GPT-5.6

Overview

GPT-5.6 is OpenAI's flagship model family, generally available since July 9, 2026. Unlike previous releases it ships as three named tiers rather than one model with size variants: Sol (frontier flagship), Terra (balanced everyday model), and Luna (fastest and cheapest).

The unusual part is that the tiers are not distinguished by context or capability ceiling. All three carry the same 1,050,000 token context window, the same 128,000 token output cap, and the same February 16, 2026 knowledge cutoff. What separates them is speed, depth of reasoning, and price — a 5x spread from Luna to Sol. OpenAI's own developer guidance treats tier selection as a runtime routing decision rather than an architectural one: start on Luna, escalate to Terra for multi-step work, and reserve Sol for tasks where failure is expensive.

These are reasoning models in the large language model sense: they spend internal reasoning tokens before answering, and depth is controlled by a reasoning_effort parameter that now ranges from none through max.

GPT-5.6 does not retire the older families. GPT-5.5, GPT-5.5 Pro, GPT-5.4 and GPT-5.4 Pro all remain listed and callable. Note that GPT-5.5 is priced identically to Sol at $5.00 / $30.00 while carrying a cutoff eleven weeks earlier — the migration is a model-ID change with no cost penalty.

Capabilities

• Long-horizon agentic work: The stated focus of the release. OpenAI reports Sol setting a new high on Agents' Last Exam, a benchmark for long-running professional workflows.
• Programmatic Tool Calling: The model writes JavaScript that runs in an isolated V8 runtime with no network access, letting it orchestrate several tools and filter intermediate results before anything reaches the context window. This cuts both round trips and context bloat.
• Multi-agent mode (beta): An ultra setting runs four agents concurrently on independent subtasks. It is experimental and needs its own synthesis logic downstream.
• Configurable reasoning depth: Effort from none through max, traded per request against latency and token spend.
• Long-context ingestion: 1,050,000 tokens on every tier — large codebases and document sets in a single request.
• Vision input: Images alongside text.
• Expanded hosted tools: Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search.

Technical Specifications

• API model IDs: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna
• Alias: gpt-5.6 routes to Sol — pin an explicit tier in production
• Context window: 1,050,000 tokens (all tiers)
• Max output: 128,000 tokens (reasoning tokens count toward this)
• Knowledge cutoff: February 16, 2026
• Input modalities: Text, images
• Output modalities: Text
• Audio / video: Not supported
• Reasoning effort: none through max
• APIs: Chat Completions, Responses, and Realtime
• Features: Streaming, function calling, structured outputs, Programmatic Tool Calling. Fine-tuning is not supported.
• Rate limits (Sol, Tier 5): 15,000 requests per minute, 40,000,000 tokens per minute

The GPT-5 Family Today

The Pro tier has no GPT-5.6 equivalent yet: gpt-5.5-pro remains the high-compute option, and it is the one model in the lineup with no cached-input row on OpenAI's pricing page — every token is billed at full rate.

Use Cases

• Long-running agentic coding: Repository-scale work where Sol's reasoning depth and the expanded shell/apply-patch tooling matter more than per-token cost.
• Tiered agent architectures: Luna as the subagent doing extraction and classification, Terra for multi-step orchestration, Sol reserved for the steps where a wrong answer is expensive. This is the deployment shape OpenAI's own guidance recommends.
• High-volume classification and extraction: Luna at $1.00 per million input tokens, with the full 1M context still available.
• Whole-codebase analysis: Ingesting a large repository and its history in a single 1M+ token prompt.
• Professional knowledge work: Research synthesis, financial and legal analysis, long-form document review.
• Tool-orchestrating agents: Programmatic Tool Calling for workflows that would otherwise burn context shuttling intermediate JSON back and forth.
• Document and screenshot understanding: Image input for charts, scans, and UI captures.

Performance / Benchmarks

OpenAI-reported figures for Sol at the July 2026 release:

Agents' Last Exam — a UC Berkeley RDI benchmark of 1,500+ tasks from real professional work across 55 industries, graded by deterministic scripts rather than an AI judge:

OpenAI's prose claims "a new high of 53.6" and a 13.1-point lead, which its own table contradicts — only 53.6 reconciles with 40.5 plus 13.1. Berkeley RDI attributes the 53.6 to Codex running Sol at high reasoning effort, a different configuration from the bare-model row above.

