Numbers first. The benchmark data OpenAI published on July 9, 2026 tells a clearer story than the marketing copy: Sol leads on computer use and complex browsing by a wide margin, while Terra and Luna cluster surprisingly close behind it on most other evaluations. Luna, the cheapest tier, clears GPT-5.5 on Agents' Last Exam, HealthBench Professional, and DeepSWE.
What follows is every major benchmark category from the official GPT-5.6 launch data — including the one evaluation where GPT-5.6 still trails Claude significantly.
How to Read GPT-5.6 Benchmark Results
Before diving in, a few framing notes:
- Three tiers, one generation. Sol is the flagship; Terra is balanced (GPT-5.5-competitive); Luna is the fastest and most affordable. All three run the same core architecture but at different capability and cost levels.
- Two optional reasoning modes.
maxextends reasoning time;ultracoordinates four parallel agents by default. Where benchmarks show these modes, it is noted. - Cost and latency are plotted alongside score. OpenAI's benchmark charts show score vs. API cost, making efficiency — not just raw score — the primary comparison axis.
Agentic and Professional Work
Agents' Last Exam
Agents' Last Exam evaluates long-running professional workflows across 55 fields — one of the most demanding real-world agent evaluations available.
| Model | Score |
|---|---|
| GPT-5.6 Sol | 52.7% |
| GPT-5.6 Terra | 50.4% |
| GPT-5.6 Luna | 50.3% |
| GPT-5.5 | 46.9% |
| Claude Fable 5 | 40.5% |
| Claude Opus 4.8 | 45.2% |
| Gemini 3.1 Pro Preview | 32.1% |
All three GPT-5.6 tiers exceed GPT-5.5 and Claude Fable 5. The gap between Sol and Luna is narrow (2.4 points), which means Luna delivers near-Sol performance on professional agentic tasks at one-fifth the output cost.
Artificial Analysis Intelligence Index v4.1
A broad aggregate intelligence measure covering agentic work, coding, scientific reasoning, and general capabilities.
| Model | Index Score |
|---|---|
| GPT-5.6 Sol | 58.9 |
| Claude Fable 5 | 59.9 |
| Claude Opus 4.8 | 55.7 |
| GPT-5.6 Terra | 55.0 |
| GPT-5.5 | 54.8 |
| GPT-5.6 Luna | 51.2 |
| Gemini 3.1 Pro Preview | 46.5 |
Sol comes within one point of Claude Fable 5 on this index, while completing tasks in 61% less time at roughly half the estimated cost, according to OpenAI's efficiency analysis.
Coding and Software Engineering
GPT-5.6 makes its most pronounced gains in coding agent tasks. OpenAI describes Sol as its best coding model to date.
Artificial Analysis Coding Agent Index v1.1
| Model | Index Score |
|---|---|
| GPT-5.6 Sol | 80 |
| Claude Fable 5 | 77.2 |
| GPT-5.6 Terra | 77.4 |
| GPT-5.5 | 76.4 |
| GPT-5.6 Luna | 74.6 |
| Claude Opus 4.8 | 72.5 |
| Gemini 3.1 Pro Preview | 42.7 |
Sol sets a new state of the art here — above Claude Fable 5 — while using less than half the output tokens. Terra also edges Fable 5 by 0.2 points.
DeepSWE v1.1 (Long-horizon engineering in real codebases)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 72.7% |
| GPT-5.6 Terra | 69.6% |
| GPT-5.6 Luna | 67.2% |
| GPT-5.5 | 67.0% |
| Claude Fable 5 | 69.7% |
| Claude Opus 4.8 | 59.0% |
Terminal-Bench 2.1 (Complex command-line workflows)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 88.8% |
| GPT-5.6 Sol Ultra | 91.9% |
| GPT-5.6 Terra | 87.4% |
| GPT-5.6 Luna | 84.7% |
| GPT-5.5 | 85.6% |
| Claude Fable 5 | 83.1% |
| Claude Opus 4.8 | 78.9% |
Ultra mode pushes Sol's Terminal-Bench score to 91.9% — the highest recorded in this evaluation.
SWE-Bench Pro
| Model | Score |
|---|---|
| Claude Mythos 5 | 80.3% |
| Claude Fable 5 | 80.0% |
| GPT-5.6 Sol | 64.6% |
| GPT-5.6 Terra | 63.4% |
| GPT-5.6 Luna | 62.7% |
| GPT-5.5 | 59.4% |
SWE-Bench Pro is the one benchmark where GPT-5.6 trails Claude models significantly. This evaluation focuses on resolving real GitHub issues in their original environment — a task profile where Claude Mythos 5 retains an advantage.
Computer Use and Browsing
BrowseComp (Agentic browsing tasks)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 90.4% |
| GPT-5.6 Sol Ultra | 92.2% |
| Claude Mythos 5 | 88.0% |
| GPT-5.6 Terra | 87.5% |
| GPT-5.6 Luna | 83.3% |
| GPT-5.5 | 84.4% |
| Claude Opus 4.8 | 84.3% |
Sol sets a new state of the art on BrowseComp. Ultra mode pushes this to 92.2%.
