GPT-5.5's Hidden Cost Surge: What OpenRouter Found

When OpenAI launched GPT-5.5 in April, the headline numbers looked manageable. List price: $12.50 per million input tokens, $50 per million output tokens. A modest bump over the GPT-5 generation on paper. But a controlled cost analysis published by OpenRouter on May 8 reveals that real-world costs jumped between 49% and 92% for developers who upgraded. The culprit isn't the per-token rate - it's how many tokens GPT-5.5 actually generates.

The Token Multiplier Effect

OpenRouter studied a "switcher cohort" - users whose top model was GPT-5.4 before the launch, who then moved to GPT-5.5 for the same workflows. Because both models share the same tokenizer family, the comparison is direct: same prompts, same tasks, different output behavior.

The findings split cleanly by prompt length. For long prompts above 10K tokens, GPT-5.5 produces 19–34% fewer completion tokens, which partially offsets the higher per-token price. But for shorter prompts - the bulk of production traffic - the pattern reverses. Under 2K tokens, completion lengths stay roughly the same. In the 2K–10K range, completions are 52% longer. Since output tokens cost four times as much as input tokens on OpenAI's pricing tier, those extra completions add up fast.

What This Means for the Leaderboard

On LMRank, GPT-5.5 sits at 9.4 out of 10, just above GPT-5's 9.3. That's a 0.1-point improvement for what can be nearly double the effective cost. For context, here's how the top-tier landscape looks on both score and price:

• Claude Opus 4.7 - 9.6 score, $15/$75 per million. Still the quality king, priced accordingly.
• GPT-5.5 - 9.4 score, $12.50/$50 per million. The new default, now with a hidden tax.
• Gemini Ultra 2 - 9.2 score, $10/$40 per million. The only 1M-context flagship under $50 output.
• DeepSeek V4 - 9.0 score, $0.50/$2 per million. Near-frontier quality at open-weight pricing.

The jump from GPT-5 to GPT-5.5 is real - the model scores higher on reasoning and coding benchmarks - but OpenRouter's data suggests most developers are paying for that improvement twice: once at the list price, and again in bloated completion lengths.

The Smarter Alternatives

If you're running high-volume production workloads, the math gets brutal quickly. A 70% effective cost increase on a model that only moved 0.1 points on our aggregate score is a poor trade for anything outside mission-critical reasoning.

Two alternatives on the LMRank leaderboard deserve a harder look:

GPT-5.5 Instant scores 8.6 - just 0.8 points behind its bigger sibling - at $2 per million input and $10 per million output. That's an 84% discount on input and an 80% discount on output. For chat applications, coding assistants, and high-volume pipelines, the quality gap is often invisible in practice.

DeepSeek V4 is the dark horse. It scores 9.0, sits four ranks below GPT-5.5, and costs $0.50/$2 per million - roughly 4% of GPT-5.5's list price and an even smaller fraction of its real-world cost. If your use case doesn't require GPT-5.5's specific tool-calling edge, DeepSeek V4 is now the clearest value play in the top ten.

The Bottom Line

OpenRouter's analysis is a reminder that list prices are marketing, and token efficiency is engineering. GPT-5.5 is still one of the three best general-purpose models on the market. But "best" and "most cost-effective" are no longer the same thing. For teams that upgraded automatically on launch week, the May 8 data is a prompt to audit your inference bills - and maybe your model routing logic.

At lmrank.com, we track scores, pricing, and real-world context windows for every major LLM. See how GPT-5.5 stacks up against the field on its model detail page - or browse categories to find the right tool for your budget.

Sources

• openrouter.ai