GPT-5.4 nano

GPT-5.4 mini and nano are smaller, faster versions of GPT-5.4 optimized for coding, tool use, multimodal reasoning, and high-volume API and sub-agent workloads.

OpenAI's GPT-5.4 nano, released March 17, 2026, is positioned for coding assistants, subagents, and computer-control workflows, running more than twice as fast as its predecessor. The model scores 52.4% on SWE-Bench Pro and 39.0% on OSWorld-Verified, a major jump from GPT-5 nano's results. OpenAI recommends it for classification, data extraction, ranking, and lighter coding subagent tasks. The article notes pricing for the new GPT-5.4 mini and nano models is up to 4x higher than their GPT-5 predecessors, while delivering substantial capability gains. In OpenAI's Codex setup, GPT-5.4 mini consumes roughly 30% of the full GPT-5.4 quota for simpler subtasks, enabling lower-cost multi-agent orchestration patterns.

Vals AI independently evaluates GPT-5.4 Nano as a model from OpenAI released March 17, 2026, listing a 400k context window and 128k max output tokens. Reported token pricing is $0.20 per million input and $1.25 per million output tokens, with text, image, and file inputs supported but video not supported. Independent benchmark measurements include 84.01% on LiveCodeBench, 77.92% on LegalBench, 77.53% on GPQA Diamond, 77.17% on MMLU Pro, and 49.03% overall accuracy. Scores on Finance Agent (38.22%), Legal Research Bench (6.25%), and Tax Agent Bench (26.58%) show weaker performance on specialized agentic and legal tasks compared to coding evaluations.

BenchmarkList presents GPT-5.4 Nano-specific results across agentic evaluations, including a Terminal-Bench Hard success rate of 42.4% (rank 26/32, 92nd percentile, dated 2026-06-10), Tau3-Banking Pass@1 of 27.4% (rank 38/174, 79th percentile, dated 2026-09-02), GDPval-AA Score Elo of 1,104 (rank 79/340, 77th percentil On competitive comparisons, GPT-5.4 Nano's 27.4% on Tau3-Banking trails field leaders like GLM-5.3 at 50.3% (−22.9 pp), and its GDPval-AA Elo of 1,104 sits 756.75 points below Opus 5 (1,861 Elo). The benchmark tracker also surfaces a Hindsight LLM Memory Leaderboard entry for the model. These independently sourced scor

EdTech Innovation Hub reports that OpenAI released GPT-5.4 nano alongside GPT-5.4 mini on March 19, 2026, as smaller models designed for high-volume workloads in production environments. Sourced from an OpenAI for Business LinkedIn post, the models extend GPT-5.4 capabilities into more efficient formats for coding assi GPT-5.4 nano is described as the smallest and lowest-cost model in the series, optimized for classification, data extraction, ranking, and coding subagents. The article situates the release within a broader architectural shift toward multi-model systems, in which larger models handle planning and decision-making while

According to the Roboflow Playground page, GPT-5.4 nano was developed by OpenAI and released on March 17, 2026, as the efficiency-optimized entry in the GPT-5.4 family. The page highlights an expanded 400,000-token context window enabling large document batches or extensive logs to be processed in a single pass. The mo GPT-5.4 nano supports both text and image inputs but is positioned as a text-first model rather than a specialized visual reasoning tool — the page notes intensive vision reasoning and UI navigation are better delegated to GPT-5.4 mini. Compared to GPT-5 nano, this version offers a significant leap in reliability for s

The AI Release Tracker catalog lists GPT-5.4 nano as one of OpenAI's 44 tracked model releases, dated Tuesday, March 17, 2026, alongside GPT-5.4 mini. The page situates the model within OpenAI's broader release cadence, which averages a new model approximately every 37 days, and provides release counts by year and day The tracker notes that GPT-5.4 nano is one of two tracked 2026 releases in the GPT-5.4 line, with the broader family including GPT-5.4, GPT-5.4-Pro, GPT-5.4 mini, and GPT-5.4 nano. The page characterizes 2 of the 44 tracked OpenAI models as open-weight or open-source, with the rest proprietary. As an aggregator, the pa

An independent intelligence and performance index lists GPT-5.4 nano as a distinct variant within OpenAI's catalog, reporting measured output speed of 166 tokens per second for the xhigh configuration. GPT-5.4 nano (Non-reasoning) achieves a latency of 0.76 seconds to first answer token, placing it among the lowest-lat The same index ranks GPT-5.4 mini (xhigh) at 191 t/s output speed, placing both GPT-5.4 variants among the fastest OpenAI models while GPT-5.4 nano specifically targets speed-critical, cost-sensitive workloads. This aggregator data provides independent corroboration of nano's variant-level performance characteristics w

OpenRouter's provider page documents GPT-5.4 nano as the lightweight, cost-efficient variant of the GPT-5.4 family, designed for speed-critical and high-volume tasks such as classification, data extraction, ranking, and sub-agent execution. It supports text and image inputs, offers a 400K context window, and lists a re The page provides live performance and routing data across providers: OpenAI direct hosting delivers roughly 0.64s p50 latency and 94 tokens/second throughput, while Azure trails at about 1.42s latency and 45 tps, with 99.83% uptime reported for Azure. Rolling 30-day weighted averages show effective input pricing aroun