For reference: Claude Fable 5 charges $10/$50.

The bare gpt-5.6 alias routes to Sol. These are not code-name placeholders — OpenAI has committed to "durable capability tiers that advance on their own cadence."

This is the most important AI model launch of 2026, and the benchmark data reveals a more nuanced picture than any headline captures. Here is the complete technical breakdown every engineering team needs.

1. The Three-Tier Architecture: What Changed

The Old Way: One Model, One Price Point

GPT-5.5 shipped as a single model with a single price. Teams either paid the premium or didn't use OpenAI. Model selection was binary: use GPT-5.5 or fall back to GPT-4o.

The New Way: Tiered by Reasoning Effort

GPT-5.6 introduces explicit effort levels configurable per request — from none to max, plus a new ultra mode that spawns parallel sub-agents. The tiers are not just price-discrimination; they represent different reasoning-depth defaults baked into the architecture.

# New API shape — effort is now a first-class parameter
response = client.responses.create(
    model="gpt-5.6-sol",
    input="Implement a distributed rate limiter in Go",
    reasoning={"mode": "pro", "effort": "high"}
)

# Ultra spawns 4 parallel agents by default
response = client.responses.create(
    model="gpt-5.6-sol",
    input="...",
    reasoning={"mode": "ultra"}  # 4x parallelization
)

The Naming Philosophy

• Sol (sun) — the brightest, most capable tier
• Terra (earth) — grounded, balanced, everyday
• Luna (moon) — reflects Sol's capabilities at a fraction of the cost

2. Benchmark Analysis: Where Each Tier Wins and Loses

2.1 Agents' Last Exam (Long-Horizon Agentic Work)

Bar: 55 professional fields, multi-hour workflows requiring planning, tool use, and adaptation.

Key insight: Sol beats Fable 5 by 13.1 points — a 32% improvement. But the within-family gap (Sol to Luna) is only 3.3 points. Luna costs 1/5 of Sol. For agentic workflows, Terra delivers 94% of Sol's performance at half the price.

Recommendation: For long-horizon agents, start with Terra. Escalate to Sol only when you hit the quality ceiling.

2.2 Coding Benchmarks: Where the Frontier Splits

Artificial Analysis Coding Agent Index

Sol leads by 2.8 points over Fable 5, using <50% of the output tokens and roughly 1/3 less cost. Terra matches Fable 5 at half the price.

Terminal-Bench 2.1 (Terminal-First Development)

Sol Ultra costs ~3x single-agent Sol ($5 vs $1.70 per run) for an extra +3.1 points. For most teams, single-agent Sol at 88.8% is the sweet spot.

SWE-Bench Pro (Repo-Level Coding)

This is the critical divergence. OpenAI questions SWE-Bench Pro's validity (~30% of tasks have broken tests), but the gap is large enough that it cannot be dismissed entirely. Fable 5 remains the leader for repo-level code generation — complex multi-file changes, refactoring, and legacy code integration. GPT-5.6 Sol excels at agentic coding — terminal interaction, tool coordination, multi-step builds.

The takeaway: These models are not interchangeable. The frontier has split. Choose your model by task type, not by overall rank.

2.3 The Luna Long-Context Cliff (MRCR)

This is the most important finding in the entire GPT-5.6 benchmark suite.

Luna drops 50 percentage points on the Multi-step Reasoning with Context Retrieval benchmark. This is not a graceful degradation — it is a cliff.

What this means:

• Luna is unsuitable for document analysis, large-codebase reasoning, multi-document synthesis
• Luna cannot reliably follow long instructions or maintain context across extended conversations
• If your workflow involves context windows >32K tokens, do not use Luna

This is the hidden cost of Luna's 80% price reduction. The tier is optimized for short-context, high-frequency tasks: classification, extraction, structured output, simple completion.

2.4 Knowledge Work & Computer Use (Sol's Cleanest Lead)

• BrowseComp: Sol at 92.2% (new SOTA) — browsing-based research and synthesis
• OSWorld 2.0: Sol at 62.6% — surpasses Opus 4.8 using 85% fewer output tokens
• Design: Sol can infer layout, typography, spacing, and slide master rules from reference decks

These benchmarks where Sol dominates are precisely the workflows that benefit from deep reasoning + tool use — the architecture sweet spot.

2.5 Cybersecurity (ExploitBench)

GPT-5.6 spent 12 days in government-mandated pre-launch review under Trump's AI Cybersecurity Executive Order. The ExploitBench score shows a 53% improvement over GPT-5.5, but Mythos 5 (Anthropic's unrestricted variant) still leads.

