Thinking & Reasoning

Some AI models think out loud before they answer. PebbleChat surfaces this reasoning — the chain of thought the model uses to work through a problem — in a collapsible Thinking panel above the final response.

What reasoning-capable models do differently

Traditional language models generate a response token by token, committing to each word as they go. Reasoning-capable models add a thinking step first: they work through the problem internally, explore alternatives, check their own work, and only then commit to a final answer. The thinking can run for many seconds or even minutes on hard problems.

Models that support explicit thinking in PebbleAI include:

• Claude Opus (extended thinking mode)
• Claude Sonnet (extended thinking on some versions)
• OpenAI o1 / o3 (reasoning models — reasoning is always on)
• Reasoning-capable Bedrock models marked with a thinking indicator in the model selector

For models that don’t support thinking (Claude Haiku, GPT-4o, Gemini Pro), no thinking panel is shown — they just respond in the normal way.

What the Thinking panel shows

When a reasoning-capable model is working, PebbleChat surfaces:

1. An animated “Thinking…” indicator that appears as soon as the model starts
2. The elapsed thinking time — so you know it’s working, not stuck
3. A collapsible panel with the model’s full reasoning trace once thinking is complete

Expand the panel to see:

• Step-by-step problem decomposition — how the model broke the question down
• Alternatives it considered and rejected — often more interesting than the final answer
• Self-corrections — places where the model noticed its own mistake and backtracked
• Verification — final checks the model performed before committing

The panel stays collapsed by default to keep the chat clean, but you can expand it any time.

Why it matters

Trust for hard problems

The best reason to look at the Thinking panel is when you’re about to act on something important. If PebbleChat says “yes, this contract clause is risky” or “this query should be safe to run in production”, expanding the reasoning lets you verify the model actually thought it through — or catch the case where it skipped a step.

Catching errors early

Occasionally the model’s reasoning reveals a flaw you need to correct. If you see in the Thinking panel that the model misread your intent, you can stop the generation and clarify before the final response solidifies the misunderstanding.

Learning from the model

Reasoning traces are a fantastic teaching tool. Watching how a good model decomposes a hard problem — what it considers, what it rules out, what it verifies — is useful whether you’re learning a technical topic or getting better at your own prompting.

Cost transparency

Thinking costs tokens. For expensive reasoning models (Opus with extended thinking, o1, o3), the Thinking panel is usually where most of the cost of a response goes. Seeing how much thinking a particular prompt triggered helps you calibrate what to ask which model.

Turning thinking effort up or down

Some models let you control how hard they think. In the composer, you’ll see a thinking effort selector (sometimes showing “Balanced” or “Low/Medium/High”) for models that support it.

• Low effort — fast, cheap, less thorough; good for quick questions
• Balanced — the default; most questions should use this
• High effort — slow, expensive, most thorough; save for truly hard problems

You change effort per message, before you send. Like the model selector, it’s a dropdown on the right side of the composer.

When a model doesn’t support effort controls, the selector is disabled and shows “Thinking effort unavailable for this model”.

Thinking and Activity Stream

Thinking and the Activity Stream are different things that can co-exist:

• Thinking is internal reasoning — the model working through the problem conceptually
• Activity Stream is external actions — web searches, tool calls, document retrievals

A response from a reasoning model that also triggered research will have both a Thinking panel and a Research & Tools Activity Stream. You can expand either independently.

When not to look at the Thinking panel

Most responses don’t need the Thinking panel. For:

• Routine drafts (emails, summaries, reformatting)
• Conversational back-and-forth
• Simple factual questions
• Anything where you’d rather have a fast answer than a justified one

… the Thinking panel is overhead you can ignore. Leave it collapsed, trust the response, and move on.

When to always check the Thinking panel

• High-stakes decisions — legal, financial, medical, safety-critical content
• Security-sensitive recommendations — cryptography choices, auth design, secret management
• Irreversible actions — anything you can’t undo
• Novel problems — when the model is out of its training depth and might be making things up

Related

• Model Selection — which models support thinking
• Activity Stream — external tool/research steps, separate from thinking
• Messages & Responses — how thinking fits into the response lifecycle