What are the verified coding and reasoning benchmarks for GPT 5.4 Mini?

GPT 5.4 Mini achieves SWE-bench Verified: 61.2% (Thinking), LiveCodeBench: 64.5%, AIME: 88%. Verified sources: OpenAI GPT-5.4-mini Performance Report (OpenAI).

How do I run GPT 5.4 Mini with Swarm?

Configure GPT 5.4 Mini via CLI using: swarm ctl model set --provider openai --model openai/gpt-5.4-mini, or switch models interactively inside the terminal TUI with /models.

Does Swarm isolate GPT 5.4 Mini code edits from other agents?

Yes. Swarm provisions isolated Git worktrees for every session and subagent, preventing file race conditions and keeping your main working tree clean until verification passes.

Can GPT 5.4 Mini be assigned to specific subagent roles?

Yes. In ~/.swarm/config.json, you can configure GPT 5.4 Mini as Primary/Planner (DAG task decomposition), Coder (implementation in isolated worktrees), or Finder (read-only symbol search).

GPT 5.4 Mini — Specs, Context & Capabilities

Q: What is the recommended Swarm multi-agent topology for GPT 5.4 Mini?

Swarm recommends GPT 5.4 Mini as Coder Subagent (Tier 3: Elite High-Throughput Subagent). High throughput and low pricing make it the premier engine for concurrent Coder and Finder jobs. Best paired with: Claude 3.7 Sonnet as Lead Planner.

GPT 5.4 Mini Benchmarks & Agentic Accuracy

Verified primary research & developer preference metrics

Primary Sources Verified

SWE-bench Verified

61.2%

ThinkingOpenAI Technical Report ↗

LiveCodeBench

64.5%

ThinkingLiveCodeBench ↗

LMSYS Coding Elo

1390Elo

Coding ArenaLMSYS Arena ↗

AIME 2024 / 2025

88%

ThinkingOpenAI Technical Report ↗

Pricing ($ / 1M tokens)

$0.5/ $2

$0.12 cached readStandard API Pricing

What This Means in Real Software Development

Breaks the 60% barrier on SWE-bench Verified (61.2%) for a lightweight, sub-dollar model.
High throughput ensures rapid test authoring, code compilation, and error recovery.
Strong 88.0% on AIME demonstrates deep logic for competitive algorithms and data structures.
The ideal model for running 5+ concurrent Coder subagents in Swarm.

Swarm Multi-Agent Topology

Tier 3: Elite High-Throughput SubagentRole: Coder Subagent

Best Paired With: Claude 3.7 Sonnet as Lead Planner

High throughput and low pricing make it the premier engine for concurrent Coder and Finder jobs.

Git Worktree Isolation Advantage:

Executes parallel bug fixes across isolated worktrees without file contention.

What Developers & Practitioners Are Saying

Real-world field observations, production strengths, and developer consensus

Field Consensus

"Highly tuned compact model featuring refined code synthesis and robust error recovery."

Verified Real-World Strengths

61.2% SWE-bench Verified
Reliable git diff generation with minimal syntax regressions
Fast turn latency

Production Caveats & Gaps

Still benefits from isolated test runners to verify changes

Primary Sources & Methodology Citations

OpenAI GPT-5.4-mini Performance Report — OpenAI (March 2026)

GPT 5.4 Mini model overview

GPT 5.4 Mini is listed in the public Swarm catalog through OpenAI. This page summarizes the verified catalog fields used by Swarm without claiming real-time access for every provider account.

Provider: OpenAI
Family: Gpt 5
Context window: 400K tokens
Thinking support: off · low · medium · high · xhigh
Availability status: Available
Catalog ID: openai/gpt-5.4-mini

File input
Image input
Text input
Text output
Reasoning
Tools
Structured output
JSON
Batch

Run GPT 5.4 Mini in Swarm

Verified CLI workflow & multi-agent architecture

Quickstart Guide →

bash — swarm run

$swarm ctl model set --provider openai --model openai/gpt-5.4-mini

# Or switch interactively inside the terminal TUI:

>/models

Subagent Role Matrix

Assign GPT 5.4 Mini to specialized system roles in ~/.swarm/config.json:

Primary / Planner: Orchestrates full task DAGs and multi-stage pipelines.
Coder Subagent: Generates code and unit tests inside private Git worktrees.
Finder Subagent: Performs read-only code search, symbol discovery, and mapping.

System Agent Roles →Why Swarm Architecture →