GPT-5.3 Codex

OpenAI announced GPT-5.3-Codex on February 5, 2026 as its most capable agentic coding model, unifying the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge strengths of GPT-5.2 in a single system that is also 25% faster. The model is described as the first that helped build itself, with early versions used by the Codex team to debug training, manage deployment, and evaluate results. GPT-5.3-Codex sets new highs on SWE-Bench Pro and Terminal-Bench, with strong performance on OSWorld and GDPval. It achieves state-of-the-art SWE-Bench Pro results across four languages using fewer output tokens than prior models, and far exceeds earlier marks on Terminal-Bench 2.0. OpenAI positions the release as expanding Codex from a code-writing reviewer into an agent capable of nearly any developer or professional computer task.

According to BenchLM's benchmark aggregation page (updated September 8, 2026), GPT-5.3 Codex was released on February 5, 2026, with a 400K token context window and is tracked across 418 destinations and 427 benchmarks. The model receives a composite capability score of 64.2 out of 100, ranking it 41st out of 231 tracke Speed measurements place GPT-5.3 Codex at 139 tokens per second (field median 92 tok/s) with a first-token latency of 62.36 seconds, making it notably faster than peers on throughput. BenchLM notes that Reasoning, Knowledge, Math, Multilingual, Multimodal, and Instruction Following categories are not rank-eligible for

GitHub announced on May 17, 2026 that GPT-5.3-Codex became the base model for all Copilot Business and Copilot Enterprise organizations, replacing GPT-4.1 in that role. The model is also GitHub's first long-term support (LTS) offering built in partnership with OpenAI, guaranteeing availability for a full 12 months from The change applies only to Copilot Business and Enterprise plans; individual Copilot Pro, Pro+, and Free tiers follow the standard deprecation timeline. GitHub noted that enterprise telemetry showed a significantly high code survival rate for GPT-5.3-Codex, which informed its selection as the default. Organizations tha

The llm-stats benchmark aggregator lists GPT-5.3 Codex scores sourced to OpenAI's official materials, including SWE-Lancer IC-Diamond at 0.81/1, Cybersecurity CTFs at 0.78/1, Terminal-Bench 2.0 at 0.77/1, and LiveBench performance figures, all marked xhigh effort. The page also tracks capability tiers, real-tasks performance across prompt categories, and how scores hold up at increasing conversation depths. The aggregator catalogues additional evaluation tracks beyond OpenAI's headline benchmarks, spanning real-world freelance engineering tasks, offensive security challenges, and contamination-limited reasoning tests. Scores are attributed to the model card and official blog rather than independent re-runs, and the page itself notes an October 13, 2026 timestamp reflecting ongoing tracking of the model's performance.

DataCamp's explainer frames GPT-5.3-Codex as OpenAI's counter-release to Anthropic's Claude Opus 4.6, merging the previously separate GPT-5.2 reasoning LLM and GPT-5.2-Codex coding agent into one general-purpose agent. The new model nearly doubles the OSWorld-Verified score, sets fresh highs on SWE-Bench Pro and Terminal-Bench, and runs 25% faster thanks to inference-stack improvements. The piece highlights that GPT-5.3-Codex was used to actively monitor and debug its own training runs in an SRE-like role, going beyond predecessors that merely generated training data. Combined with the recently released Codex app for macOS, the model enables interactive, real-time collaboration without losing task context across long-running workflows.

OpenAI's GPT-5.3-codex vaulted to the top of the Microsoft Foundry AI Model Leaderboard shortly after its February 5, 2026 debut, ranking No. 1 on the platform's Quality rating, which averages accuracy across reasoning, knowledge, and coding benchmark datasets. GPT-5.3-codex appeared on Microsoft Foundry, an enterprise The article notes that enterprise developers using Microsoft Foundry may have to wait to access newer GPT-5.4 and its Codex variant, since GPT-5.3-codex was the current top offering on the Foundry platform at the time of reporting. Microsoft Foundry evaluates models on quality, speed, safety, cost, and throughput, prov

The OpenAI Help Center Model Release Notes page provides official first-party context for GPT-5.3-Codex and its place in the OpenAI catalog. The page records that GPT-5.3-Codex became available in the Responses API via OpenAI's changelog, confirming its distribution as a released OpenAI coding model, and it situates th The same Model Release Notes index documents the broader OpenAI model lifecycle relevant to GPT-5.3-Codex users, including the May 28, 2026 continuation of older-model retirements covering OpenAI o3 and GPT-4.5 in ChatGPT to make room for newer capable models, and a Model Spec update on August 18, 2026 adding clarity o

OpenAI's March 5, 2026 announcement for GPT-5.4 states that the new model "incorporates the industry-leading coding capabilities of GPT-5.3-Codex" while improving tool use, software environment handling, and professional tasks such as spreadsheets, presentations, and documents. This directly positions GPT-5.3-Codex as The same announcement notes that GPT-5.4 is the first general-purpose model released with native, state-of-the-art computer-use capabilities, supports up to 1M tokens of context for long-horizon agentic tasks, and adds tool search for more efficient tool selection; it also describes GPT-5.4 as OpenAI's most token-effic