What it is
GPT-5.5 is OpenAI's frontier large language model released in April 2026, succeeding GPT-5.4 and preceding GPT-5.6. It targets complex professional work and compute-intensive agentic tasks, shipping with a one-million-token context window, image input, and a native suite of agentic tools — computer use, hosted shell, Model Context Protocol (MCP), web search, and Skills — accessible directly through the Chat Completions and Responses API. A Pro variant (GPT-5.5 Pro) is positioned for the heaviest workloads.
Release and safety artifacts
The model launched on April 24, 2026, accompanied by a system card published the day before — a formal safety and capability disclosure document that has become a standard release artifact in the GPT-5 series. OpenAI subsequently launched a Bio Bug Bounty program specifically targeting GPT-5.5, inviting external researchers to probe biosecurity-related vulnerabilities. This structured external red-teaming, focused on a high-stakes domain, reflects a deliberate investment in post-deployment safety evaluation beyond internal testing.
Behavioral and API changes from GPT-5.4
Two behavioral shifts distinguish GPT-5.5 from its predecessor. First, reasoning effort defaults to medium rather than maximum, giving practitioners explicit control over the cost-latency-quality tradeoff without requiring per-call overrides. Second, prompt caching is extended-only, a narrowing from prior GPT-5 series behavior that affects how long-context workflows should be structured. Both changes are consequential for production integrations built on earlier assumptions.
Capability profile and benchmark position
The events bundle does not disclose internal architecture. Externally, GPT-5.5 occupies the upper-frontier tier: Anthropic cited it as a peer to Claude Opus 4.8 and Kimi K2.7 when contesting the severity of the Claude Fable 5 jailbreak incident, and the Artificial Analysis Intelligence Index placed it at 55 — above GLM-5.2 (51, open-weights) and just below Claude Opus 4.8 (56) at the time of those measurements.
Known gaps are equally important for practitioners to track:
- SWE-Interact: GPT-5.5 solves roughly 50% of single-turn coding baseline tasks but only ~25% of multi-turn interactive tasks where requirements are revealed progressively. Failure modes include over-agentic coding, requirement forgetting, and early abandonment under ambiguity.
- ActiveVision: At maximum reasoning effort, GPT-5.5 scores 10.6% on active visual observation tasks (zeroing on 11 of 17 sub-tasks) versus a 96.1% human baseline. The benchmark tests iterative attention redirection based on intermediate hypotheses — a closed-loop visual cognition capability the model structurally lacks.
Agentic toolset and the Responses API
GPT-5.5 is the first GPT-5 series model to ship computer use and a hosted shell as first-class built-ins rather than external integrations. Combined with MCP and web search, this makes it possible to build agents that browse, execute code, and interact with external services without a separate orchestration layer. The Responses API, which also carries multi-agent orchestration in beta (fully released in GPT-5.6), is the intended surface for these workflows.
GPT-Live-1: GPT-5.5 as a reasoning backend
On July 8, 2026, OpenAI released GPT-Live-1 and GPT-Live-1 mini, replacing Advanced Voice Mode with a full-duplex voice system. The architecture separates a real-time conversational voice model from a reasoning tier: when deeper analysis is needed, the voice model delegates to GPT-5.5 or GPT-5.5 Thinking in the background while continuing to speak. User-selectable effort levels (Instant, Medium, High) route to the appropriate GPT-5.5 variant. The result is a substantial performance jump — GPQA rose from 45.3% (prior AVM) to 84.2% at high reasoning, and BrowseComp improved from 0.7% to 75.2%. Human raters preferred GPT-Live-1 in 75.7% of comparisons. This deployment pattern — a lightweight real-time model delegating to a heavier reasoning model — is a notable architectural template for latency-sensitive agentic applications.
Competitive context
GPT-5.5 entered a market where Claude Opus 4.8 (Anthropic) and GLM-5.2 (Z.ai, open-weights, 753B MoE) were the primary alternatives at comparable capability tiers. Sakana AI's Fugu-Ultra orchestrator — which routes across Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 — achieved state-of-the-art results on several benchmarks by combining all three, suggesting that GPT-5.5's strengths are complementary to rather than dominant over its peers. Cognition's Devin Fusion multi-model routing system also claimed a 35% cost reduction versus GPT-5.5 and Opus 4.8 by dynamically selecting between them.
Supersession by GPT-5.6
GPT-5.6 (Sol, Terra, Luna tiers) launched on July 9, 2026, and immediately became the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. GPT-5.6 Sol Ultra scores 91.9% on TerminalBench 2.1 and is priced at roughly half of Anthropic's comparable tier. GPT-5.5 remains active as the GPT-Live-1 reasoning backend and in existing API integrations, but it is no longer OpenAI's flagship.
Regulatory backdrop
GPT-5.5's release window coincided with the first major government interventions in frontier model access. The U.S. government suspended Anthropic's Claude Fable 5 and Mythos 5 in June 2026 — citing a jailbreak that Anthropic argued produced results already achievable by GPT-5.5 and other publicly available models. GPT-5.6's launch was itself preceded by a government-mandated capability preview, and early access was restricted to vetted partners at the Trump administration's request. This regulatory environment — in which Washington exerts direct control over frontier model releases — is the governance context in which GPT-5.5 operated and GPT-5.6 launched.
