82.7%. That’s GPT-5.5’s score on Terminal-Bench 2.0, a benchmark designed to test AI’s ability to independently plan, use tools, and execute command-line workflows in a sandboxed environment. OpenAI launched GPT-5.5 on April 23, 2026, positioning it as its most capable agentic AI model to date — one engineered from the ground up to reduce the need for human hand-holding across complex tasks. The score alone doesn’t tell the full story. What does is the fact that it’s 7.6 points ahead of GPT-5.4 and 13.3 points above Claude Opus 4.7. But here’s the twist: despite the leap in capability, OpenAI doubled the API price — to $5 per million input tokens and $30 per million output tokens. That’s not a typo. And it’s not a temporary rate. This is the new baseline.
Key Takeaways
- GPT-5.5 launched April 23, 2026, as OpenAI’s first retrained base model since GPT-4.5, co-designed with NVIDIA’s GB200 and GB300 NVL72 systems
- It leads on Terminal-Bench 2.0 (82.7%) and SWE-Bench Pro (58.6%), and hits 74.0% on MRCR v2 at one million tokens — up from 36.6%
- API pricing has doubled: $5/$30 per million input/output tokens, with GPT-5.5 Pro costing $30/$180
- No score on MCP Atlas, Scale AI’s tool-use benchmark, where Claude Opus 4.7 leads at 79.1%
- OpenAI claims 20% lower effective cost due to token efficiency, a claim validated by independent lab Artificial Analysis
Agentic by Design, Not Just Hype
OpenAI isn’t just calling GPT-5.5 “agentic” because the term is trending. This is the first base model retrained since GPT-4.5 with agent-like behavior baked into its training pipeline. The company says GPT-5.5 can now plan, use tools, validate its own outputs, and work through tasks with minimal human intervention. That’s different from previous models that required multiple prompts, manual corrections, and external orchestration layers to achieve similar results.
The model is rolling out to Plus, Pro, Business, and Enterprise users via ChatGPT and Codex, with API access going live on April 24. The timing suggests a deliberate push to embed GPT-5.5 into production workflows — especially those involving automation, code generation, and long-form task execution.
On SWE-Bench Pro, which tests the ability to resolve real GitHub issues, GPT-5.5 hits 58.6% — a meaningful jump from prior versions. More telling is Expert-SWE, OpenAI’s internal benchmark for tasks with a median human completion time of 20 hours. GPT-5.5 scores 73.1%, up from 68.5% on GPT-5.4. That improvement suggests the model isn’t just faster — it’s handling deeper, more complex reasoning chains.
Long-Context Breakthrough — But Only One Way
Where GPT-5.5 shines most dramatically is in long-context retrieval. On MRCR v2 at one million tokens — a benchmark that asks the model to find a specific answer buried in a massive document — GPT-5.5 scores 74.0%, more than double GPT-5.4’s 36.6%. That’s not incremental. It’s a leap.
For developers working with legal documents, technical manuals, or enterprise data archives, this kind of performance could eliminate entire preprocessing steps. No more chunking, no more metadata tagging just to make content searchable. The model can now navigate the full context and pull out precise answers — assuming the prompt is well-structured.
But this strength doesn’t generalize across all long-context tasks. OpenAI hasn’t released results for other reasoning-heavy benchmarks at this scale, and there’s no indication that GPT-5.5 maintains the same accuracy when summarizing or synthesizing across million-token spans. Retrieval is one thing. Understanding is another.
The NVIDIA Co-Design Edge
GPT-5.5 was co-designed with NVIDIA’s GB200 and GB300 NVL72 rack-scale systems. That’s not just marketing fluff. It means the model’s architecture, communication patterns, and memory usage were optimized in tandem with the hardware. The result? Lower latency, higher throughput, and better scaling during distributed inference — especially for tool-using agents that make multiple API calls in sequence.
This level of hardware integration used to be reserved for hyperscalers. Now it’s becoming table stakes for next-gen models. If you’re running agent workflows at scale, being on compatible infrastructure could mean the difference between usable latency and system collapse.
Benchmarks Tell Half the Story
OpenAI’s benchmark table is unusually transparent — and revealing. It includes a blank cell for MCP Atlas, Scale AI’s Model Context Protocol benchmark for tool use. No score for GPT-5.5. Meanwhile, Claude Opus 4.7 scores 79.1%, the highest on record. OpenAI included that row anyway. That’s either confidence or quiet concession.
The omission suggests GPT-5.5 may struggle with certain structured tool-calling patterns — perhaps those requiring strict schema adherence or multi-step API coordination outside OpenAI’s ecosystem. Or maybe the model simply wasn’t tested. Either way, it’s a red flag for teams building agents that interact with third-party SaaS tools.
On BrowseComp, OpenAI’s own agentic web-browsing benchmark, GPT-5.5 Pro leads at 90.1%. But internal benchmarks are inherently optimistic. The real test is whether it can consistently book travel, extract pricing tables from dynamic sites, or debug live frontend issues without falling into infinite loops.
- Terminal-Bench 2.0: GPT-5.5 — 82.7%, GPT-5.4 — 75.1%, Claude Opus 4.7 — 69.4%
- SWE-Bench Pro: GPT-5.5 — 58.6%, GPT-5.4 — prior version not specified
- MRCR v2 (1M tokens): GPT-5.5 — 74.0%, GPT-5.4 — 36.6%
- Expert-SWE: GPT-5.5 — 73.1%, GPT-5.4 — 68.5%
- MCP Atlas: Claude Opus 4.7 — 79.1%, GPT-5.5 — no score
Pricing Math That Demands Scrutiny
Here’s the part that makes engineers pause: API pricing has doubled. GPT-5.5 costs $5 per million input tokens and $30 per million output tokens. GPT-5.5 Pro, available to higher-tier users, costs $30 input and $180 output — six times more than standard output pricing for the previous model.
OpenAI argues the efficiency gains offset the increase. According to the company, GPT-5.5 completes the same Codex tasks with fewer tokens than GPT-5.4. That claim was validated by Artificial Analysis, an independent testing lab. Their tests showed a 20% higher effective cost after accounting for reduced task iterations and retries.
But that 20% premium only makes sense if your workflow actually converges faster. For exploratory coding, debugging, or iterative design, the model might save tokens. For simple summarization or translation? You’re paying more for no gain. At 10 million output tokens per month, GPT-5.5 costs $300 versus Claude Opus 4.7’s $250. That difference only pays off if fewer retries and shorter task chains outweigh the base rate hike — and that math varies by use case.
And let’s be clear: this isn’t just about cost per token. It’s about lock-in. The more you invest in agent workflows tuned to GPT-5.5’s behavior, the harder it becomes to switch — even if a competitor delivers better tool use or lower prices.
What This Means For You
If you’re building agent-based systems — especially for software engineering, data analysis, or enterprise automation — GPT-5.5’s gains in planning, tool use, and long-context retrieval are real and worth testing. The jump on MRCR v2 alone could justify a trial run for teams drowning in document processing. But don’t assume efficiency gains will cover the cost. Run your own benchmarks. Measure token usage, task completion rate, and iteration count — not just accuracy.
For API budget holders, the pricing shift means tighter cost controls. The $300/month tab at 10 million output tokens is no longer theoretical. Monitor usage closely, especially with GPT-5.5 Pro, where output tokens hit $180 per million. And consider fallback strategies: mixing models, caching results, or using smaller agents for simple steps. The era of cheap, high-volume prompting is over — at least for OpenAI’s flagship models.
So here’s the real question: if GPT-5.5 is truly agentic, why does OpenAI still need us to manually validate its tool-use gaps?
Sources: AI News, original report
