GPT-5.5 Cost Analysis: How Reduced Verbosity Softens the 2x Price Hike

While GPT-5.5 features a 100% increase in list price, OpenRouter's analysis shows actual user costs rose by only 49-92%. This is because the model generates significantly fewer completion tokens for prompts longer than 10K tokens. Users with shorter prompts, however, experience a higher cost increase as the model's verbosity remains high for those tasks.

Key Points

• OpenAI doubled GPT-5.5 prices to $5.00/M for input and $30/M for output tokens.
• Actual user costs increased by 49-92%, failing to reach the full 100% price hike due to model efficiency.
• GPT-5.5 is significantly less verbose for long prompts, generating up to 34% fewer tokens for prompts over 128K.
• Short prompts under 10K tokens see the highest cost impact because completion lengths did not decrease.
• The analysis utilized a switcher cohort methodology to ensure a direct comparison of the same user workflows across versions.

Sentiment

The community is moderately skeptical. While many acknowledge GPT-5.5 offers genuine improvements, particularly for agentic coding, there is widespread concern that the price increase is not justified by proportional gains. Commenters frequently reframe the discussion away from per-token savings toward task-completion economics, and several advocate for open-source alternatives as a more cost-effective path.

In Agreement

• GPT-5.5's reduced verbosity does translate to real cost savings compared to a naive 2x price increase, as OpenRouter's data shows
• GPT-5.5 on low reasoning matches GPT-5.4 on medium reasoning at lower cost, making it a practical upgrade path
• GPT-5.5 represents a genuine quality improvement for agentic coding and complex instruction following, with several independent benchmarks confirming this
• Per-token cost analysis is a useful starting point for understanding the real-world impact of pricing changes

Opposed

• Cost per token is misleading because it ignores multi-turn interactions and task completion efficiency — cost per completed engineering task is the metric that matters
• The analysis lacks methodological rigor: no sample size, no distribution data, no control for number of turns in agentic workflows, and no understanding of task boundaries
• LLM progress has plateaued and price increases reflect providers squeezing customers rather than delivering proportional value improvements
• Higher reasoning levels can actually produce worse code due to scope creep and over-engineering, undermining the value proposition of expensive frontier models
• Open-source models like GLM 5.1, Kimi K2.6, and Xiaomi now offer competitive quality at far lower cost, making frontier model pricing hard to justify
• Newer models are being overfitted for coding at the expense of general capabilities, with regressions observed in domains like NLP and linguistics