GPT-5.5 launch

The most consequential fact about GPT-5.5 is not the model itself, but the calendar. On April 16, Anthropic shipped Claude Opus 4.7 and took the SWE-bench Pro coding crown at 64.3%. Exactly seven days later, OpenAI pushed out GPT-5.5 — codenamed 'Spud' internally and positioned as the first fully retrained base model since GPT-4.5. The last point release, GPT-5.4, was only six weeks old. That cadence is the telltale: OpenAI is no longer pacing itself against product cycles, it is pacing itself against Anthropic's release calendar.

The competitive context explains the urgency. Reports of a 'Code Red' posture inside OpenAI have circulated since December 2025, and Anthropic's ARR reportedly jumped from roughly $9B to $30B during that window, pulled upward by developer spend on Claude for agentic coding. A seven-day turnaround from a rival's flagship to your own retrained base model is only possible if the new model was already in final staging — meaning the decision OpenAI is optimizing for is not 'when is it ready' but 'when does it ship relative to Opus.' The launch, in other words, is a response, not a plan.

Read the headline coverage and GPT-5.5 is 'the smartest model yet.' Read the benchmark table and the story is more careful. On Terminal-Bench 2.0 — the agentic coding benchmark OpenAI leaned on hardest — GPT-5.5 posts 82.7%, edging Anthropic's gated Claude Mythos Preview at 82.0%. That is a 0.7-point margin. VentureBeat's write-up captured the framing tension precisely when it wrote that GPT-5.5 'narrowly beats' Mythos. On BrowseComp, GPT-5.5 Pro genuinely leads at 90.1% versus Opus 4.7 at 79.3%, and OSWorld-Verified (78.7%) supports the computer-use narrative.

But the benchmark the developer community treats as gospel — SWE-bench Pro — tells the opposite story. GPT-5.5 scores 58.6%; Opus 4.7 scores 64.3%; gated Mythos posts 77.8%. On GPQA Diamond, GPT-5.5 (93.6%) trails Opus 4.7 (94.2%) and Mythos (94.5%). The pattern is consistent: OpenAI wins the agentic/terminal/browsing suite, Anthropic still wins code-repository-grade engineering tasks. The 60% reported hallucination drop and the GDPval result — matching or beating industry professionals in 84.9% of comparisons — are real gains, but they do not change the coding ranking. The 'smartest model' claim is carrying a lot of weight that the SWE-bench Pro number does not support.

GPT-5.5's API list price is $5 per million input tokens and $30 per million output tokens — double GPT-5.4's $2.50/$15. GPT-5.5 Pro lands at $30/$180. On the surface, this reads as a simple premium. The pitch OpenAI and NVIDIA are making is that the per-query cost actually falls once you account for the stack underneath. GPT-5.5 is trained and served on NVIDIA GB200/GB300 NVL72 systems, which NVIDIA claims deliver '35x lower cost per million tokens and 50x higher token output per second per megawatt' versus the previous generation. Codex itself wrote load-balancing heuristics during development that delivered a reported 20%+ token-generation speedup. Greg Brockman frames the net as 'a faster, sharper thinker for fewer tokens.'

The tell that this is a vertical play, not a list-price bump, is the NVIDIA Codex pilot. NVIDIA rolled GPT-5.5-powered Codex to over 10,000 employees across engineering, legal, and marketing on launch day. An anonymous NVIDIA engineer told the company blog that 'debugging cycles that once stretched across days are closing in hours.' That is the economic story OpenAI needs the doubled price to ride on: if a senior-engineer debugging cycle collapses from three days to three hours, the per-token cost is not the budget line that matters — the seat count is. For buyers who can't extract that productivity delta, the doubled API price is just a doubled API price.

GPT-5.5 is not positioned as a better ChatGPT. OpenAI's own launch language calls it 'a new class of intelligence for real work and powering agents, built to understand complex goals, use tools, check its work, and carry more tasks through to completion.' The feature set follows: computer use, 400K context in Codex at launch with 1M coming to API, long-horizon tool-calling, and scientific-research workflows that Mark Chen says show 'meaningful gains.' The bundled rollout across ChatGPT, Codex, and a forthcoming AI browser makes it explicit — OpenAI is trying to collapse the chat product, the coding agent, and the browsing agent into one runtime.

Developer YouTube has reacted accordingly: the top-performing review coverage frames GPT-5.5 as a step-change for computer-use and agentic workloads rather than as a chat-quality bump, and the benchmark-comparison framing versus Opus 4.7 and Gemini 3.1 Pro dominates the conversation. On X, sentiment skews mixed-to-positive, with official accounts and AI-leaker voices leaning into the agent narrative while parts of the dev community remain skeptical — the tension is not about capability, it is about whether the cadence (5.3 → 5.4 → 5.4-mini/nano → 5.5 in under two months) is a sign of momentum or of model fatigue. The quiet second-order question: if the 'smartest model' is now the agent runtime, the competitive surface stops being 'best benchmark' and becomes 'best integration' — and that is a product fight, not a model fight.

A detail that got little headline coverage is the most policy-relevant one. On CyberGym, GPT-5.5 scores 81.8% versus Opus 4.7 at 73.1% and Mythos at 83%. Anthropic has already used cyber-exploitation risk as the explicit reason it gated Claude Mythos Preview rather than releasing it broadly. OpenAI published a System Card alongside GPT-5.5 describing targeted red-teaming, and the Bank of England publicly voiced cyber-risk concerns in the days around launch.

The structural question the industry has not resolved: two of the three frontier labs now concede that their top coding model is also a top offensive-security tool, and they are making opposite release decisions about it. Anthropic gates. OpenAI ships to Plus, Pro, Business, Enterprise, Edu, and Go. If that divergence holds, the 'frontier' becomes less a capability line and more a release-policy line — and regulators, not researchers, start deciding where it sits.