New coding models landed this week. Set the effort ceiling on purpose.

If your team uses Claude Code or Codex, check whether the model doing the work changed this week. In at least one product, it did, whether or not anyone on the team decided it should.

Anthropic released Claude Opus 5.5, and OpenAI released GPT-6 Sol and GPT-6 Luna within about an hour of it. Draftbit’s product note for the same day says that in Draftbit, Claude Code now starts on Opus 5.5 and Codex now starts on GPT-6 Sol. Simon Willison wrote that he now uses those two as his defaults in both tools.

For a technical leader, the model is only half of the change. The other half is the effort setting, and that is where this week’s most useful warning came from.

What happened

Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. The company says cache reads make up the majority of agentic and coding work costs, and that at default settings the model will cost 40% less than Opus 5 on typical workloads. Those are company-reported figures. Its benchmarks treat medium as the default effort and also report results at higher settings, including max.

Willison ran his usual test prompt, an SVG of a pelican riding a bicycle, on Opus 5.5 at the “max” thinking level. The model never answered. It reached the 128,000-token output limit while it was still reasoning about the drawing. He tried again and got the same result. Each failed attempt cost him $2.56 and took nearly 20 minutes. His conclusion was that he does not trust “max” not to do the same on more interesting work.

That is one person’s test on one prompt, not a lab benchmark. It is still worth acting on, because the failure is cheap to prevent and expensive to discover in production.

Why the ceiling matters more than the model

An interactive session fails in front of someone. They see the spinner, they stop it, they try again at a lower setting. An unattended agent, such as a nightly job, a review bot or a triage loop, fails quietly. It spends the full output budget, returns nothing, and the next run does the same.

The cost of one runaway call has a simple upper bound: the output token limit times the output price.

A worked example

Take the figures from this week’s sources:

• Output limit: 128,000 tokens
• Output price for Opus 5.5: $20 per million tokens
• Worst case for one call: 0.128 × $20 = $2.56

That is exactly the cost Willison reported per attempt, consistent with each one running all the way to the ceiling.

Now picture a hypothetical nightly job that sends 40 tickets through an agent, one call each, at max effort. If every call over-thinks the same way, the worst case is 40 × $2.56 = $102.40 a night, and about 40 × 20 minutes of wall-clock time, with no answers at the end. Over a month of nights, that is roughly $3,000 of spend with nothing to show for it. The numbers are illustrative; the arithmetic is the point. Run it with your own call counts before you decide the setting.

A decision rule for effort levels

Write this into your coding-agent policy, in your own words:

1. Unattended work runs at the default effort or below unless an eval shows a higher level wins on that exact task.
2. Max is interactive only. A person starts it, watches it, and can stop it.
3. Every unattended job has a budget: a token cap per call, a call cap per run, and a timeout shorter than the model’s worst case.
4. An empty answer is an error, not a retry. If a call returns no usable output, the job stops and alerts someone. It does not loop.
5. Model or effort changes go through the eval suite first, the same way a dependency upgrade goes through CI.

The morning-after checklist

When a default model changes underneath your tools, check these before the next unattended run:

• Which jobs pin a model, and which inherit the tool’s default?
• Which effort level does each unattended job use, and who chose it?
• What does the new price do to last week’s bill at the same volume?
• Do cache hits still land? A prompt layout change can quietly move work from $0.20 cache reads to full-price input.
• Does every merge still get a named human approval? A new model is not a reason to change that rule.
• Did someone rerun the eval suite on the new default?

What to do this week

• List every place a coding agent runs without a person watching. Note the model and effort level for each.
• Pin the model in any job where a silent change would hurt, and set an explicit effort level instead of trusting a default.
• Add a per-call token cap and a per-run spend cap to each job.
• Run five real tasks from your backlog through the old and new defaults at medium effort, and compare the results before you widen anything.

New models will keep arriving on this schedule. A written rule for effort and budget means the next default change is a config review, not an incident.