GPT-5.5 "Spud" Doubles Its Price — And Bets Agents Are Worth It

OpenAI launched GPT-5.5 yesterday at $5/$30 per million input/output tokens — exactly double GPT-5.4's $2.50/$15 pricing and the first unambiguous frontier price hike in over a year. The model, internally codenamed "Spud" and announced by Greg Brockman as "two years of research," hits 88.7% on SWE-Bench and 92.4% on MMLU at launch. That SWE-Bench number is a genuine jump. But the pricing signal is the more significant data point.

OpenAI's positioning is explicit: GPT-5.5 targets agentic coding, computer use, multi-step autonomous workflows, and early scientific research. Brockman called it "a big step towards more agentic and intuitive computing." The $30/$180 Pro tier makes the intent clearer — this is infrastructure pricing for workloads that run long chains of actions, not a general-purpose assistant upgrade. The commodity tier of the frontier still exists; GPT-5.4 isn't going away.

The counterargument deserves honest treatment. Agent loops are genuinely token-heavy — a single autonomous coding session can consume an order of magnitude more tokens than a chat exchange. If GPT-5.5 is actually more expensive to serve because of longer context retention, multi-modal computer use, and tighter safety guardrails (API access was delayed pending additional cybersecurity review), then $5/$30 might reflect real inference economics rather than a margin grab. And if the 88.7% SWE-Bench score translates to meaningfully fewer failed iterations in production agent pipelines, a 2× price against a 2× productivity gain is cost-neutral per task completed.

The split that's emerging is harder to dismiss. Claude Opus 4.7 currently leads LMArena at 1503 Elo and ships at $5/$25 — matching GPT-5.5 on input, undercutting it on output by $5 per million tokens, without a price hike. Anthropic has spent April moving in one direction on cost. OpenAI just moved in the other. Both strategies can be rational simultaneously: Anthropic is optimizing for developer adoption volume; OpenAI is betting that enterprise buyers running autonomous agent fleets will treat per-token cost as secondary to capability and reliability at scale.

The practical consequence is that the next pricing announcements from Google and Anthropic won't be read in isolation. If Gemini 2.5 Ultra or a future Opus release lands at $6+ input, it will be seen as validating OpenAI's premium tier rather than a unilateral hike. Teams budgeting agentic workloads in 2026 are now watching two frontier providers diverge on the same capability class — and the model they standardize on this quarter may be the one that dictates their infrastructure costs for the next two years.