and costing roughly a third less on the measured tasks.

Treat both with day-one discipline. Vendor-reported and launch-week numbers give you a rough direction, but they routinely shift once independent evaluations land, and benchmark deltas of a few points rarely survive contact with your specific workload. The only benchmark that matters for your system is your own eval set. For a continuously updated cross-vendor view, our AI model leaderboard refreshes weekly with benchmark scores across open and proprietary models.

The Safety Story: Cyber Thresholds and a Government Preview

Two safety-related details in this launch are new territory, and they echo something we saw with Claude Sonnet 5's release.

First, all three models crossed OpenAI's "High" cybersecurity capability threshold on internal capture-the-flag testing: Sol at 96.7%, Terra at 91.84%, and Luna at 85.19%. OpenAI says Sol ships with its strongest safety stack to date, with tighter protections around higher-risk activity, sensitive cyber requests, and repeated misuse. If your application touches security topics, expect more refusals at this tier. We flagged the same operational reality for Claude Sonnet 5's cybersecurity safeguards.

Second, the rollout itself was staged in an unusual way. Before the public launch, OpenAI shared the models and release plans with the US government and ran a limited preview with roughly 20 organizations. This lands in the middle of reported talks between the White House and OpenAI, Google, and Anthropic on voluntary standards for frontier model releases (benchmarks, testing timelines, and access rules). Whether or not those talks formalize, GPT-5.6 shows what frontier launches may look like from now on: notify the government, run a controlled preview, then open the doors.

GPT-5.6 vs Claude Sonnet 5: The Practical Comparison

Claude Sonnet 5 launched on June 30; GPT-5.6 followed nine days later. The natural head-to-head is Terra vs Sonnet 5, since both are the "workhorse" tier where production volume concentrates.

	GPT-5.6 Terra	Claude Sonnet 5
Price (input / output per 1M tokens)	$2.50 / $15	$3 / $15
Positioning	High-volume business tasks	Default model for Free and Pro users; production workhorse
Context window	Not yet confirmed at time of writing	1M tokens (default and maximum)
Reasoning behavior	Effort settings (medium and higher on paid ChatGPT plans)	Adaptive thinking on by default; effort parameter
Notable migration gotcha	New tier naming; plan-to-model mapping still settling	New tokenizer produces ~30% more tokens for the same text

On list price alone, Terra undercuts Sonnet 5 by 50 cents per million input tokens at the same output price. But as we covered in the Sonnet 5 post, list price is not effective price: Sonnet 5's new tokenizer changes how many tokens the same text consumes, and OpenAI and Anthropic tokenizers have never counted text identically. The only honest cost comparison is running your real prompts through both and comparing the bill, not the rate card.

At the flagship level, Sol ($5 / $30) enters well below what flagship pricing looked like a year ago, continuing the industry-wide trend of frontier capability getting cheaper per token even as absolute spend rises with usage.

Which Should You Use?

If you are on GPT-5.5 today: Terra is the obvious first test, reported as GPT-5.5-competitive at roughly half the cost. Run your eval set before switching production traffic; "competitive" is a launch-week adjective, not a guarantee about your workload.
If you run agentic coding workloads: Sol's Terminal-Bench claim and the multi-agent "ultra" setting are aimed directly at you. Budget for $30/M output tokens. Agent loops are output-heavy, and that's where this tier gets expensive.
If you are cost-driven at scale: Luna at $1 / $6 competes in the small-model bracket where latency and price dominate: classification, extraction, summarization, the boring 80% of production LLM work.
If you are happy on Claude Sonnet 5: nothing here forces a move. The 50-cent input-price gap is within tokenizer-difference noise for many workloads, and switching costs (prompt tuning, eval reruns, refusal behavior differences) are real. Re-evaluate when independent benchmarks stabilize.

For what it's worth, nothing in this launch changes this site's own stack either: the chatbot here runs on Gemini with Pinecone retrieval, and a new rate card is a reason to re-run evals, not a reason to migrate during launch week.

FAQ

Is GPT-5.6 available right now?

Yes. The public rollout began July 9, 2026 across the ChatGPT apps and the API, following a limited preview period. OpenAI described full availability as arriving within roughly 24 hours of launch, so access should be broadly live by the time you read this.

Do free ChatGPT users get GPT-5.6?

Launch reporting indicates free users get Terra, while Sol is reserved for paid plans (with a Sol Pro variant and the multi-agent "ultra" setting on higher tiers). Plan mappings can shift during rollouts, so the OpenAI help center has the authoritative current matrix.

What happened to GPT-5.5?

OpenAI has not framed GPT-5.6 as an immediate removal of GPT-5.5. The more meaningful change is structural: with generation numbers and capability tiers now separated, future upgrades should arrive as tier updates rather than whole-family renames. Deprecation timelines for older models will follow OpenAI's usual API lifecycle announcements.

Are the benchmark numbers trustworthy?

They are real reported numbers, but they are launch-week numbers: vendor-run or early third-party measurements on specific task sets. History says the ordering between frontier models on any given benchmark tightens or flips as independent evaluations accumulate. Use them to decide what to test, not what to deploy.

Why did OpenAI show the models to the US government first?

OpenAI shared the models and release plans with the US government and ran a ~20-organization preview before public launch, against the backdrop of reported White House talks with OpenAI, Google, and Anthropic on voluntary frontier-release standards. It's the clearest signal yet that staged, government-notified releases are how frontier models will ship from now on.

References

Key Takeaways

GPT-5.6 launched publicly on July 9, 2026 as a three-tier family: Sol (flagship, $5/$30), Terra (balanced, $2.50/$15), and Luna (fast, $1/$6) per million tokens.
The naming system now separates generation (the number) from capability tier (the name). Sol, Terra, and Luna are meant to persist across future generations, each advancing on its own cadence.
Terra is the head-to-head rival to Claude Sonnet 5: same $15 output price, slightly cheaper input. But tokenizer differences mean the rate card alone cannot tell you which is cheaper for your workload.
All three models crossed OpenAI's "High" cyber capability threshold, and the release was staged through a US-government-notified preview of ~20 organizations, a template for how frontier launches may work going forward.
Day-one benchmark claims (Terminal-Bench 2.1 SOTA for Sol; 80.0 on the Artificial Analysis Coding Agent Index) are worth testing against, not deploying on. Run your own evals; check our weekly-updated leaderboard for the cross-vendor picture.

Quiz

Question 01

What does the new GPT-5.6 naming system signify, according to OpenAI?

B is correct. The post explains OpenAI's stated design: the generation number (5.6) and the tier names (Sol, Terra, Luna) are now separate axes, so a tier can be updated on its own schedule without renaming the whole family.

Question 02

Which GPT-5.6 model is priced closest to Claude Sonnet 5, and how do their list prices compare?

C is correct. The post's comparison table shows Terra ($2.50 / $15) sitting directly against Claude Sonnet 5 ($3 / $15): identical output price, with Terra slightly cheaper on input.

Question 03

What was unusual about how GPT-5.6 was released, according to the post?

C is correct. The post notes the staged rollout: a narrow preview group of about 20 organizations and disclosure of release plans to the US government preceded the July 9 public launch, context that connects to ongoing talks about voluntary frontier-model standards.