Skip to content
CrucibleMark

GPT 5.6 Luna

GPT-5.6 Luna is the cheapest and fastest tier of OpenAI’s three-tier GPT-5.6 series (Sol, Terra, Luna) for high-volume, latency-sensitive tasks — available since July 30, 2026 at $0.20 / $1.20 per million tokens, roughly 80 percent below Sol. The 1-million-token context variant with 128,000 output tokens delivers frontier-adjacent agentic performance according to OpenAI, but falls off noticeably against its larger siblings on context recall beyond 512,000 tokens.

OpenAI Version 5.6 Commercial use permitted Dense 1000 K Context 02/2026 $0.2 / $1.2 per 1M

  • Proprietary
  • Frontier
  • OpenAI
  • Text
  • Vision
  • Real-Time

Sovereign Risk: MEDIUM The model is developed and hosted by a US-based company. Due to US jurisdiction, it is potentially subject to the CLOUD Act, which represents a moderate risk of data access by US authorities. Since the weights are proprietary and not distributed, there is no additional risk from disclosure of the weights themselves.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 95 40 62
HTTP Fetch & Extract 75 80 77.5
Tool Failure Handling (404) 80 100 87
Web Search & Tool Selection 95 80 88.3
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 70 85

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
1.97
First Request
MCP
1.88
Protocol Latency
Synthesis
5.26
Response Generation
Total
54.73
Sum of All Phases
Token
11377
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

Created

Deployment Verdict

Conditional deploy: tool execution is strong, but tool-call validity is not consistently reliable, and synthesis fidelity in the honeypot is too weak for trust-critical pipelines.

Tool Execution Profile

GPT 5.6 Luna demonstrates genuine tool intelligence, not just rigid retrieval. In the Web Search and Tool Selection test — which requires distinguishing between search and direct fetch without an explicit hint — it selects the correct tool with high confidence. This points to workable orchestration in open MCP flows. In the URL Construction test, which requires deriving the target URL from its own knowledge and then fetching it correctly, it performs adequately but with less precision. That is exactly where the limit becomes visible: it usually identifies the correct operation type, but does not produce a fully reliable call at every step. The fact that the tool call was rated invalid overall matters more for production than the solid P1 average. This is not a comprehension problem at the task level — it is a protocol and execution risk at the edges.

Synthesis Fidelity

How well does it consolidate tool results? Solidly, but not reliably enough for strict fact pipelines. In HTTP Fetch & Extract and Multilingual Search & Synthesis it consolidates results usably and mostly in a structured way. The overall score remains only in the solid mid-range, however, because output quality varies noticeably across tasks. For analytical assistance this is often sufficient. For deterministic handoffs to downstream systems it is too inconsistent.

Does it stay within the tool result or fall back on training data? This is the core problem. In the EU License Research honeypot — designed to verify whether current licensing restrictions genuinely come from web sources rather than model knowledge — synthesis falls off clearly. It does not hallucinate overtly, but the trust signal is weak: the model does not reliably demonstrate that it strictly confines its answer to the retrieved tool context. For compliance, regulatory, and current policy use cases, this is a warning sign.

Error Resilience

Good for production. In the 404 test — which checks whether a failing tool call is handled with transparency rather than fallback hallucination — the model responds cleanly. It does not fabricate page content and communicates the failure correctly. This behavior is exactly what keeps a tool pipeline stable when external sources drop out.

Operational Profile

Call 1: 1.97s. MCP latency: 1.88s. Call 2: 5.26s. Total: 54.73s.
Cost-efficient, but not fast in end-to-end runs. Attractive on price; limited for latency-sensitive multi-step pipelines.

Conclusion & Recommendation

Suitable for agentic research and routing pipelines where tool selection, fault tolerance, and cost matter more than hard factual fidelity in the final consolidation. Not suitable for compliance, legal, or policy workflows where the model must remain strictly bound to current tool evidence. I would deploy it as a cost-efficient orchestrator with downstream validation — not as the final trust-granting authority.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.