GPT-5.4 mini

Overview

GPT-5.4 mini is the mid-tier model in the GPT-5.4 family, released on March 17, 2026 with GPT-5.4 nano. OpenAI positioned it as the go-to option for developers who want near-flagship output quality without paying flagship prices: at $0.75 per million input tokens, it costs 70% less than the standard GPT-5.4 ($2.50/M).

The performance numbers back that framing up. Mini scores 54.4% on SWE-Bench Pro against GPT-5.4's 57.7% - a 3.3-point gap that most production coding workflows won't notice. On GPQA Diamond, which tests graduate-level science reasoning, it reaches 88.0% vs the flagship's 93.0%. OSWorld-Verified puts it at 72.1%, which crosses the 72.4% human baseline on desktop navigation tasks.

Agentic benchmarks are where mini especially stands out. On Toolathlon it scores 42.9%, a 16-point jump over the prior-gen GPT-5 mini (26.9%). MCP Atlas puts it at 57.7%, up from 47.6% for the predecessor. For developers building tool-calling pipelines and agentic AI workflows, those gains matter more than the headline science or coding scores.

Mini is also available to ChatGPT's free and Go tiers - the first GPT-5 family model accessible without a subscription.

Key Specifications

Benchmark Performance

The benchmark picture is consistent across categories: mini lands within a few percentage points of the flagship, well ahead of its predecessor GPT-5 mini.

The 3.3-point gap on SWE-Bench Pro against GPT-5.4 is the number to watch. For pure coding agent throughput at scale, it's a strong justification for choosing mini over the flagship. The larger gap on GPQA Diamond (5 points) means the flagship is the right call for anything requiring deep scientific or mathematical reasoning.

For full competitive context, see our coding benchmarks leaderboard and agentic AI benchmarks leaderboard.

GPT-5.4 mini is accessible via the OpenAI API and included in ChatGPT free tier starting March 17, 2026. Source: unsplash.com

Key Capabilities

Coding. SWE-Bench Pro at 54.4% is competitive for a sub-$1/M input model. The previous GPT-5 mini series had a larger gap from the flagship; OpenAI has substantially closed it here. For code review, generation, and autocomplete tasks, mini is a viable drop-in at lower cost.

Computer use. OSWorld-Verified at 72.1% isn't far behind the flagship's 75.0%. At 2.9 points back but 70% cheaper, mini is the practical default for computer use deployments that don't need to push the absolute ceiling. For comparison, it still beats the human baseline of 72.4%.

Agentic tool use. The Toolathlon and MCP Atlas numbers show the clearest generational improvement. The 16-point Toolathlon jump over GPT-5 mini suggests that OpenAI has specifically tuned mini for tool-calling and multi-step agent workflows - which maps to their stated positioning for the model.

Speed. OpenAI states mini runs more than 2x faster than GPT-5 mini. Community throughput estimates put it around 180-190 tokens per second at normal priority, roughly 3x the rate of the older GPT-5 mini. For latency-sensitive production apps, that's the spec that matters.

Benchmark evaluation setup for testing model performance across coding, reasoning, and agentic tasks. Source: unsplash.com

Pricing and Availability

GPT-5.4 mini is available through the OpenAI API with standard, cached, and batch pricing tiers.

The cached input rate at $0.075/M is the standout number. At 90% off standard pricing, it makes mini extremely cost-effective for applications that repeatedly process similar long context - RAG systems, document analysis pipelines, and chat applications with long system prompts.

Compared to its siblings in the GPT-5.4 family: nano runs $0.20/$1.25 (for classification and sub-agent tasks) while the flagship sits at $2.50/$15.00. Mini fills the middle slot - more capable than nano for complex reasoning, far cheaper than the flagship for most production workloads.

Against competitors: Gemini 3.1 Flash-Lite costs $0.40/$1.50 per million tokens, cheaper on base pricing but weaker on agentic benchmarks. Claude Haiku pricing lands around $0.80/$4.00, close to mini's base rate but without the 90% caching discount.

