GPT-5.5 is OpenAI's most capable model as of April 2026, released on April 23, 2026. OpenAI President Greg Brockman described it as "a new class of intelligence" — specifically designed for complex, multi-step agentic work where the model plans, uses tools, checks its own output, and keeps going until a task is done.
What It Does Well
GPT-5.5 is strongest in agentic coding, computer use, knowledge work, and early scientific research. OpenAI emphasizes its ability to handle underspecified problems — given a messy, multi-part task, it plans rather than asks, navigates ambiguity, and self-corrects. It also matches GPT-5.4's per-token latency despite being significantly more capable, and uses fewer tokens to complete the same Codex tasks.
Core strengths:
- Writing and debugging code
- Online research with tool use
- Data analysis
- Creating documents and spreadsheets
- Operating software across tools
- Reasoning across long context windows
Model Variants
- GPT-5.5 — available to all paid ChatGPT users (Plus, Pro, Business, Enterprise) and in Codex
- GPT-5.5 Pro — extended reasoning variant; rolling out to Pro, Business, and Enterprise users in ChatGPT only; uses parallel test-time compute
- GPT-5.5 Thinking — available to all paying users; dedicated reasoning mode
API availability arrived April 24, 2026, one day after the ChatGPT rollout, as gpt-5.5 in the Responses API and Chat Completions API. Pricing is higher than previous models.
Context in the GPT-5.x Series
GPT-5.5 follows a rapid release cadence: GPT-5 (August 7, 2025) → GPT-5.1 → GPT-5.2 (December 11, 2025) → GPT-5.3-Codex (February 5, 2026) → GPT-5.4 (March 5, 2026) → GPT-5.5 (April 23, 2026). This is the first release to explicitly follow Anthropic's Opus 4.7, which OpenAI directly benchmarks against — GPT-5.5 outperforms Opus 4.7 on most standard benchmarks, though GPT-5.4 Pro still outperforms default GPT-5.5 on some tasks.
Safety and Cybersecurity
OpenAI introduced stricter cybersecurity classifiers with this release, citing GPT-5.5's meaningful uplift in cyber capability. On the CyberGym benchmark, GPT-5.5 scores 81.8% (vs. Anthropic's Mythos model at 83.1%). The company evaluated the model across its full Preparedness Framework, conducted targeted red-teaming for advanced cybersecurity and biology capabilities, and worked with nearly 200 early-access partners before release. OpenAI's stated approach is "trusted access plus robust safeguards that scale with capability" rather than restricting access.
Key Benchmarks
Specific benchmark numbers for GPT-5.5 are not yet comprehensively published as of this writing — OpenAI's system card notes it outperforms GPT-5.4 on agentic tasks and Opus 4.7 on most standard benchmarks.
The framing here is deliberate: GPT-5.5 is positioned as a step toward computers that run themselves, not just respond to prompts. The "less guidance needed" emphasis is the real signal — OpenAI is building toward ambient agents, and GPT-5.5 is the first model they're explicitly pitching that way.
The benchmark race with Anthropic is now explicit. OpenAI directly comparing to Opus 4.7 in their press materials marks a shift from ignoring competitors to naming them. For the MythOS audience (Obsidian + Claude users), this matters as competitive context — the models powering agentic workflows are now converging, and the differentiation is increasingly at the systems layer, not the model layer.
The cybersecurity angle is worth tracking. The Mythos model controversy (Anthropic limiting its rollout due to cyber capabilities) appears to have accelerated OpenAI's cybersecurity narrative. They're framing advanced cyber capability as a feature to be deployed defensively, not a liability to be suppressed.
GPT-5.5 Pro's parallel test-time compute architecture is the one to watch technically — it's the same approach driving the highest-ceiling performance on reasoning-heavy tasks.