OpenAI AI Updates: April 24, 2026

1. GPT-5.5 Launches with State-of-the-Art Agentic Coding Scores and Double the Price

OpenAI. OpenAI released GPT-5.5, positioning it as an agentic model that autonomously switches between tools for coding, research, and data analysis. Headline numbers: 82.7% on Terminal-Bench 2.0 (versus Claude Opus 4.7’s 69.4%), 35.4% on FrontierMath Tier 4, MRCR v2 long-context jumping from 36.6% to 74.0%, and state-of-the-art on 14 benchmarks. It trails on SWE-Bench Pro (58.6% vs Claude’s 64.3%). The API price is $5 per million input tokens and $30 per million output — double GPT-5.4’s rates, with the Pro variant at $30/$180. Plus, Pro, Business, and Enterprise ChatGPT and Codex users get it immediately; API access is arriving “very soon.” Source

2. GPT-5.5 System Card Reports a 60% Hallucination Drop

OpenAI. Alongside the model release, OpenAI published the GPT-5.5 system card detailing safety evaluations, refusal behavior, and capability testing. The headline claim is a roughly 60% reduction in hallucination rate compared to the previous generation, plus token-efficiency gains that let GPT-5.5 match GPT-5.4’s latency while using fewer tokens to complete equivalent tasks. The card is the primary reference for teams evaluating the model against internal safety and reliability baselines. Source

3. Codex CLI 0.124.0 Adds Amazon Bedrock Support and Stable Hooks

OpenAI. The Codex CLI 0.124.0 release brings TUI reasoning controls, multi-environment management, stable hooks, and automatic approval reviews. Paired with 0.123.0 from earlier in the day, it also adds a full Amazon Bedrock provider with AWS profile support, enhanced MCP diagnostics, and improved realtime handoffs for background agents. The Bedrock integration is notable: Codex CLI can now be driven by models hosted in a customer’s own AWS account, which matters for teams who can’t or won’t send code to OpenAI’s public API. Source

4. OpenAI Open-Sources a Privacy Filter Model for Text Redaction

OpenAI. OpenAI released Privacy Filter, an open-source model purpose-built for detecting and redacting personal information in text. The move is unusual for OpenAI — a purpose-built open-weight release aimed squarely at a compliance workflow rather than at general capability — and gives teams an alternative to running regex rules or sending sensitive text through a frontier model just to strip identifiers. Source

5. OpenAI Launches a Trusted Access Program Giving Microsoft Its Most Capable Models for Cyber Defense

OpenAI. OpenAI’s new Trusted Access for Cyber program grants Microsoft access to its most capable models specifically for cybersecurity work, in exchange for Microsoft’s entire cybersecurity team helping protect OpenAI’s models, infrastructure, and mutual customers. The partnership funnels through Microsoft’s Secure Future Initiative and explicitly extends to open-source software protections. It is a notable reassertion of the OpenAI-Microsoft relationship after months of public friction and comes amid broader industry debate about how useful frontier models really are at finding vulnerabilities. Source