GPT-5.6 Capabilities Explained

Generally available as of July 2026: GPT-5.6 (Sol, Terra, and Luna) is generally available in the API and Codex — no waitlist and no vetted-org gating (OpenAI).

GPT-5.6 is OpenAI's newest model family, announced June 26, 2026 (OpenAI). It ships as three tiers built for different jobs: Sol for the hardest problems, Terra for high-volume business tasks, and Luna for fast, low-cost everyday work.

This guide explains GPT-5.6 in plain English. You will learn what each tier does, what it costs, how to get access today, the best business uses, the limits to plan around, and how it compares to Claude Fable 5.

What Is GPT-5.6?

GPT-5.6 is OpenAI's newest model family, offered in three tiers named Sol, Terra, and Luna (OpenAI). OpenAI announced it on June 26, 2026.

Each tier targets a different job. You pick the tier that matches the task, rather than using one model for everything.

GPT-5.6 is generally available today. Any organization can use it through the API and Codex — no waitlist and no vetting (OpenAI).

• Sol — the hardest problems, such as complex coding and security research. It adds new "max" and "ultra" reasoning modes, and "ultra" uses subagents to speed up complex work.
• Terra — high-volume business tasks, such as customer support, internal tools, and document analysis.
• Luna — fast, low-cost everyday work, such as summarization, drafting, and routine automation.
• Delivery — available via the OpenAI API and Codex.
• Status — generally available in the API and Codex as of July 2026 (OpenAI).

Key Features and Capabilities

GPT-5.6's main feature is the three-tier split, so you can trade off power against cost. Sol handles the hardest work, Terra handles scale, and Luna handles speed.

Sol adds two new reasoning modes, called "max" and "ultra" (OpenAI). The "ultra" mode uses subagents to speed up complex work.

Terra and Luna aim at everyday business volume. Terra fits support and document analysis, while Luna fits quick drafting and routine automation.

• Sol reasoning — new "max" and "ultra" modes for complex coding and security research.
• Ultra subagents — "ultra" mode uses subagents to speed up complex work (OpenAI).
• Terra scale — built for customer support, internal tools, and document analysis.
• Luna speed — built for summarization, drafting, and routine automation at low cost.
• Access surfaces — the API and Codex.

New in GPT-5.6 (What Changed at GA)

General availability brought more than open access. GPT-5.6 adds five API features that let it handle more of a job end to end, instead of handing steps back to your code (OpenAI).

The theme is coordination. GPT-5.6 can now run its own glue logic, split work across helper agents, think harder when quality matters, carry reasoning between turns, and reuse cached context to keep repeat calls cheap.

For most teams the takeaway is simple: these upgrades let GPT-5.6 take on more tool-heavy, multi-turn work with less custom plumbing on your side.

• Programmatic Tool Calling — GPT-5.6 runs lightweight JavaScript in a hosted runtime to coordinate tools, pass results between calls, and cut intermediate steps.
• Multi-agent (beta) — it can spawn concurrent subagents inside a single request to split work, explore approaches in parallel, and synthesize the result.
• Pro mode + max reasoning effort — Pro mode does more model work before returning one final answer for quality-first tasks; max effort gives it more room to explore and verify.
• Persisted reasoning — it carries reasoning forward across turns, improving multi-turn quality and cache efficiency when the goal stays stable.
• Prompt cache breakpoints — you mark the reusable prompt prefixes you want cached so repeated context runs more cheaply.

Pricing and Access

GPT-5.6 pricing is set per million tokens and rises with each tier's power (OpenAI). Luna is the cheapest, Terra sits in the middle, and Sol is the most expensive.

Access is open today. GPT-5.6 is generally available via the API and Codex, with no waitlist or vetting (OpenAI).

It reached general availability in July 2026 (OpenAI). Any business can buy it directly through the API.

• Sol — $5 per million input tokens, $30 per million output tokens (OpenAI).
• Terra — $2.50 per million input tokens, $15 per million output tokens (OpenAI).
• Luna — $1 per million input tokens, $6 per million output tokens (OpenAI).
• Access — generally available via the API and Codex (OpenAI).
• Timeline — generally available as of July 2026 (OpenAI).

Best Uses for Business

The best use of GPT-5.6 depends on your role and the task in front of you. Match the tier to the job to control both quality and cost.

Engineering and security teams get the most from Sol. Support and operations teams fit Terra, while high-volume, low-stakes work fits Luna.

One non-obvious tradeoff: routing everything to Sol wastes money on tasks Luna handles well. A tier router that picks the cheapest tier that clears a quality bar is usually the real win.

• Engineering leads — use Sol for complex coding and hard debugging, with "ultra" mode on the toughest tickets.
• Security teams — use Sol for security research and deep code review.
• Customer support — use Terra for ticket triage, drafted replies, and knowledge-base answers.
• Operations and legal ops — use Terra for document analysis and internal tools over large volumes.
• Marketing and admin — use Luna for summarization, first-draft copy, and routine automation.
• Cost control — route each task to the cheapest tier that meets your quality bar.

Limitations and Risk Factors

Access is no longer a limit — GPT-5.6 is generally available via the API and Codex (OpenAI). The real limits now are cost control and matching the right tier to each task.

Because the tiers differ widely in price and speed, sending every task to Sol wastes money. Matching the tier to the job is the main discipline to get right.

A common failure mode is over-reliance on Sol's "ultra" mode. Subagent runs use more tokens and time, so reserve them for work that truly needs them.

