OpenAI's July 9 release didn't ship one model — it shipped three, each optimized for a different point on the cost-performance curve. Sol is the flagship. Terra undercuts GPT-5.5's price while matching or beating it on every major benchmark. Luna is faster and cheaper than both, and still clears GPT-5.5 on the hardest long-horizon agentic evaluations.
The three-tier structure makes model selection a real decision: pick the wrong tier and you're either overpaying by 5× or leaving accuracy on the table. Here's the data to get it right.
What Is GPT-5.6?
GPT-5.6 is OpenAI's newest model generation, designed around a clear principle: more useful work from every token. Rather than releasing a single flagship and calling it done, OpenAI structured GPT-5.6 as a tiered family where the generation number (5.6) tracks the release, while the tier names (Sol, Terra, Luna) describe durable capability levels that can each advance on their own cadence.
The family also introduces two reasoning modes — max and ultra — that let you push further when a task demands it. max gives the model more time to reason and explore alternatives. ultra goes further by coordinating four parallel agents by default, trading higher token spend for stronger results and shorter wall-clock time on demanding tasks.
The Three Tiers at a Glance
| Tier | Positioning | Input price | Output price |
|---|---|---|---|
| Sol | Flagship — highest intelligence and efficiency | $5 / 1M tokens | $30 / 1M tokens |
| Terra | Balanced — GPT-5.5-competitive at roughly half the cost | $2.50 / 1M tokens | $15 / 1M tokens |
| Luna | Fastest — lowest cost, best latency | $1 / 1M tokens | $6 / 1M tokens |
All three are available through the OpenAI API, ChatGPT (Plus, Pro, Business, Enterprise), and Codex starting from launch day.
Sol: The Flagship Tier
Sol is OpenAI's highest-capability model in the GPT-5.6 family. It sets new state-of-the-art results on several major benchmarks:
- Agents' Last Exam (55 professional domains, long-horizon agentic work): 52.7% — up from GPT-5.5's 46.9%
- Artificial Analysis Coding Agent Index: 80 index score — above Claude Fable 5 (77.2) while using less than half the output tokens
- BrowseComp (complex agentic browsing): 90.4%
- Terminal-Bench 2.1 (command-line workflows): 88.8%
- OSWorld 2.0 (computer use): 62.6% — while using 85% fewer output tokens than competitors at similar scores
Sol also supports ultra mode in ChatGPT Work (Pro and Enterprise) and Codex (Plus and above), which runs four parallel agents by default for the most demanding tasks.
Best for: Long-horizon agentic workflows, complex code reviews, multi-step research, knowledge work requiring high accuracy, tasks where output quality directly affects business outcomes.
Terra: The Balanced Tier
Terra is designed to deliver performance competitive with GPT-5.5 at a significantly lower cost. According to OpenAI, agents running GPT-5.5 can often match results by switching to Terra at half the cost and 16% fewer output tokens.
Key benchmark scores for Terra:
- Agents' Last Exam: 50.4% — above GPT-5.5 (46.9%)
- Coding Agent Index: 77.4 — above Claude Fable 5 (77.2)
- BrowseComp: 87.5%
- Terminal-Bench 2.1: 87.4%
Terra is the default model for Free and Go users in ChatGPT Work and Codex. For API developers, it represents the practical sweet spot for production workloads where sustained accuracy matters but Sol-level cost is not justified.
Best for: Production coding agents, everyday knowledge work, content generation pipelines, moderate-complexity multi-step tasks, workflows that previously ran on GPT-5.5.
Luna: The Fast Tier
Luna is optimized for speed and cost. At $1 input / $6 output per million tokens, it is the most affordable model in the family — and it still punches above what you might expect at that price point.
Luna's benchmark scores:
- Agents' Last Exam: 50.3% — still above GPT-5.5 (46.9%)
- Coding Agent Index: 74.6
- BrowseComp: 83.3%
- Terminal-Bench 2.1: 84.7%
That means Luna outperforms GPT-5.5 on agentic benchmarks while costing roughly one-quarter as much per token. For tasks that don't require Sol's peak intelligence, Luna's cost profile changes the math on what's practical to run at scale.
