82.7%. That’s GPT-5.5’s score on Terminal-Bench 2.0, a benchmark designed to test an AI’s ability to plan, execute, and coordinate tools in a sandboxed command-line environment. On April 23, 2026, OpenAI launched GPT-5.5 with a sharp focus on agentic capabilities, calling it “a new class of intelligence for real work and powering agents.” The number isn’t just a benchmark win—it’s the foundation of a strategic pivot: this model isn’t meant to respond. It’s meant to act.
Key Takeaways
- GPT-5.5, launched April 23, 2026, is OpenAI’s first retrained base model since GPT-4.5 and is optimized for agentic workflows.
- It scores 82.7% on Terminal-Bench 2.0, outperforming GPT-5.4 (75.1%) and Claude Opus 4.7 (69.4%).
- API pricing is twice that of GPT-5.4: $5 per million input tokens, $30 per million output tokens.
- Despite higher rates, OpenAI claims 20% higher effective efficiency due to fewer task iterations, a claim validated by Artificial Analysis.
- The model shows major gains in long-context reasoning: 74.0% on MRCR v2 (1M tokens), up from 36.6%.
Agentic by Design, Not by Prompt
OpenAI didn’t just upgrade GPT-5.5. They rebuilt it. This is the first retrained base model since GPT-4.5, co-designed with NVIDIA’s GB200 and GB300 NVL72 rack-scale systems. That co-design detail matters—not because of hardware synergy alone, but because it signals a fundamental shift in how OpenAI is thinking about AI deployment. GPT-5.5 isn’t a chatbot with better recall. It’s a system trained from the ground up to plan, use tools, verify its own outputs, and complete multi-step tasks without constant human intervention.
The difference is operational. Before, developers had to chain prompts, inject feedback loops, and babysit workflows as models stumbled through complex tasks. Now, OpenAI says, you can hand off a goal—like “debug this CI/CD pipeline” or “build a Python module that scrapes and summarizes daily news”—and the model will run it down. That’s not a minor improvement. That’s the shift from AI as assistant to AI as executor.
And the benchmarks back the claim. On SWE-Bench Pro, which tests GitHub issue resolution, GPT-5.5 reaches 58.6%. That may not sound high, but it represents a meaningful jump in single-pass success. More importantly, OpenAI introduced Expert-SWE, an internal benchmark where tasks carry a median estimated human completion time of 20 hours. GPT-5.5 scores 73.1% there—up from 68.5% for GPT-5.4. That’s not just better code generation. That’s better sustained reasoning.
Long Context, Real Gains
The MRCR v2 benchmark result is perhaps the most underdiscussed but significant part of the release. At one million tokens, GPT-5.5 scores 74.0% in retrieving a specific answer buried in a massive document. GPT-5.4? 36.6%. That’s not a marginal improvement. It’s a near doubling in effectiveness.
Why does this matter? Because real-world workflows—legal discovery, scientific literature review, compliance auditing—don’t happen in 32k-token snippets. They require navigating sprawling, complex documents. A model that can actually find a needle in a million-token haystack isn’t just smarter. It’s more useful. And for developers building document-heavy agents, this is the difference between a prototype and a production tool.
But let’s be clear: this isn’t magic. The model still fails nearly a quarter of the time. And retrieval isn’t understanding. But the direction is undeniable. GPT-5.5 isn’t just longer-context capable. It’s context-aware in a way its predecessors weren’t.
Benchmarks Tell a Mixed Story
OpenAI released a detailed benchmark table. And it includes a conspicuous absence: no score for GPT-5.5 on MCP Atlas, Scale AI’s Model Context Protocol tool-use benchmark. That slot is blank. Meanwhile, Claude Opus 4.7 posts a 79.1%—higher than GPT-5.5’s 82.7% on Terminal-Bench, though different benchmarks aren’t directly comparable.
What’s interesting isn’t just the omission. It’s that OpenAI included it. They didn’t hide the gap. They framed it. That’s either confidence or desperation, depending on your view. But the message is clear: OpenAI knows tool use is a competitive battlefield. And they’re not claiming dominance across all dimensions.
