GPT‑5.6 Sol vs GPT‑5.6 Terra
Terra costs exactly half as much per token. Sol is the frontier tier. The practical question is not which model scores higher—it does—but when Sol’s margin is worth paying 2× for.
Last updated July 10, 2026 · Official OpenAI data + independent context
GPT‑5.6 Sol
GPT‑5.6 Terra
Exactly half the listed token price
Sol and Terra share the same context length, maximum output, API endpoints, image-input support, tool support, cache lifetime, and pricing mechanics. The difference is capability tier and token rate: Sol is precisely 2× Terra on fresh input, cached input, cache writes, and output.
| API specification | GPT‑5.6 Sol | GPT‑5.6 Terra |
|---|---|---|
| Input / 1M tokens | $5.00 | $2.50 |
| Output / 1M tokens | $30.00 | $15.00 |
| Cached input / 1M | $0.50 | $0.25 |
| Cache write / 1M | $6.25 | $3.125 |
| Context / max output | 1.05M / 128K tokens | 1.05M / 128K tokens |
| Long-context requests | Over 272K input: 2× input, 1.5× output for full request | Same policy |
| Reasoning | Highest tier; max supported | Higher reasoning tier |
Published benchmark gaps: small in some work, large in others
All figures in the following tables are from OpenAI’s GPT‑5.6 release tables. They compare the same model family, so this is a cleaner comparison than cross-vendor benchmark charts. Still, benchmark results are proxies—not a substitute for running your own agent harness and task set.
| Coding & agent evaluation | Sol | Terra | Terra as % of Sol |
|---|---|---|---|
| Artificial Analysis Coding Agent Index v1.1 | 80.0 | 77.4 | 96.8% |
| SWE‑bench Pro | 64.6% | 63.4% | 98.1% |
| DeepSWE v1.1 | 72.7% | 69.6% | 95.7% |
| Terminal‑Bench 2.1 | 88.8% | 87.4% | 98.4% |
| Agents’ Last Exam | 52.7% | 50.4% | 95.6% |
| Artificial Analysis Intelligence Index v4.1 | 58.9 | 55.0 | 93.4% |
Professional work and computer use
The gap grows somewhat on long-horizon professional workflows, design/computer-use tasks, and browse-heavy work. Even there, Terra stays close enough to be an attractive first-pass tier when the workflow can retry or escalate.
Professional reasoning
| Evaluation | Sol | Terra |
|---|---|---|
| GDPval‑AA v2 | 1,747.8 Elo | 1,593.0 Elo |
| Management consulting tasks | 43.2% | 37.2% |
| Big Finance Bench | 53.0% | 51.0% |
| AA Intelligence Index | 58.9 | 55.0 |
| HealthBench Professional | 60.5% | 57.7% |
Computer use & browsing
| Evaluation | Sol | Terra |
|---|---|---|
| OSWorld 2.0 | 62.6% | 50.2% |
| BrowseComp | 90.4% | 87.5% |
| BenchCAD | 70.6% | 62.3% |
| BenchCAD with Python tool | 83.4% | 78.2% |
| MMMU Pro with tools | 84.6% | 82.0% |
Where Sol’s premium is easiest to justify
The largest published differences appear on high-difficulty scientific, security, and abstract-reasoning evaluations. These may not map to ordinary product features, and cyber benchmarks are particularly sensitive; use them as signs of capability headroom, not deployment instructions.
| High-difficulty evaluation | Sol | Terra | Gap |
|---|---|---|---|
| GeneBench Pro | 28.7% | 23.3% | +5.4 |
| MedChemBench (internal) | 48.3% | 35.0% | +13.3 |
| SEC‑Bench Pro | 71.2% | 57.7% | +13.5 |
| ExploitBench | 73.5% | 52.9% | +20.6 |
| FrontierMath Tier 4 | 83.0% | 68.3% | +14.7 |
| ARC‑AGI‑3 | 7.78% | 0.80% | +6.98 |
Radar: capability headroom vs cost efficiency
This radar is a transparent decision aid, not a universal quality score. The axes summarize the official tables above; “price efficiency” is inverted so Terra’s half-price advantage extends further.
● GPT‑5.6 Sol ● GPT‑5.6 Terra
- Terra nearly overlays Sol on Terminal‑Bench, SWE‑bench Pro, and most professional work.
- Sol’s lead widens on the hardest science, security, and abstract-reasoning tasks.
- Terra’s value is structural: the per-token price is exactly half, not a temporary promotion.
What the independent view adds
Artificial Analysis reports Sol at 59 and Terra at 55 on its Intelligence Index at max reasoning. Its Coding Agent Index places Sol at 80 and Terra at 77. It also estimates Terra’s cost per Intelligence Index task at $0.55, about 50% below Sol’s $1.04. That is consistent with the list-price ratio, but it is an estimated task-cost measurement—not a guaranteed bill for your application.
Routing guide
The task is repeatable, retryable, or high volume: extraction, structured output, support workflows, routine code transformations, first-pass review, summarization, and agent substeps with a human or Sol escalation path.
The failure cost is high or the task is unusually hard: repo-wide debugging, complex terminal agents, difficult scientific analysis, complex design/computer-use jobs, and workflows that must succeed with fewer retries.
Output length or tool loops dominate cost. Cheap tokens do not automatically mean cheaper resolutions. Compare success rate, retries, output tokens, tool calls, latency, and human review time on a held-out set of real tasks.
Verdict
Terra is not a “mini” model in the usual sense. On the published coding and professional tables it often retains 94–98% of Sol’s score for half the token price. That makes it the natural baseline for most production traffic. Sol is the deliberate escalation tier. It is worth its 2× rate when you buy down failure risk on difficult agentic, scientific, security, or abstract-reasoning work.
The best architecture is rarely a single model choice: start with Terra, capture task-level outcomes, and route complex or failed cases to Sol. The measurement that matters is cost per successful task in your environment—not cost per million tokens in isolation.
Sources & methodology
- OpenAI — GPT‑5.6 launch and complete evaluation tables: all same-family benchmark figures and core pricing in this article.
- OpenAI API docs — GPT‑5.6 Sol and GPT‑5.6 Terra: context, maximum output, caching, and long-context pricing.
- Artificial Analysis — GPT‑5.6 benchmarks across intelligence, speed and cost: independent index scores and estimated per-task costs.
- All benchmark values are published results and may depend on model effort, scaffolding, tool access, and evaluation configuration. Charts preserve native scores and are not composite rankings.