Mini is available to ChatGPT Free, Go, Plus, Team, and Pro subscribers, and through the OpenAI API. For a full cross-model cost comparison, see our cost efficiency leaderboard.

Strengths

• Close to flagship quality at 30% of the cost. 3.3 points behind GPT-5.4 on SWE-Bench Pro, $1.75 cheaper per million input tokens
• Strong agentic performance. 16-point Toolathlon gain over predecessor; MCP Atlas at 57.7% makes it a practical choice for multi-step tool workflows
• 90% caching discount. $0.075/M cached input is one of the best cache rates available and changes the economics for context-heavy applications
• Crosses human baseline on computer use. 72.1% OSWorld-Verified beats the 72.4% human benchmark, though the flagship's 75.0% still leads
• 2x faster throughput than GPT-5 mini. 180-190 tokens/second vs ~80 for the older model
• Broad availability. Free tier ChatGPT, all paid tiers, and full API access including batch

Weaknesses

• Smaller context window than the flagship. 400K vs GPT-5.4's 1M token context - the flagship is necessary for tasks requiring full large codebase or document corpus ingestion
• Science reasoning gap widens. 5-point GPQA Diamond gap (88.0% vs 93.0%) is more meaningful for math and science-heavy workflows than the coding gap
• No computer use in nano comparison. OSWorld-Verified numbers for nano aren't available, so the true cost-performance tradeoff for computer use between the two is unclear
• No parameter disclosure. Architecture details remain undisclosed, making it harder to assess on-prem or distillation feasibility
• Pricing beats Gemini Flash-Lite on standard rate, but not on base. Gemini 3.1 Flash-Lite at $0.40/$1.50 undercuts mini on standard pricing, though mini's caching rate compensates at scale

Related Coverage

• OpenAI's New Mini and Nano Slash GPT-5.4 Pricing - Launch coverage
• GPT-5.4 - The flagship this model is derived from
• GPT-4o mini - The prior-generation budget workhorse
• Cost Efficiency Leaderboard - Cross-model cost comparisons
• Coding Benchmarks Leaderboard - Full coding rankings
• Agentic AI Benchmarks Leaderboard - Tool-use and agent task rankings
• Small Language Model Leaderboard - Where mini fits among efficient models

FAQ

What is GPT-5.4 mini best for?

Coding agents, tool-calling pipelines, and computer use tasks where you need near-flagship performance without flagship pricing. Toolathlon at 42.9% and SWE-Bench Pro at 54.4% make it strong for agentic workflows at $0.75/M input.

How does GPT-5.4 mini compare to GPT-5.4?

Mini scores within 3.3 points on SWE-Bench Pro (54.4% vs 57.7%) and 2.9 points on OSWorld computer use (72.1% vs 75.0%), at 30% of the cost. The larger gap is on GPQA Diamond: 88.0% vs 93.0% for graduate-level science tasks.

What is the cached input price for GPT-5.4 mini?

$0.075 per million tokens - a 90% discount from the standard $0.75/M rate. At that price, context-heavy RAG and long system prompt applications become significantly cheaper to run.

Is GPT-5.4 mini available on the free ChatGPT tier?

Yes. Mini is the first GPT-5 family model available to free ChatGPT users, with Go tier subscribers.

What context window does GPT-5.4 mini support?

400,000 tokens. This is smaller than GPT-5.4's 1M token window but larger than GPT-4o mini's 128K context.

How fast is GPT-5.4 mini?

Community measurements put throughput around 180-190 tokens per second, which OpenAI describes as more than 2x faster than GPT-5 mini.

Sources

• OpenAI's New Mini and Nano Slash GPT-5.4 Pricing - Awesome Agents
• OpenAI API Pricing
• GPT-5.4 mini and nano - OpenAI