OSWorld 2.0 (Computer use)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 62.6% |
| Claude Opus 4.8 | 54.8% |
| GPT-5.6 Terra | 50.2% |
| GPT-5.5 | 47.5% |
| GPT-5.6 Luna | 45.6% |
Sol's 62.6% on OSWorld 2.0 surpasses Claude Opus 4.8 (54.8%) while using 85% fewer output tokens.
Science and Health
GeneBench Pro (Genomics and quantitative biology)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 28.7% |
| GPT-5.6 Terra | 23.3% |
| Claude Opus 4.8 | 16.0% |
| GPT-5.5 | 12.0% |
| GPT-5.6 Luna | 10.8% |
Note: Claude Fable 5 is not included in this benchmark — Anthropic has confirmed it does not answer advanced biology questions.
HealthBench Professional
| Model | Score |
|---|---|
| Claude Fable 5 | 60.9% |
| GPT-5.6 Sol | 60.5% |
| GPT-5.6 Terra | 57.7% |
| GPT-5.6 Luna | 55.7% |
| GPT-5.5 | 49.5% |
| Claude Opus 4.8 | 53.0% |
Long Context
GPT-5.6 handles long-context recall with a clear tier split — Sol and Terra maintain strong performance at 1M tokens; Luna drops substantially.
OpenAI MRCR v2 — 8 needle, 512K–1M context
| Model | Score |
|---|---|
| GPT-5.6 Sol | 73.8% |
| GPT-5.6 Terra | 72.5% |
| GPT-5.5 | 74.0% |
| GPT-5.6 Luna | 41.3% |
Luna is not the right choice for tasks requiring reliable recall across very long contexts.
Academic Reasoning
GPQA Diamond (Graduate-level science questions)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 94.6% |
| Claude Mythos Preview | 94.6% |
| Claude Fable 5 | 92.6% |
| GPT-5.5 | 93.6% |
| GPT-5.6 Terra | 92.9% |
| GPT-5.6 Luna | 92.3% |
The entire GPT-5.6 family scores above 92% on GPQA Diamond. The gap between Sol and Luna here is only 2.3 points.
Abstract Reasoning
ARC-AGI-3
| Model | Score |
|---|---|
| GPT-5.6 Sol | 7.78% |
| Claude Opus 4.8 | 1.5% |
| GPT-5.5 | 0.43% |
| GPT-5.6 Terra | 0.80% |
| GPT-5.6 Luna | 0.18% |
ARC-AGI-3 remains an extremely hard benchmark for all frontier models. Sol's 7.78% is a significant jump from GPT-5.5's 0.43%, but absolute scores remain low across the board.
What the Benchmarks Tell You About Model Selection
A few patterns emerge from reading the full benchmark set together:
1. Luna is more capable than its price suggests. On most agentic and professional benchmarks, Luna outperforms GPT-5.5 despite costing four times less per million tokens. For workloads that previously ran on GPT-5.5, Luna deserves a real evaluation before defaulting to Terra.
2. Terra is the practical production tier. It edges Claude Fable 5 on coding agent benchmarks, exceeds GPT-5.5 across the board, and does so at half the cost. For sustained production workloads, Terra is the default choice.
3. Sol's advantage concentrates in computer use and complex browsing. OSWorld 2.0 (62.6% vs 50.2% for Terra) and BrowseComp are where Sol's lead is most pronounced. If your agent interacts with real GUI environments or needs reliable complex web research, Sol's edge matters.
4. SWE-Bench Pro is the exception. Claude Mythos 5 and Fable 5 retain a substantial lead here. For agents whose primary task is resolving GitHub issues in production codebases, that benchmark should weigh heavily in model selection.
5. Luna fails at long context. The drop from 72.5% (Terra) to 41.3% (Luna) on 512K–1M MRCR tasks is significant. If your agent processes long documents or maintains long conversation histories, Luna is the wrong choice.
Building an Agent Stack Around GPT-5.6
Benchmarks measure reasoning. But reasoning is only half of what a production agent needs. Most real workflows also require the agent to create media, search the web, inspect files, or publish output — capabilities that are outside any language model's native scope.
AnyCap provides that capability layer: a single CLI that adds image generation, video production, web search, web crawl, persistent storage, and hosted page publishing to any AI agent. Your agent chooses the GPT-5.6 tier that fits the reasoning task; AnyCap handles the rest.
The two complement each other cleanly. A Terra-powered research agent can call AnyCap to pull current data from the web, generate charts, and publish a finished report — within a single agent session, without rebuilding infrastructure for each capability.
All benchmark data sourced from OpenAI's official GPT-5.6 announcement, published July 9, 2026 at openai.com. Scores reflect standard evaluation conditions unless otherwise noted.