Still, Terminal-Bench 2.0 is their turf. It tests command-line workflows that require planning, tool coordination, and error recovery—all in a sandbox. GPT-5.5’s 82.7% isn’t just a number. It’s a statement: we’ve built something that operates more like a developer than a chatbot.
Where OpenAI’s Edge Matters
- Terminal-Bench 2.0: GPT-5.5 scores 82.7% (GPT-5.4: 75.1%, Claude Opus 4.7: 69.4%)
- SWE-Bench Pro: 58.6% (improved single-pass resolution)
- Expert-SWE: 73.1% (tasks with 20-hour human median effort)
- MRCR v2 (1M tokens): 74.0% (GPT-5.4: 36.6%)
- MCP Atlas: No score (Claude Opus 4.7: 79.1%)
Double the Price, But Is It Worth It?
Here’s the thing that’s going to sting for builders: GPT-5.5 API pricing is exactly twice that of GPT-5.4. Input tokens: $5 per million. Output tokens: $30 per million. For heavy users, that’s an immediate 100% cost increase on paper. GPT-5.5 Pro—available to Pro, Business, and Enterprise tiers—costs $30 per million input and $180 per million output. That’s enterprise-grade pricing, no debate.
OpenAI’s defense? Efficiency. They claim GPT-5.5 completes the same Codex tasks with fewer tokens and fewer retries. And according to Artificial Analysis, an independent testing lab, that checks out. Their validation shows effective costs are only about 20% higher once you factor in reduced iteration.
But that 20% isn’t guaranteed. It depends on your workload. For a developer running a simple query engine, the cost jump might not pay off. But for an agent orchestrating a multi-step debugging session? Fewer retries mean faster resolution, less latency, and lower operational friction. The math only works if your use case actually benefits from agentic behavior. If you’re still doing one-off prompts, you’re overpaying.
And let’s talk numbers. At 10 million output tokens per month, GPT-5.5 standard costs $300. Claude Opus 4.7 costs $250. That 20% premium only makes sense if GPT-5.5’s superior agentic performance reduces task overhead enough to justify it. For some teams, it will. For others, it won’t.
GPT-5.5 Pro and the Parallel Compute Edge
GPT-5.5 Pro isn’t just more expensive. It’s architecturally different. OpenAI applies additional parallel test-time compute on harder problems. The result? It leads OpenAI’s BrowseComp benchmark—a measure of agentic web browsing—at 90.1%. That’s not just better. It’s best-in-class for publicly available models.
What does “parallel test-time compute” mean in practice? It means the model can explore multiple reasoning paths simultaneously when stuck. Think of it as thinking out loud—but in parallel universes. When faced with a complex web automation task, it doesn’t just try one approach. It simulates several, evaluates outcomes, and picks the most promising. That’s expensive. That’s why it’s locked behind Pro, Business, and Enterprise tiers.
But for teams building AI agents that need to navigate dynamic websites, extract data, or perform research, this is a game-changer. The jump from 80% to 90% in browsing accuracy isn’t linear. It’s the difference between an agent that fails on JavaScript-heavy pages and one that just… works.
What This Means For You
If you’re a developer building agents, GPT-5.5 is worth testing—fast. The gains in tool use, planning, and long-context retrieval aren’t incremental. They’re operational. But don’t assume efficiency gains will offset costs automatically. Run your actual workloads. Measure token usage, retry rates, and end-to-end task completion. The 20% effective cost increase is only a win if your agents finish faster and with less human oversight.
For startups and small teams, the pricing jump is a real constraint. At $30 per million output tokens, experimentation gets expensive. Consider using GPT-5.5 selectively—reserve it for complex, multi-step tasks—and stick with cheaper models for simpler work. And watch the MCP Atlas gap. If your agents rely on third-party tool integration, Claude Opus 4.7 might still be the better choice, at least for now.
OpenAI didn’t just release a new model. They released a statement: the future is agentic, and we’re pricing it accordingly. But adoption won’t be about benchmarks. It’ll be about whether that agentic promise actually reduces friction in real systems. The model is capable. The question is whether it’s worth it.
Sources: AI News, Artificial Analysis