Best for: High-volume routing, triage, summarization, labeling, lightweight monitoring tasks, streaming responses where latency matters, parallel agent coordination at scale.
Side-by-Side Benchmark Comparison
| Benchmark | Sol | Terra | Luna | GPT-5.5 |
|---|---|---|---|---|
| Agents' Last Exam | 52.7% | 50.4% | 50.3% | 46.9% |
| Coding Agent Index | 80 | 77.4 | 74.6 | 76.4 |
| BrowseComp | 90.4% | 87.5% | 83.3% | 84.4% |
| Terminal-Bench 2.1 | 88.8% | 87.4% | 84.7% | 85.6% |
| OSWorld 2.0 | 62.6% | 50.2% | 45.6% | 47.5% |
| GPQA Diamond | 94.6% | 92.9% | 92.3% | 93.6% |
All scores sourced from OpenAI's official GPT-5.6 launch page (July 9, 2026).
How to Choose: A Practical Decision Framework
The right tier depends on the task profile, not just cost sensitivity.
Use Sol when:
- The task is long-horizon (multi-hour agent runs)
- Errors are expensive to catch or correct
- You need
ultramode for parallel agent orchestration - The output directly ships to end users without human review
Use Terra when:
- You previously ran GPT-5.5 and want the same or better results at lower cost
- The task is well-defined but still requires reliable multi-step reasoning
- You're building a production pipeline that needs to run continuously
Use Luna when:
- The task is high-volume and latency-sensitive
- You need a cheap, fast model for triage, classification, routing, or summarization
- Luna runs alongside Sol or Terra in a multi-agent setup (routing decisions, monitoring)
- You're prototyping and want to minimize costs before switching to a higher tier
Where the Capability Layer Fits In
Choosing the right GPT-5.6 tier handles the reasoning side of your agent stack. But reasoning alone doesn't finish most real-world tasks. Agents frequently need to create images, inspect media, search the web, generate video, or publish output — capabilities that no language model provides natively.
That's the gap AnyCap closes. AnyCap is the missing capability layer for AI agents: one CLI, one auth flow, and access to image generation, video production, web search, web crawl, persistent file storage, and page hosting. Your agent handles the reasoning with whichever GPT-5.6 tier fits the task; AnyCap handles everything the model alone cannot do.
The pairing is additive, not exclusive. A Sol-powered agent writing a technical report can call AnyCap to search the web for current data, generate diagrams, and publish the finished document as a hosted page — all without leaving the agent workflow.
Pricing Summary
| Tier | Input | Output | Cache write | Cache read (90% off) |
|---|---|---|---|---|
| Sol | $5.00 | $30.00 | $6.25 | $0.50 |
| Terra | $2.50 | $15.00 | $3.13 | $0.25 |
| Luna | $1.00 | $6.00 | $1.25 | $0.10 |
GPT-5.6 also introduces more predictable prompt caching: explicit cache breakpoints and a 30-minute minimum cache lifetime. Cache writes are billed at 1.25× the uncached input rate; cache reads continue to receive a 90% discount.
The Bottom Line
GPT-5.6's three-tier structure gives developers a genuine choice rather than a forced trade-off between one expensive flagship and nothing else. In practice, most production agent workflows will land on Terra — it covers the majority of tasks at a price that makes continuous operation sustainable. Sol steps in for the hardest problems where accuracy cannot be compromised. Luna handles the volume work: routing, triage, parallel coordination, and anything where throughput and latency matter more than peak intelligence.
The tier your agent uses for reasoning is one half of the equation. The other half is what your agent can actually do once it has the answer — and that's where a capability layer like AnyCap picks up where the model leaves off.
All benchmark data is sourced from OpenAI's official GPT-5.6 announcement (openai.com, July 9, 2026). Prices are per 1 million tokens at standard API rates.