GPT

OpenAI GPT

OpenAI

OpenAI's GPT-5.4 brings a 1M token context window, 5-level reasoning control, and 75% accuracy on OSWorld computer use tasks. The mini and nano variants offer aggressive pricing for high-volume deployments.

chatgpt.com
Last updated: April 20, 2026
$2.50/M
GPT-5.4 Input
1M
Context Window
128K
Max Output
94.3%
GPQA Diamond
★★★★★
Coding
★★★★★
Reasoning
★★★★
Writing
★★★★
Speed

Available Models

ModelInput $/1MOutput $/1MContextBest For
GPT-5.4 (flagship)$2.50$15.001MComplex reasoning, coding, agents
GPT-5.4-mini$0.75$4.50400KHigh-volume chat, lightweight tasks
GPT-5.4-nano$0.20$1.25400KEdge inference, classification
GPT-5.4-pro$10.00$40.001MMaximum quality, complex tasks
o3 (reasoning)$10.00$40.00200KDeep reasoning, math, science
o4-mini$1.10$4.40200KFast reasoning at lower cost

Consumer Plans

PlanPriceIncludes
Free$0/moGPT-5.4-mini access
Go$8/moBasic GPT-5.4 access
Plus$20/moFull GPT-5.4, DALL-E, advanced tools
Pro$200/moUnlimited access, o3, GPT-5.4-pro
Business$25/user/moTeam features, admin controls

Strengths & Weaknesses

Strengths
  • Leading GPQA Diamond score: 94.3% (OpenAI Docs)
  • 1M token context window
  • 5-level reasoning control (none to xhigh)
  • 75% OSWorld computer use accuracy
  • Multimodal: vision, realtime audio, image gen
Weaknesses
  • Can be verbose on simple queries
  • SWE-bench trails Claude (57.7% vs 80.8%) (NxCode)
  • Pro variant pricing is steep ($200/mo)

Best For

  • Professional workflows and complex reasoning
  • Computer use and autonomous subagents
  • Realtime voice and image generation
  • High-volume production with nano/mini tiers
  • Multi-modal applications (vision + audio + text)

Latest Release NEW

GPT-5.4 — March 2026

  • 1M+ token context window — largest in the GPT family
  • 75% OSWorld computer use accuracy for agentic tasks
  • 5-level reasoning control from quick answers to deep analysis
  • Native coding improvements with built-in code execution
  • LMSYS Arena Elo ~1495 — top-tier across benchmarks

Previous Releases

GPT-5.3 Codex — January 2026

  • Code-specialized variant with enhanced software engineering capabilities
  • Introduced multi-step tool use for complex coding tasks
  • Predecessor to GPT-5.4's integrated coding features

GPT-4o — May 2024

  • First natively multimodal GPT model — text, image, audio in one model
  • Significantly faster and cheaper than GPT-4 Turbo
  • Still widely used in production and available via API