← Back to feed

If you are setting detail: low on image inputs to reduce API costs, OpenRouter’s empirical study published July 7 suggests you are paying more for worse results on reasoning models.

Tested across five models from OpenAI and Google on MMMU-Pro Vision, detail: low on GPT-5.5 produced 65.2% accuracy against 79.0% with detail: auto — a 13.8 point gap. Cost: 5.1 cents per question on low detail versus 4.5 cents on auto. Lower image quality, worse accuracy, higher bill.

Why This Happens

The detail: low parameter was designed before extended reasoning existed as a commercial feature. On a classic non-reasoning model, downsampling an image means cheaper input tokens and roughly proportional accuracy loss. The tradeoff was straightforward.

Reasoning models break that assumption. When GPT-5.5 receives a degraded image, it extends its chain-of-thought to fill the information gap — consuming 1.6x more thinking tokens than it would on a clean auto-detail input. Reasoning tokens are priced higher per token than image tokens on most tier-1 APIs. The model compensates its way into a more expensive outcome.

OpenRouter found this pattern holds across every model it tested. detail: auto produced better results on all five, and in several cases also cost less.

The Better Lever

OpenRouter’s recommendation: send higher-quality images at auto or high detail, and adjust reasoning effort level instead.

For tasks that do not require maximum reasoning depth, dialing back the thinking budget achieves better cost efficiency without forcing the model to speculate from degraded inputs. The interaction between image quality and inference mode matters more than either variable in isolation.

For non-reasoning models, detail: low still reduces costs — though the accuracy penalty is measurable there too. The optimisation strategy diverges by model architecture, not just by task.

Developer Economics Implication

Production multimodal pipelines processing thousands of image-heavy requests against GPT-5.5 or Gemini’s top tiers are likely optimising the wrong parameter. Cost-per-task benchmarks at the model level do not automatically transfer to deployed system economics when reasoning mode interacts with input quality.

The practical takeaway is narrow but actionable: audit any production pipeline using detail: low against a reasoning model. The setting that was saving money may now be the line item costing the most.

Benchmark Parameters

• Models: 5 (OpenAI and Google, specific models not fully disclosed)
• Benchmark: MMMU-Pro Vision
• Conditions: Temperature 0, one epoch
• GPT-5.5 result: 65.2% (low) vs 79.0% (auto), 5.1¢ vs 4.5¢ per question