Fast vs reasoning models

Two models, same brand, same chat box — one answers instantly, the other takes a minute and shows a “thinking” spinner. The difference is not quality versus speed with a clear winner. It is a trade: time and cost for deliberate problem-solving — and choosing correctly is a skill.

What “reasoning” means here

A reasoning model spends extra computation before committing to an answer: it works through steps, checks itself, explores alternatives. Sometimes you see a summary of that thinking; often you just wait longer. A fast model answers in one pass: quick, cheap, and surprisingly good at anything that does not need deliberation.

The trade is visible in the pricing. Current example (September 2026, official docs): OpenAI’s lineup spans a flagship (GPT-6 Astra, $10 in / $50 out per million tokens) through balanced and cost tiers (GPT-5.6 Terra at $2/$12; GPT-5.6 Luna at $0.20/$1.20), with reasoning effort adjustable per request. Anthropic spans Fable 5.1 (demanding reasoning, slower, $10/$50) through Sonnet 5 (speed + intelligence, $2/$10) to Haiku 4.5 (fastest, $1/$5), with “adaptive” thinking steered by an effort setting. Names and prices change — treat this as the shape of the market, and check current docs before deciding.

When the fast tier wins

• Rewriting, summarizing, translating, extracting — one-pass transformations.
• Classification and triage at volume.
• Quick drafts you will edit anyway.
• Anything interactive: latency is the feature.

When to pay for thinking

• Multi-constraint problems: “check this contract against these five rules.”
• Planning and analysis where a wrong answer costs more than the waiting.
• Debugging or logic that needs several steps to verify.
• Drafts you cannot easily check by eye.

Calibration

• Escalate on doubt. If a quick pass looks shaky, re-run it with the reasoning tier — cheaper than starting heavy for everything.
• Reasoning is not truth. A long chain can still hallucinate; verify specifics as always.
• Bulk + synthesize. Fast models for the twenty extractions; a reasoning model for the synthesis that matters.

A bad example

Ask: “Analyze this 30-page vendor contract against five constraints — jurisdiction, penalties, renewal, data terms, deadlines.”

The cheapest, fastest tier may skate over clause interactions. A missed penalty clause is expensive; the extra think-time is not.

A better example

Fast tier: “Extract every clause mentioning penalties, with section numbers.” Then reasoning tier: “Using only the clauses above, find conflicts and rank the top three risks.”

Extraction is a pattern task; conflict analysis is a reasoning task. Two tiers, two jobs, better result than either alone.

Practice

Transfer

• Daily drafting: fast tier, always.
• Decisions with money, law or health stakes: reasoning tier plus a source check.
• High-volume pipelines: fast first, escalate only the exceptions.

Next

Next: Why answers vary — what changes between two runs of the same model. The reasoning model glossary entry is the one-line version.