Jun 26, 2026
OpenAI announced GPT-5.6 Sol, Terra, and Luna together as a preview family.
Source: OpenAI announcementUpdated June 28, 2026
GPT-5.6 Sol
Can you use GPT-5.6 Sol, what does it cost, and how do its scores compare?
Start here before building around the name. OpenAI describes GPT-5.6 as a limited preview family: Sol is the flagship, Terra is the lower-cost capability tier, and Luna is the fastest low-cost tier. This page separates official facts from leaderboard data: access comes from OpenAI's Help Center, pricing comes from the preview access note, safety and capability scores come from the System Card, and Claude/Gemini/Grok/DeepSeek/Qwen comparisons come from Artificial Analysis model pages.
Limited preview
This is not a normal self-serve launch. Treat access as account-specific until your API model list or OpenAI contact confirms it.
Source: OpenAI Help CenterSol / Terra / Luna
Sol is the flagship; Terra is the lower-cost capability option; Luna is the fastest and most cost-efficient tier.
Source: OpenAI announcement$5 / $30
Published Sol API price per 1M input/output tokens in OpenAI's preview access note; prompt-cache rules need live confirmation.
Source: OpenAI Help CenterHigh capability
OpenAI treats the family as High capability in Cybersecurity and Biological/Chemical risk, but not Critical.
Source: GPT-5.6 System Card96.7%
Sol saturates OpenAI's internal CTF cyber evaluation; this is not a generic all-task score.
Source: GPT-5.6 System CardNumbers are linked to their source. External leaderboard values were captured on June 28, 2026 and should be rechecked before buying or routing traffic.
Release and Access Status
What users need to know first
If it is not in your API model list, you probably cannot use it yet.
For a builder, the only useful access check is concrete: open your OpenAI project, inspect the available model IDs, and confirm whether Sol, Terra, or Luna is enabled. A press release or screenshot from another account is not deployable access.
Check API, Codex, and product surface separately.
Do not assume ChatGPT, API, and Codex expose the same preview models. Before writing code, record the exact model ID, account or org where it appears, rate limits, and whether tool access or safety gating changes the task you want to run.
Use Sol only for steps where failures are more expensive than tokens.
A practical routing plan starts with three buckets: run Luna for extraction and routing, Terra for repeated work that still needs quality, and Sol only for hard reasoning, code edits, cyber-defense review, or expensive failures that justify premium pricing.
Price and Model Family
Sol is not the whole family
GPT-5.6 Sol
Flagship model for difficult reasoning, coding, cyber-defense review, and agentic tasks where quality matters more than token price.
- Published API price
- $5 input / $30 output per 1M tokens.
- Access posture
- Limited preview; verify eligibility and exact model ID in your OpenAI account before implementation.
- Best use
- Use for upper-bound evals, difficult code changes, high-value analysis, safety review, and cases where one wrong answer costs more than the model call.
GPT-5.6 Terra
The middle tier to test when Sol is too expensive for every call but Luna loses too much quality on your own tasks.
- Published API price
- $2.50 input / $15 output per 1M tokens.
- Access posture
- Same preview-family status; confirm whether the same account surface exposes Terra before routing.
- Best use
- Use as the default comparison against Sol on repeated workflows; keep it only if task-level pass rate stays close enough.
GPT-5.6 Luna
The low-cost tier for tasks that can be checked cheaply: extraction, classification, routing, summaries, and first-pass drafts.
- Published API price
- $1 input / $6 output per 1M tokens.
- Access posture
- Preview access still needs verification; do not use Luna performance as a proxy for Sol.
- Best use
- Use for cheap preprocessing, classification, extraction, summaries, and routing before escalating to Terra or Sol.
Pricing shown here is the public per-1M token preview price captured on June 28, 2026 from OpenAI's Help Center note. Do not copy old cached-token assumptions into a budget; recheck the access note and your account limits before production traffic.
Benchmark and Evaluation Cards
Numbers plus interpretation
96.7%
Internal CTF cyber evaluation
The System Card says GPT-5.6 Sol reaches 96.7% on OpenAI's curated internal Capture-the-Flag task set.
This is a cyber-specific result under a tool-enabled harness. It is useful for judging cyber capability and safeguards, not for ranking writing quality, business analysis, or everyday chat.
Source: GPT-5.6 System Card60.5
HealthBench Professional
The System Card reports a 60.5 length-adjusted HealthBench Professional score for Sol; the separate HealthBench Consensus score is 95.5.
Use this to understand relative model behavior on medical QA benchmarks, not as medical advice. Tool access, browsing, answer length, and policy constraints all change the result.
Source: GPT-5.6 System CardHigh, not Critical
CVE-Bench, VulnLMP, ExploitBench context
The System Card separates CTF tasks, CVE-Bench, VulnLMP, ExploitBench, ExploitGym, and SEC-Bench Pro instead of collapsing cyber into one score.
When you evaluate your own security workflow, split it the same way: triage, reproduction, exploit primitive development, and full exploit chains have different permission gates.
