Opus 5.5 at max burned 128,000 tokens on a single pelican

Simon Willison gave the new Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna the same task: produce an SVG of a pelican riding a bicycle. This does not measure general intelligence. It does expose differences between reasoning-effort levels within a model family because the prompt remains fixed.

A week of model launches exposed a price war below the absolute frontier

GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens in the API. GPT-6 Luna costs $0.10 and $0.50 respectively. In Willison's comparison, both are roughly half the price of their GPT-5.6 counterparts.

Claude Opus 5.5 is 20% cheaper than Opus 5 at $4 per million input tokens and $20 per million output tokens. Cache reads fell 60% to $0.20. The price fight is therefore concentrated below GPT-6 Astra and Claude Fable 5.1, which remain at $10 and $50.

The same absurd task revealed behavior that benchmark tables hide

Willison's pelican is not a vendor leaderboard. As a repeatable probe, however, it exposed an operational failure: Opus 5.5 at max effort reasoned until it hit the 128,000 output-token limit and never returned the finished SVG. The same thing happened twice. Each attempt cost $2.56 and took nearly 20 minutes.

That is more useful to an application developer than another fraction of a benchmark point. If an application allows a model to raise effort freely, it also needs limits on time, cost and steps. Otherwise it may purchase more expensive silence.

One pelican does not invalidate the entire max setting

The test is deliberately odd, and it cannot establish that Opus 5.5 fails on long coding jobs. Anthropic reports stronger agentic-coding results and lower costs on typical workloads. Those numbers still come from the vendor and early testers, not from a universal guarantee.

The episode identifies a specific failure mode: more reasoning does not automatically produce a better result. Without a suitable stop condition, a model can spend its entire allowance planning a detail that delivers no value to the user.

Production logs must measure the cost of a completed task

Future comparisons should track success rate, total tokens, elapsed time and cost per completed task across effort levels. The list price per million tokens says nothing about how many an agent will consume before producing an answer.

For teams, the useful signal is the share of tasks completed within a fixed budget and the frequency of human intervention. A cheaper model on the price sheet can cost more in production if it keeps circling above one pelican.

Lilith's verdict

On the price sheet, max effort looks like a bigger engine. Without a limiter, Opus 5.5 twice burned $2.56 and nearly 20 minutes before the pelican even left the garage.