• Tier sprawl — three tiers mean you must route each task to the right one or overspend.
• Model choice — picking Sol for simple work wastes budget that Luna or Terra would cover.
• Cost creep — Sol and "ultra" mode use more tokens, which raises the bill fast.
• Compliance diligence — confirm the specific model is named in your BAA before sending regulated data (see below).
• Human review — all three tiers can make mistakes, so keep a person in the loop on high-stakes output.

Compliance and Data Handling

GPT-5.6 inherits OpenAI's platform compliance coverage once a model is in scope. That includes SOC 2, ISO 27001, and a HIPAA BAA available on the API and Enterprise (OpenAI).

Coverage is not automatic for every model. Confirm the specific model is named in your BAA before you send regulated data (OpenAI).

For regulated fields — healthcare, legal, finance — start with a scoped pilot. Keep protected data out until the model is confirmed in your agreement.

• Platform coverage — SOC 2 and ISO 27001 apply once a model is in scope (OpenAI).
• HIPAA — a BAA is available on the API and Enterprise plans (OpenAI).
• BAA check — confirm GPT-5.6 is named in your BAA before sending regulated data (OpenAI).
• Safe default — run non-regulated pilots first, then expand once BAA coverage is confirmed.

How GPT-5.6 Compares

GPT-5.6's main alternative for most businesses is Claude Fable 5, which is generally available worldwide as of July 1, 2026 (Anthropic). GPT-5.6 is now generally available too, so the choice comes down to price, tiers, and data handling rather than access.

On price, Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens (Anthropic). Every GPT-5.6 tier is cheaper per token, and you can buy GPT-5.6 openly through the API today.

The practical choice today is simple: both are deployable now, so weigh Claude Fable 5's single public model against GPT-5.6's tier-level cost control across Sol, Terra, and Luna.

• Availability — both are generally available: Claude Fable 5 worldwide, GPT-5.6 via the API and Codex (Anthropic; OpenAI).
• Tiers — GPT-5.6 offers three tiers; Claude Fable 5 is a single public model (OpenAI; Anthropic).
• Price — GPT-5.6 tiers run $1–$5 input and $6–$30 output; Claude Fable 5 is $10 input and $50 output (OpenAI; Anthropic).
• Data handling — Claude Fable 5 carries a mandatory 30-day safety retention, so full zero-data-retention is not available on it (Anthropic).
• When to choose GPT-5.6 — you want to route cheap tasks to Luna and hard tasks to Sol for tier-level cost control.
• When to choose Claude Fable 5 — you need a capable model you can deploy today across your business.

How to Access GPT-5.6

You access GPT-5.6 today through the OpenAI API and Codex (OpenAI). It is generally available, so any organization can sign up — no vetting and no waitlist.

Getting started is straightforward: set up API access, then design your tier routing and confirm compliance needs (OpenAI).

A short pilot helps. Test Luna for volume work, Terra for support and documents, and Sol for the hardest coding before you commit spend.

• Get access — sign up through the OpenAI API; it is generally available with no vetting (OpenAI).
• Use the right surface — access runs through the OpenAI API and Codex (OpenAI).
• Generally available — GPT-5.6 is live in the API and Codex as of July 2026 (OpenAI).
• Confirm compliance — verify the model is named in your BAA before regulated data (OpenAI).
• Pilot by tier — map each workflow to Sol, Terra, or Luna before scaling.

Getting Started with GPT-5.6 (And Is It Free?)

No, GPT-5.6 is not free — there is no free tier, and it is a paid model, generally available via the API and Codex (OpenAI). You pay per token on each tier, so cost scales with how much you use.

Getting started is a short, ordered process. You set up API access, pick the right tier for each task, then pilot before you scale.

Design your tier routing before you spend. Route cheap, high-volume work to Luna, support and documents to Terra, and only the hardest problems to Sol.

• Step 1 — Set up access. GPT-5.6 is generally available, so sign up through the OpenAI API — no vetting or waitlist (OpenAI).
• Step 2 — Choose your surface. Access runs through the OpenAI API and Codex (OpenAI).
• Step 3 — Map tasks to tiers. Send everyday work to Luna, business volume to Terra, and hard coding or security work to Sol.
• Step 4 — Run a small pilot. Test one workflow per tier and measure quality and token cost before scaling.
• Step 5 — Confirm compliance. Verify the model is named in your BAA before you send any regulated data (OpenAI).

Recent Updates on GPT-5.6 Capabilities and Model Variants

GPT-5.6 capabilities have recently expanded, as OpenAI made three model variants—Sol, Terra, and Luna—generally available, each designed to balance performance and cost while strengthening cybersecurity safeguards.

According to OpenAI's June 26, 2026 announcement, Sol acts as the flagship GPT-5.6 model, with Terra and Luna introduced as alternatives optimized for different workloads and pricing structures. All three focus on enhanced cyber capabilities, a multi-layer safeguard system, and increased reliability through automated adversarial testing. This layered stack approach aims to identify and neutralize potential risks proactively. Notably, the review and testing protocols now emphasize iterative red-teaming, which OpenAI describes as using automated systems to simulate adversarial misuse and evaluate the model’s ability to withstand sophisticated attacks. This marks a shift from GPT-5.5, where red-teaming relied more heavily on manual workflows and static defense layers. In practice, these features address emergent risks in both user data handling and model output, reflecting OpenAI's ongoing updates to align with evolving security demands.

• Three model variants released: Sol (flagship), Terra, and Luna.
• Focus on enhanced cyber capabilities and robust automated safeguards.
• Layered safeguard stack is designed to identify and neutralize risks.
• Automated red-teaming now supplements manual security assessments.
• The models address emergent data handling and model output risks.