OpenAI launched Trusted Access for Cyber — verified individuals get access to defensive capability (vulnerability triage, patch validation). Hardware-backed passkeys required by September 1, 2026.

3. Cost-Performance: The Luna Disruption

DeepSWE v1.1 (Cost-Efficiency Leaderboard):

Luna delivers 5.3x more benchmark points per dollar than Opus 4.8 and 7.5x more than Fable 5. For high-volume, short-context workloads, Luna is the most cost-efficient model ever shipped by OpenAI.

But cost-efficiency is not cost-effectiveness if the model cannot do the task. The MRCR cliff means Luna's cost advantage only applies within its narrow capability envelope.

Practical Cost Scenarios

Assumes 1K input + 2K output tokens per request average.

4. Ultra Mode: Multi-Agent Inference Comes to the API

Ultra mode is not a model — it is an inference-time multi-agent orchestration layer. By default, it spawns 4 parallel sub-agents, each working on sub-tasks, coordinated by a meta-agent.

Performance gains vs single-agent Sol:

• BrowseComp: +1.8 points
• SEC-Bench Pro: +3.1 points
• Terminal-Bench 2.1: +3.1 points

Cost: ~3x single-agent Sol.

When to use Ultra:

• Tasks that decompose naturally into independent sub-tasks
• Browsing/research (where parallel page visits help)
• Security analysis (multiple attack surface vectors investigated simultaneously)
• Code review (parallel analysis of multiple files)

When NOT to use Ultra:

• Tasks requiring tight sequential reasoning (one step's output determines the next step's input)
• Short requests (the orchestration overhead dominates)
• Budget-constrained workloads

OpenAI also released a multi-agent beta in the Responses API, letting developers build custom agent orchestration instead of using the default 4-agent Ultra topology.

5. What OpenAI Did Not Report

The omitted benchmarks tell a story of their own:

OpenAI now reports performance as curves across reasoning-effort levels, not single scores. The charitable read: single-shot academic benchmarks don't capture real-world model usage. The cynical read: OpenAI leads on agentic benchmarks and trails on academic ones. Both can be true.

6. Engineer's Tier Selection Guide

Use Luna When:

• ✦ Classification, extraction, summarization (context < 32K tokens)
• ✦ High-volume structured output generation
• ✦ Simple RAG retrieval (single-document grounding)
• ✦ Cost-constrained batch processing
• ✦ ✦ NEVER use for: long-context reasoning, multi-document analysis, code generation for complex repos

Use Terra When:

• ✦ Everyday coding and debugging
• ✦ Single-file code generation and review
• ✦ Agentic workflows (you get 94% of Sol's performance at 50% of the cost)
• ✦ Medium-complexity RAG and tool use
• ✦ General-purpose replacement for GPT-5.5 / Claude Opus

Use Sol When:

• ✦ Long-document analysis (>64K tokens)
• ✦ Complex multi-file code generation
• ✦ Browsing-based research (BrowseComp SOTA)
• ✦ Computer-use tasks (OSWorld SOTA)
• ✦ Cybersecurity analysis
• ✦ Any task where the marginal cost of getting it right exceeds the marginal cost of Sol's 2x premium over Terra

Use Sol Ultra When:

• ✦ Research that benefits from parallel exploration
• ✦ Multi-perspective code review
• ✦ Security assessments
• ✦ Deadline-constrained complex tasks where cost is secondary

7. The Bigger Picture: What GPT-5.6 Means for Engineering Teams

The Reasoning Spectrum

OpenAI's effort-level architecture (none → low → medium → high → max → ultra) represents a fundamental shift in how we think about model selection. Instead of choosing a model, teams will increasingly choose a reasoning budget — and the API dynamically allocates compute within that budget.

This is the direction the industry is heading: continuous capability scaling rather than discrete model upgrades. The same model architecture serves both a $1/MTok Luna request and a $15/MTok Sol Ultra request.

The Platform Play: Programmatic Tool Calling

GPT-5.6 can write and run lightweight programs that coordinate tools, process intermediate results, and adapt workflows — available in the Responses API with Zero Data Retention (ZDR) support. This is OpenAI's answer to Anthropic's tool-use architecture: not just calling tools, but writing programs that orchestrate tools.

The Gatekeeping Precedent

GPT-5.6 is the first flagship model to undergo a formal government pre-launch review — 13 days from restricted preview to general availability. This sets a precedent. Future model launches may face similar delays, and engineering teams should build model-agnostic orchestration layers that can switch providers on short notice.

Summary Decision Matrix

GPT-5.6 is the most capable model family OpenAI has ever shipped — but capability without routing intelligence is wasted budget. The winning engineering teams in 2026 will be those that build smart tier-routing layers, not those that default to the most expensive model for every task.