When OpenAI launched GPT-5.6, it also published comparisons between Sol, Terra and Luna and GPT-5.5, Claude Fable 5, Claude Opus 4.8 and Gemini 3.1 Pro Preview. The table is useful for understanding broad positioning, but scores from different organizations and test conditions should not be treated as one universal ranking.
The short version: GPT-5.6 Sol targets the hardest coding, knowledge-work and agent tasks; Terra is the better everyday default; and Luna focuses on speed and cost. GPT-5.5 remains a reliable previous-generation baseline. Claude Fable 5 leads some professional evaluations, while Claude Opus 4.8 and Gemini 3.1 Pro Preview are competitive in particular workloads.
Professional evaluations
| Evaluation | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 | Claude Fable 5 | Claude Opus 4.8 | Gemini 3.1 Pro Preview |
|---|---|---|---|---|---|---|---|
| Agents’ Last Exam | 52.7% | 50.4% | 50.3% | 46.9% | 40.5% | 45.2% | 32.1% |
| GDPval-AA v2 | 1,747.8 Elo | 1,593 Elo | 1,591.8 Elo | 1,493.7 Elo | 1,759.6 Elo | 1,600.1 Elo | 962.3 Elo |
| Internal management consulting tasks | 43.2% | 37.2% | 35.4% | 31.3% | 35.5% | 31.6% | 13.2% |
| Big Finance Bench | 53% | 51% | 36% | 49% | — | 44% | — |
| Artificial Analysis Intelligence Index v4.1 | 58.9 | 55 | 51.2 | 54.8 | 59.9 | 55.7 | 46.5 |
Sol, Terra and Luna all score above GPT-5.5 on Agents’ Last Exam. Claude Fable 5 leads GDPval-AA v2, with Sol close behind. Fable 5 also has a slightly higher overall Artificial Analysis Intelligence Index score than Sol, but that does not mean it is better for every workload.
Coding and terminal tasks
| Evaluation | GPT-5.6 Sol | GPT-5.6 Sol Ultra | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 | Claude Fable 5 | Claude Opus 4.8 | Gemini 3.1 Pro Preview |
|---|---|---|---|---|---|---|---|---|
| Artificial Analysis Coding Agent Index | 80 | — | 74.6 | 76.4 | — | 77.2 | 72.5 | 42.7 |
| SWE-Bench Pro | 64.6% | — | 63.4% | 62.7% | 59.4% | 80.3% | 77.8% | 54.2% |
| DeepSWE v1.1 | 72.7% | — | 69.6% | 67.2% | 67% | — | 69.7% | 11.8% |
| Terminal-Bench 2.1 | 88.8% | 91.9% | 87.4% | 84.7% | 85.6% | 88% | 83.1% | 70.7% |
Sol is in the top tier on the Coding Agent Index, DeepSWE and Terminal-Bench, while ultra raises its Terminal-Bench score to 91.9%. Claude Fable 5 and Claude Opus 4.8 score higher on SWE-Bench Pro. Coding quality is not one-dimensional: repository understanding, tool use, test repair, token cost and long-running stability also matter.
Science and health tasks
| Evaluation | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 | Claude Opus 4.8 | Gemini 3.1 Pro Preview |
|---|---|---|---|---|---|---|
| GeneBench Pro | 28.7% | 23.3% | 10.8% | 12% | 16% | 3.1% |
| LifeSciBench | 59.9% | 56% | 51.2% | 50.4% | 53.6% | — |
| MedChemBench (internal) | 48.3% | 35% | 30.4% | 35.5% | — | — |
| HealthBench Professional | 60.5% | 57.7% | 55.7% | 49.5% | 53% | — |
Sol is clearly ahead of GPT-5.5 on these evaluations, with Terra and Luna also improving over the previous generation. Science and health are high-stakes domains, however; scores cannot replace expert review, source checking and safety procedures.
Computer use and browsing
| Evaluation | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 | Claude Fable 5 | Claude Opus 4.8 | Gemini 3.1 Pro Preview |
|---|---|---|---|---|---|---|---|
| OSWorld 2.0 | 62.6% | 50.2% | 45.6% | 47.5% | — | 54.8% | — |
| BrowseComp | 90.4% | 87.5% | 83.3% | 84.4% | 88% | 84.3% | 85.9% |
| BenchCAD | 70.6% | 62.3% | 63.1% | 44.4% | 38.4% | 35.5% | 27.3% |
| BenchCAD (Python tool) | 83.4% | 78.2% | 73.9% | 55.8% | 65% | 51.8% | — |
Sol posts strong results on OSWorld 2.0, BrowseComp and BenchCAD, while Terra and Luna also beat GPT-5.5. For browsing, desktop automation, chart generation and frontend work, computer-use success can matter more than a pure text benchmark.
Cybersecurity evaluations
| Evaluation | GPT-5.6 Sol | GPT-5.6 Sol Ultra | GPT-5.6 Terra | GPT-5.6 Luna | GPT-5.5 |
|---|---|---|---|---|---|
| Capture-the-Flag Challenges | 96.7% | — | 91.8% | 85.2% | 88.1% |
| SEC-Bench Pro | 71.2% | 74.3% | 57.7% | 48.9% | 45.8% |
| CyberGym | 84.5% | — | 81.8% | 77.9% | 81.8% |
| ExploitBench | 73.5% | — | 52.9% | 33.2% | 47.9% |
| ExploitGym | 33.7% | — | 23.2% | 12.4% | 15.1% |
GPT-5.6 improves substantially on these cybersecurity evaluations. The capabilities are dual-use, so OpenAI combines Trusted Access, real-time checks and account-level controls for higher-risk work. Use them only in authorized environments and prioritize defensive auditing, patch validation and blue-team testing.
Price and performance must be considered together
GPT-5.6 API pricing is $5 input/$30 output for Sol, $2.50/$15 for Terra and $1/$6 for Luna per million tokens. GPT-5.5, Claude and Gemini have different pricing, cache discounts, context limits and product access. A benchmark table alone cannot determine procurement.
Build a small evaluation set from your own codebase, documents, tool calls and structured outputs. Track quality, latency, input and output tokens, cache hit rate and human rework time.
Which model should you choose?
- Complex agents, cross-file coding, professional research and high-value work: start with GPT-5.6 Sol.
- Everyday office work, normal coding and most default requests: start with GPT-5.6 Terra.
- Classification, summarization, extraction and high-throughput processing: test GPT-5.6 Luna.
- If GPT-5.5 is stable in production, run an A/B test before migrating.
- For code repair, compare Claude Fable 5, Claude Opus 4.8 and GPT-5.6 on your own repository rather than relying on one benchmark.
- For browsing, desktop use or frontend generation, focus on OSWorld, BrowseComp, BenchCAD and real interaction success.
GPT-5.6 is not about replacing every model with Sol. Its three tiers cover different budgets and difficulty levels. Cross-model evaluations narrow the options; your own data, tools and cost model should make the final decision.
Source: OpenAI GPT-5.6 official release.