GPT-5.5 Takes AA’s Launch Lead at Double GPT-5.4’s Token Price
GPT-5.5 launched April 23, with API access added April 24 at $5/$30 per million tokens. Artificial Analysis reported an xhigh Intelligence Index lead, but roughly 40% lower token use limited its measured cost increase to about 20%.
GPT-5.5 launched in paid ChatGPT and Codex on April 23, with API access added April 24. OpenAI listed the standard API model at $5 per million input tokens and $30 per million output tokens, double GPT-5.4’s launch token prices.
The price comparison alone misses token efficiency. In its April 23 evaluation, Artificial Analysis reported that GPT-5.5 at xhigh led its Intelligence Index by three points. It also reported roughly 40% fewer tokens than GPT-5.4 across that suite, producing an approximately 20% increase in the total cost of running the index despite doubled unit prices.
Claims vs checks
OpenAI positioned GPT-5.5 for coding, research, data analysis, documents and computer operation. Its reported 82.7% Terminal-Bench 2.0 result is a vendor evaluation, distinct from Artificial Analysis’s Terminal-Bench Hard tests. OpenAI says its evaluations use xhigh effort in a research environment.
Artificial Analysis’s launch review reported leads in Terminal-Bench Hard, GDPval-AA and APEX-Agents-AA. Its AA-Omniscience findings were mixed: stronger knowledge accuracy alongside a higher hallucination rate than the cited Claude and Gemini configurations. A composite lead does not establish superiority on every task.
Availability
The launch announcement named Plus, Pro, Business and Enterprise for ChatGPT and Codex rollout. GPT-5.5 Pro was offered in ChatGPT to Pro, Business and Enterprise. The April 24 update separately confirmed both API variants.
Limits
These are launch-window prices and results. Effort settings, test harnesses, and total token use matter; the ranking should not be read as a permanent hierarchy or a measurement of default ChatGPT behavior.
Correction (October 14, 2026): Removed unsupported base-retraining conclusions and retrospective Arena figures; distinguished OpenAI’s Terminal-Bench 2.0 result from AA’s Terminal-Bench Hard.