A.I.
Frontier models
GPT-5.6 Sol
Flagship for complex professional work; pre-Astra default for many teams.
Training
Knowledge cutoff: February 16, 2026. Parameter count, FLOPs, and train-time RL recipe for Sol: UNKNOWN.
Inference
reasoning.effort includes none through max (broader than Astra, which has no none). Reasoning tokens are supported.
Context
1.05M context window. 128K max output tokens.
Methods
Functions, web search, file search, and computer use are named on the models page. Exact RL recipe: UNKNOWN.
Benchmarking
OpenAI-published Sol numbers from the GPT-6 Astra announce (OpenAI's Sol column / prose). Do not merge with Anthropic's Sol column from the Fable page; harnesses differ.
| Eval | GPT-5.6 Sol (OpenAI) | GPT-6 Astra | Notes |
|---|---|---|---|
| Terminal-Bench 4.0 | 37.3% | 57.9% | |
| Terminal-Bench Science 0.1 | 22.4% | 64.6% | |
| OSWorld 2.0 (latency sim prose) | 65.7% @ ~75 min/task | 72.6% @ ~40 min | Astra ~47% less time |
| AutomationBench | 18.1% | 41.4% | |
| BrowseComp | 90.4% | 91.5% | |
| BenchCAD | 83.3% | 95.9% | |
| DeepSWE v1.1 | 72.7% | 74.1% | |
| FrontierMath Tier 4 (v2) | 83.0% | 97.6% | |
| GPQA Diamond | 94.6% | 96.0% | |
| Humanity's Last Exam (w/ tools) | - | 57.2% | Sol not listed for this row |
Health (system card Table 6, length-adjusted, Sol as baseline): HealthBench Professional 60.5; HealthBench 57.0; HealthBench Hard 33.1.
Footnote: Anthropic also printed GPT-5.6 Sol columns on the Fable 5.1 announce for some evals (e.g. Terminal-Bench 4.0 37.3%, AutomationBench 19.6%, CursorBench 3.2.0 67.2%). Those are Anthropic's harness results, not merged into the table above.
Hardware
Per-SKU hardware for Sol: UNKNOWN. OpenAI's published Stargate Abilene / NVIDIA GB200 naming applies to GPT-5.5 training, not a Sol-specific map. Buyers purchase API tokens.
Sources
- GPT-5.6 Sol model page
- GPT-5.6 Luna model page (Luna cached input $0.02 confirmed here, 4 Sep 2026)
- Models
- GPT-6 Astra announce (Sol comparison columns)
- GPT-6 Astra System Card (HealthBench Sol baselines)
- Building the compute infrastructure for the intelligence age (company infra context only)
Sibling: GPT-5.6 Terra
Balance of intelligence and cost. Id: gpt-5.6-terra. Pricing: $2 / $12 per million tokens. Context / output: 1.05M / 128K. Cutoff: February 16, 2026. Same tool family; reasoning.effort none…max. Benchmarking and per-SKU hardware: UNKNOWN.
Sibling: GPT-5.6 Luna
Cost-sensitive, high-volume (roughly the nano tier of earlier GPT-5 families). Id: gpt-5.6-luna. Pricing: $0.20 / $1.20 per million tokens; cached input $0.02. Context / output: 1.05M / 128K. Cutoff: February 16, 2026. reasoning.effort supports none, low, medium (default), high, xhigh, and max (confirmed on the Luna model page, 4 Sep 2026). Benchmarking and per-SKU hardware: UNKNOWN.