Source: GPT-5.6 System Card11.3h / 270h+
METR time-horizon uncertainty
METR discussion reports that treatment of environment exploitation can swing Sol's time-horizon estimate from about 11.3 hours to more than 270 hours.
For agent tasks, do not trust a single success rate. Log tool calls, hide parts of the test, prevent environment shortcuts, and review failures manually.
Source: METR evaluation context750 tok/s
Cerebras serving speed
OpenAI says GPT-5.6 Sol will run on Cerebras at up to 750 tokens per second for select customers as capacity expands.
Fast streaming helps agent loops only after the model can complete the task. Measure end-to-end time, including tool calls, retries, refusals, and human review.
Source: OpenAI announcementExternal Model Comparison
Scores captured June 28, 2026
| Representative model | Provider | Benchmark | Score | Price | Speed / latency | Score page |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | OpenAI | OpenAI System Card | CTF 96.7% / HealthBench Pro 60.5 | $5 input / $30 output | 750 tok/s; TTFT N/A | |
| Claude Opus 4.8 (max) | Anthropic | AA Intelligence Index | 56 | $3.85 blended | 59 tok/s; TTFT 13.51s | |
| Gemini 3.1 Pro Preview | AA Intelligence Index | 46 | $1.74 blended | 140 tok/s; TTFT 33.92s | ||
| Grok 4.3 (high) | xAI | AA Intelligence Index | 38 | $0.64 blended | 139 tok/s; TTFT 18.28s | |
| DeepSeek V4 Pro (Max) | DeepSeek | AA Intelligence Index | 44 | $0.18 blended | 78 tok/s; TTFT 1.78s | |
| Qwen3.7 Max | Alibaba | AA Intelligence Index | 46 | $1.43 blended | 203 tok/s; TTFT 2.50s |
External rows use Artificial Analysis model pages: Intelligence Index, blended price, median output speed, and time-to-first-token. GPT-5.6 Sol is listed separately with OpenAI System Card metrics because it was not available as an Artificial Analysis leaderboard row at capture time.
Preparedness classification
OpenAI treats the GPT-5.6 preview family as High capability in Cybersecurity and Biological/Chemical risk. That means release and tool access need stronger controls than a normal text model rollout.
Source: GPT-5.6 System CardAgentic behavior needs logs
The System Card discusses metagaming and chain-of-thought monitoring; METR highlights how environment handling changes task-horizon estimates. Treat eval integrity as part of the deployment design.
Source: GPT-5.6 System CardDeployment controls are part of adoption
Before giving Sol tools, define allowed actions, require confirmation for destructive steps, log every tool call, cap spend, and keep a lower-cost fallback for routine work.
Source: GPT-5.6 System CardQuestions People Actually Ask
Short answers with source links
Is GPT-5.6 Sol an official OpenAI model?
Yes. OpenAI announced GPT-5.6 Sol, Terra, and Luna on June 26, 2026. This site is an independent guide and links back to official sources for verification.
Can I use GPT-5.6 Sol today?
Only if your OpenAI account or organization is included in the preview. Check the API model list and Help Center access note; do not infer availability from another user's screenshots.
Does 96.7% CTF mean Sol is best at everything?
No. It is a cyber evaluation result under a specific tool-enabled setup. It is strong evidence for cyber capability and safety planning, but it says little about your product's writing, research, coding, or support workload until you test those tasks directly.
Can the external scores decide which model to use?
No. The table gives a current scoring snapshot and links back to the score pages, but production choice still depends on your prompts, tools, latency budget, price sensitivity, and safety constraints.
Evidence Index
What each source supports
Official launch
OpenAI GPT-5.6 Sol announcement
Release date, family framing, Sol/Terra/Luna positioning, limited preview context, and Cerebras speed claim.
Use this first for what OpenAI actually announced, not for second-hand summaries.
System CardGPT-5.6 Preview System Card
Preparedness categories, CTF/CVE/VulnLMP/ExploitBench context, health evals, metagaming, and safeguards.
This is the highest-value source for numbers and risk boundaries.
Access + pricingOpenAI Help Center preview access note
Who can access the preview, how to verify availability, and the published Sol/Terra/Luna token prices.
Use this for GPT-5.6 pricing; the generic API pricing page may lag or omit preview-only models.
EvaluationMETR predeployment evaluation context
Time-horizon measurement uncertainty, environment exploitation, and why agent evals need careful methodology.
Useful counterweight to simple leaderboard thinking.
External scoresArtificial Analysis model leaderboard
Claude, Gemini, Grok, DeepSeek, and Qwen scores, prices, speed, latency, plus provider pages for official model identity.
Use the linked score pages for current benchmark values, then confirm model identity and access in the provider docs.
Access noteOpenAI Help Center preview article
Account eligibility, current access caveats, model-family availability, and preview-specific operational notes.
Check this before writing code; availability is more likely to change than the model-family announcement.