Other reported figures for Sol:

The SWE-Bench Pro result runs the other way. Claude Fable 5 scores 80.3% against Sol's 64.6% — a 16-point deficit on the most-cited agentic coding benchmark. OpenAI's response was to dispute the benchmark, estimating that roughly 30% of its tasks are broken. That may well be true, but a vendor challenging the scoring of a benchmark it lost is an argument, not a result. Sol also trails Claude models on tool-use evaluations.

All figures above are vendor-reported or drawn from launch coverage; independent evaluations may differ.

Safety & Preparedness

Under OpenAI's Preparedness Framework, all three GPT-5.6 models are treated as:

• Biological and Chemical — High capability. Below the Critical threshold.
• Cybersecurity — High capability. Below Critical.
• AI Self-Improvement — Below High capability.

The June 26, 2026 preview was deliberately narrow: OpenAI opened it only to a small group of trusted partners, and shared the list of participants with the government, before the July 9 general release.

The cyber safeguard stack includes live restrictions on risky cyber assistance, threat-intelligence-driven investigation and detection, controls on scaled agentic vulnerability research through the Trusted Access for Cyber program, and user confirmations for high-risk computer actions.

Limitations

• Loses to Claude on SWE-Bench Pro by 16 points, and on tool-use evaluations. OpenAI contests the benchmark rather than the gap.
• Text output only. No audio or video generation, and no audio or video input.
• Long-prompt price cliff. Prompts above 272K input tokens are billed at a premium — a 1M-token prompt costs materially more per token than a 200K one.
• Reasoning tokens are billed as output. High effort can consume a large share of the 128K output budget before any visible answer appears.
• max effort has no cost ceiling beyond the output cap; a single request can spend the full 128K budget on reasoning.
• Multi-agent ultra mode is beta and only pays off when subtasks are genuinely independent.
• Cache writes now cost extra — 1.25x the uncached input rate — so caching a prefix used only once is a net loss.
• No Pro tier. The highest-compute option is still gpt-5.5-pro, on the older cutoff.
• No fine-tuning on any GPT-5.6 tier.
• Knowledge cutoff. Nothing after February 16, 2026 is in the weights; use web search or retrieval for current information.
• Regional processing surcharge. Routing requests through regional processing adds a 10% uplift.

Pricing & Access

API pricing (per 1M tokens)

Modifiers that apply to GPT-5.6:

• Cached reads: 90% off the input rate.
• Cache writes: 1.25x the uncached input rate — new in this release, and a reason to place stable instructions before request-specific data.
• Long prompts: more than 272K input tokens are billed at a premium rate.
• Regional processing: adds a 10% uplift for models released after March 5, 2026.

Availability

ChatGPT (Plus, Pro, Business, and Enterprise), ChatGPT Work, Codex, and the API, with tier availability varying by platform. OpenAI is also serving Sol on Cerebras hardware at up to 750 tokens per second.

Deprecations to watch

On June 11, 2026 OpenAI announced a December 11, 2026 shutdown for a batch of older models:

An earlier wave shuts down sooner: the -chat-latest and -codex aliases across GPT-5, 5.1, and 5.2 retire on July 23, 2026, as do the o3-deep-research and o4-mini-deep-research snapshots. The gpt-5.2-chat-latest and gpt-5.3-chat-latest aliases follow on August 10, 2026.

Ecosystem & Tools

Official SDKs

• Python SDK
• Node SDK
• REST API via the Responses, Chat Completions, and Realtime endpoints

Hosted tools

Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, tool search, and Model Context Protocol (MCP) servers are available without extra plumbing.

Platforms

• OpenAI Developer Platform — models, pricing, and API reference
• Platform documentation
• OpenAI Cookbook — worked examples

Community & Resources

• GPT-5.6 announcement - OpenAI's general-availability post
• GPT-5.6 Sol model reference - Context, pricing, modalities, rate limits
• GPT-5.6 Terra model reference
• GPT-5.6 Luna model reference
• GPT-5.6 preview deployment safety page - System card for Sol, Terra, and Luna
• GPT-5.5 deployment safety page - The previous flagship
• API pricing
• Model deprecations
• Developer forum