GPT-5.4 Nano

GPT-5.4 Nano is OpenAI’s most affordable GPT-5.4 variant for high-volume standard tasks such as classification, extraction, and ranking. With a context window of 272,000 tokens and up to 128,000 tokens of output, the model is well-suited for batch processing and sub-agent routing. Available exclusively via the OpenAI API at low cost.

OpenAI Version 5.4-nano Commercial use permitted Dense 272 K Context 08/2025 $0.2 / $1.25 per 1M

Proprietary
Frontier
OpenAI
Text
Vision
Instruction-Tuned
Real-Time

Sovereign Risk: MEDIUM OpenAI is a US-based company and subject to the CLOUD Act. When using the API, input data leaves the local network — government access to processed data is legally possible.

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	35	20	26
HTTP Fetch & Extract	80	35	57.5
Tool Failure Handling (404)	80	60	73
Web Search & Tool Selection	35	20	28.3
URL Construction & Fetch	80	80	80
Multilingual Search & Synthesis	35	20	27.5

Name

Combined

EU License Research

HTTP Fetch & Extract

57.5

Tool Failure Handling (404)

Web Search & Tool Selection

28.3

URL Construction & Fetch

Multilingual Search & Synthesis

27.5

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

Reliability

Tool Call Valid: No

Retry: Not required

Hallucination: 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: 1.19; First Request

MCP: 0.2; Protocol Latency

Synthesis: 1.94; Response Generation

Total: 20; Sum of All Phases

Token: 6653; 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.

Deployment Verdict

Conditional deploy, because tool execution works in parts, but the combination of invalid tool calls, detected hallucination, and weak overall synthesis makes the model unreliable for autonomous MCP pipelines.

Tool Execution Profile

GPT-5.4 Nano shows no stable tool judgment. On the Web Search and Tool Selection test — which checks whether web_search is chosen over fetch without a hint — it falls noticeably short. On the URL Construction and Fetch test, which measures independent derivation of a target URL, it performs solidly. This does not suggest genuine tool intelligence; it points more to a fixed pattern: known or directly constructable URLs work, open-ended research paths do not. For MCP-backed infrastructure this is critical, because dynamic pipelines demand exactly this selection capability. The fact that the tool call was marked as not valid overall sharpens the finding. The problem therefore lies not only in response quality but in protocol compliance toward the tool layer.

Synthesis Fidelity

How well does it consolidate tool results? Weakly. The P2 performance shows that GPT-5.4 Nano frequently fails to merge retrieved content with sufficient precision. Consolidation quality drops sharply in particular on EU License Research, Web Search & Tool Selection, and Multilingual Search & Synthesis. Usable extraction from clearly available fetch content is possible, but the translation into reliable answers remains inconsistent.

Does it stay within the tool result or fall back on training? The trust signal here is mixed. On the honeypot EU License Research test — which checks whether current license restrictions are actually retrieved from web sources — no hallucination was detected. That is a positive. At the same time, one hallucination was detected globally. In a tool pipeline this is not merely a quality deficiency but a security risk: once a model outputs fabricated facts as apparent tool results, the entire infrastructure loses its auditability.

Error Resilience

On the 404 test, which checks how the model handles a failing tool call, it responds acceptably. It does not hallucinate page content despite the error and therefore remains fundamentally transparent. The P2 rating is only middling, but the decisive point for production is met: it does not obscure the failure with invented fallback content.

Operational Profile

Call 1: 1.19s. Call 2: 1.94s. MCP latency: 0.20s. Total: 20.00s. Price: stated locally. Fast at the individual-call level for the performance shown, but the overall run is not efficient relative to the weak result quality.

Conclusion & Recommendation

Suitable as a low-cost auxiliary component for tightly scoped tasks with a predefined URL, simple extraction, and controlled error handling. Not suitable as an autonomously deciding model for research, tool selection, multilingual synthesis, or compliance-adjacent responses. If it enters a pipeline, then only with hard tool validation, schema checking, response guardrails, and downstream verification by a more reliable model or deterministic checks.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.