How Hard Should Your AI Think?

The first week of September was busy. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1, Google released Gemini 3.8 Flash on September 2, and OpenAI launched GPT-6 Astra on September 3. For a small team, the interesting part isn't the leaderboard reshuffle. It's that these releases all expose a reasoning-effort dial. With the same model, where you set that dial changes your bill and your answer quality at the same time.

What is the reasoning-effort dial, and why does it matter now?

Effort tells the model how long to think before answering, and thinking tokens are billed as output. That makes it a pricing switch and a quality switch in one parameter. Artificial Analysis measured Gemini 3.8 Flash at $0.24, $0.41, and $0.58 per task across low, medium, and high — roughly a 2.4x spread — with intelligence scores of 52, 57, and 59. In other words, cost can more than double while the score moves seven points.

The dial also behaves differently per product. Claude Fable 5.1 lets you switch effort per message mid-conversation without breaking your cache, while GLM-5.3 removed the option to turn thinking off entirely. Claude Code runs Fable 5.1 at High by default. If you have never touched the setting, you are already paying for a default someone else chose.

Where should a small team set the dial?

Set it per workload, not once per model. And judge it by cost per completed task, not by the per-token rate, because higher effort doesn't just make answers longer — it adds tool calls and round trips. Artificial Analysis found Claude Sonnet 5 burned roughly six times more agentic turns at maximum effort than at low effort on knowledge-work evaluations.

• Low: ticket routing, tagging, pulling line items out of receipts and orders, shortening product copy — repetitive work with a narrow right answer.
• Medium: first-draft replies to customer questions, blog outlines, code edits that follow an established pattern in your repo.
• High: comparing two contract versions, chasing anomalies in payout data, diagnosing an outage you've never seen — the small set where a wrong answer creates human cleanup.
• Cheap tier: for bulk classification and cleanup, low-cost open-weight options like GLM-5.3-Flash or Qwen3.8-Flash often hold quality at a fraction of the cost.

What can you do about it this week?

You don't need a re-platforming project. List every place your product calls a model, then write down one default effort level per place. That's a half-day of work for most small teams.

1. List all AI call sites and note, in one line, who cleans up when each one is wrong.
2. Save 20 real requests per call site as golden samples — pick the ambiguous ones, not the easy wins.
3. Run the same samples at low, medium, and high, recording accuracy, cost per completed task, and turn count.
4. If the quality gap is small, adopt the lower setting as your default and record the reason next to the config.
5. Match your eval effort to production effort. If production runs medium, your eval runs medium — otherwise the comparison is meaningless.

Set re-measurement triggers too. The obvious ones are model version changes and expiring introductory prices: Gemini Flash introductory rates double on January 1, 2027, and OpenAI moved GPT-5.6 Sol below its launch list price. New models keep widening the range, as GPT-6 Astra's broader effort dial shows. The dial isn't a new feature so much as a new responsibility — deciding how hard each task should think now matters more than which model name you pick.