--- title: OpenAI announces GPT-5.5 | daily.dev description: GPT-5.5 brings modest accuracy improvements and speed gains over its predecessor, with 52.5% fewer claimed hallucinations, though the conciseness advantage... canonical: https://daily.dev/posts/openai-announces-gpt-5-5-wge1t8qia twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: OpenAI announces GPT-5.5 | daily.dev og:description: GPT-5.5 brings modest accuracy improvements and speed gains over its predecessor, with 52.5% fewer claimed hallucinations, though the conciseness advantage... og:url: https://daily.dev/posts/openai-announces-gpt-5-5-wge1t8qia og:image: https://api.daily.dev/og/posts/wGe1t8QIA.png og:image:alt: OpenAI announces GPT-5.5 og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # OpenAI announces GPT-5.5 **[Collections](https://daily.dev/sources/collections)** · 4 min read · 0 upvotes · 0 comments ## Summary GPT-5.5 brings modest accuracy improvements and speed gains over its predecessor, with 52.5% fewer claimed hallucinations, though the conciseness advantage didn't hold up in testing. The more significant development is GPT-5.5-Cyber, a restricted variant for critical infrastructure defenders capable of pentesting, exploit development, and malware analysis — the second model to complete the UK AI Security Institute's multi-step cyber-attack simulation end to end. Benchmark data reveals alarming trends: frontier AI completes six times more attack steps than 18 months ago, multi-turn attack strategies expose far worse safety performance than single-prompt benchmarks suggest (GPT-5.4 jumped from 2.74% to 24.68% attack success rate), and a cheaper model with more scaffolding can match frontier models on vulnerability finding. Leaked model labels suggest GPT-5.6 and Claude Opus 4.8 are in development. The broader shift is toward AI agents needing governance as security principals with identity and access controls, not just tools. ## Content ## What GPT-5.5 actually is OpenAI has been incrementing model numbers in a fairly straightforward way: each step from GPT-5.0 through 5.5 brings capability improvements and better token efficiency, which translates to speed. GPT-5.5 is currently their best model, and GPT-5.5 Instant is now the default in ChatGPT, with OpenAI claiming 52.5% fewer hallucinations than its predecessor. In hands-on testing, the accuracy improvement is real. GPT-5.5 correctly answered a question about Claude Sonnet's context window where GPT-5.2 confidently hallucinated the answer. Personalization also improved slightly, with the model surfacing more behavioral patterns from memory. The one claim that didn't hold up: conciseness. GPT-5.2 was actually more concise, though GPT-5.5 is more conversational. The difference between the two models is modest enough that you'd probably miss it without a direct side-by-side comparison. For everyday use, the speed gains are the most noticeable thing. Several developers have noted they're reaching for ChatGPT more often now for quick queries, math, and plots. ## The cybersecurity angle is more interesting The more consequential story is GPT-5.5-Cyber, a restricted variant OpenAI is rolling out to a select group of critical infrastructure defenders. It's capable of pentesting, bug finding, exploit development, and malware analysis. The UK's AI Security Institute confirmed it's the second model ever to complete one of their multi-step cyber-attack simulations end to end. On expert cyber tasks, GPT-5.5 scores 71.4% versus Claude Mythos Preview's 68.6%. One benchmark had GPT-5.5 solving a challenge in 10 minutes and 22 seconds at a cost of $1.73 that took a human expert 12 hours. There's an awkward footnote here: Sam Altman had recently criticized Anthropic for restricting Mythos to select defenders, calling it "fear-based marketing." OpenAI then did essentially the same thing with GPT-5.5-Cyber. ## How this fits into the broader AI security picture A few things are worth holding together: **The threat is real but measured.** A month after Mythos launched, security practitioners say the hacking threat is more measured than initial fears suggested. Mythos can find vulnerabilities faster and with simpler prompts, but the real bottleneck is organizations' ability to validate and prioritize the flood of discovered flaws, not the discovery itself. Infrastructure and compute demands currently limit broad access, though that's expected to change quickly. **The numbers are still alarming in aggregate.** NCSC data shows frontier AI completing six times more attack steps than 18 months ago at roughly £65 per attempt. Check Point Research documented a single operator using Claude Code and GPT-4.1 in a dual-AI workflow to breach nine Mexican government agencies. Anthropic's Mythos found 181 Firefox exploits versus 2 for its predecessor. OpenAI's Codex Security has been credited with 14 CVEs. **Single-turn safety benchmarks are misleading.** Cisco tested 15 frontier models under multi-turn attack strategies and found the results significantly worse than single-prompt benchmarks suggest. Claude Opus 4.6 went from a 3.64% attack success rate in single-turn scenarios to 16.20% in multi-turn. GPT-5.4 went from 2.74% to 24.68%. Gemini 3 Pro went from 18.10% to 73.35%. The industry is measuring a narrow slice of real attacker behavior. **A smaller model with more scaffolding can match a frontier model.** A separate analysis found that a cheaper model with more prompting scaffolding achieved similar vulnerability-finding results to GPT-5.5, which raises uncomfortable questions about how quickly offensive capabilities democratize. ## What's coming Leaked signals point to a busy few months. GPT-5.6 labels have appeared in Codex routing logs alongside internal model tags (Iris Alpha, Ember Alpha, Beacon Alpha), though nothing is confirmed. On the Anthropic side, Claude Opus 4.8 has been spotted on Google Vertex. Gemini 3.5 Pro is also rumored. OpenAI separately confirmed an internal reasoning model solved an 80-year-old Erdős math conjecture, which hints at stronger reasoning capabilities that could eventually reach Codex. ## The underlying shift The competitive edge in AI is moving away from raw model quality toward systems, distribution, connectors, and security infrastructure. Anthropic released 10 finance agent templates with full production stacks. OpenAI partnered with PwC on similar finance agents. Claude is now integrated into Microsoft 365 apps. For security teams specifically, the practical implication is that AI agents need to be governed as security principals with their own identity and access controls, not just as tools. Static detection is losing ground to adaptive approaches. And the baseline threat model is shifting toward continuous low-level compromise rather than discrete incidents. ## Similar posts on daily.dev - [Frontier AI models offer sneak peak of seismic cyber shifts ahead](https://daily.dev/posts/frontier-ai-models-offer-sneak-peak-of-seismic-cyber-shifts-ahead-rxzsn25gn) · CSO Online · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#openai](https://daily.dev/tags/openai), [#ai-agents](https://daily.dev/tags/ai-agents), [#gpt](https://daily.dev/tags/gpt) [View this post on daily.dev](https://daily.dev/posts/openai-announces-gpt-5-5-wge1t8qia) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"OpenAI announces GPT-5.5","url":"https://daily.dev/posts/openai-announces-gpt-5-5-wge1t8qia","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/openai-announces-gpt-5-5-wge1t8qia"},"datePublished":"2026-04-30T14:01:45.916Z","dateModified":"2026-05-31T09:37:30.485Z","description":"GPT-5.5 brings modest accuracy improvements and speed gains over its predecessor, with 52.5% fewer claimed hallucinations, though the conciseness advantage...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/openai-announces-gpt-5-5-wge1t8qia","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,openai,ai-agents,gpt","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"OpenAI announces GPT-5.5"}]} ```