Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched (GPT-5.3-Codex, 5 Feb 2026): shipped in Codex surfaces first, API followed later in Feb (Kilo post of 25 Feb), so no launch list price (current page $1.75/$14) and price is null. 400k context, text+image input (current OpenAI model page). Maker figures: Terminal-Bench 2.0 77.3% (Opus 4.6 65.4%), SWE-Bench Pro 56.8%, OSWorld-Verified 64.7%; AA's later Terminal-Bench Hard reading was 53% (via AA's GPT-5.4 post). None is Terminal-Bench 4.0, so coding is null. No launch-time AA Intelligence Index entry opened (AA marks it Estimated): reasoning and speed null.[2]
Codex, on the other hand, actually feels good in that autonomous workflow. It is more predictable, easier to steer mid-task, and more trustworthy when the job spans multiple files and decisions.
GPT-5.3-Codex surpasses its predecessor as well as the newly released Claude Opus 4.6 (65.4 percent) by about twelve percent in the agentic coding benchmark Terminal-Bench 2.0
Compared to prior Codex models, it's more token-efficient and approximately 25% faster.
Terminal-Bench 2.0 jumped from 64.0% for GPT-5.2-Codex to 77.3% for 5.3.
| GPT-5.3-Codex: OpenAI introduces new coding model · 2026年2月6日 | 0.70 |
|---|---|
| AI Coding Agent Showdown: GPT-5.3-Codex vs Claude Opus 4.6 · 2026年2月7日 | 0.70 |
| GPT-5.3-Codex is Live in Kilo · 2026年2月25日 | 0.50 |
| GPT-5.3-Codex: A Late Review (And Why I'm Paying Attention Now) · 2026年5月24日 | 0.70 |
这个标记会跟进新闻:它显示目前已知的情况,这里列出改了什么以及起因。
已编辑由 @AIDesk
读者上传的照片,与图钉带来源的媒体分开显示。你可以为吃到或看到的内容评分,在餐厅图钉上还能直接在照片上标记菜品。
第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
OpenAI's[1][3] API model page for GPT-5.3-Codex lists low/medium/high/xhigh reasoning effort, a 400,000 context window, 128,000 max output tokens, Aug 31, 2025 knowledge cutoff and $1.75/$14 per 1M tokens.
OpenAI's[1][2] announcement post, linked from the release note entry.
Independent report giving the SWE-bench Pro (Public) 56.8%, Terminal-Bench 2.0 64.0% to 77.3% and OSWorld-Verified 64.7% (previous generation 38.2%) figures and carrying the three launch benchmark charts used on this pin.[1]
Independent launch report whose screenshots of OpenAI's[1][2][3] charts show Terminal-Bench 2.0 (GPT-5.3-Codex 77.3%, GPT-5.2-Codex 64.0%, GPT-5.2 62.2%) and OSWorld-Verified (64.7%, 38.2%, 37.9%), and the racing-game demo used as a picture here.
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: GPT-5.3-Codex Released