The New Tier

Lesson 1 of 3 in Reasoning Models in the Landscape.

A few product generations ago, getting an Large language model (LLM) to work through a problem step by step was a prompting trick — you asked for the steps and hoped. Then the behavior was trained in deliberately, and a Reasoning model became a thing you could buy: a model that spends generated tokens on Chain-of-thought (CoT) before committing to an answer. How that training works — reinforcement learning against tasks with checkable answers — is the posttraining story. This module is about what it did to the map.

What it did is this: reasoning became a tier, not a family. You will not find a “reasoning vendor” the way you can find an open-weights vendor. Instead, almost every Model family now documents some way to turn thought on, off, or partway — a mode, a dial, a tagged toggle, or a separately listed variant. The tier reaches all the way down: even small models in the verified snapshot document dual-mode reasoning.

For a navigator, that changes the question. Not “which families reason?” but “what reasoning control does this exact model document, and what style of control is it?” The matrix below is the September 2026 answer — read it for the column structure, not the cell contents, because the cells will churn and the structure will not.

Documented reasoning controls by family, quoted or paraphrased from the official model cards and vendor docs verified 2026-09-16. Controls are documented per model card — never assume a sibling model shares them. Sources: huggingface.co/Qwen/Qwen3.8-27B ; huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash ; docs.mistral.ai/getting-started/models/models_overview/ and huggingface.co/mistralai/Mistral-Medium-3.5-128B ; huggingface.co/microsoft/Phi-4-reasoning-vision-15B ; huggingface.co/openai/gpt-oss-120b ; platform.claude.com/docs/en/about-claude/models/overview ; ai.google.dev/gemini-api/docs/models ; huggingface.co/Qwen/Qwen3-0.6B ; huggingface.co/HuggingFaceTB/SmolLM3-3B (fact file: src/data/facts/model-landscape.ts).
FamilyDocumented reasoning controlControl styleDocumented on

Qwen

Thinking mode on by default, with a tunable reasoning_effort

Default-on mode + tunable dial

Qwen3.8-27B model card

DeepSeek

A “continuously controllable reasoning effort” setting — an integer from 1 to 100

Continuous dial

DeepSeek-V4.1-Flash model card

Mistral

Flagship Mistral Medium 3.5 — “our first flagship merged model” — is documented as handling instruction-following, reasoning, and coding with configurable reasoning effort per request

Per-request setting on a merged model

docs.mistral.ai models overview; Mistral Medium 3.5 card

Phi

Hybrid reasoning in a single model: <think>...</think> chain-of-thought for math and science versus <nothink> direct answers for perception tasks

Hybrid think/no-think toggle

Phi-4-reasoning-vision-15B model card

gpt-oss

Adjustable reasoning effort (low/medium/high), with “Full chain-of-thought” access — which the card notes is not intended for end users

Discrete levels

gpt-oss-120b model card

Claude

“Adaptive” thinking steered by an effort parameter (current lineup)

Adaptive mode + effort parameter

Anthropic model overview docs

Gemini

The Live API lineup lists an “Extended Thinking” Live variant alongside a default voice model pitched as the “Default Live API model for most low-latency voice agent experiences without reasoning delays”

Separate model variant

Gemini API model docs

Qwen3-0.6B (SLM)

Documents “seamless switching between thinking mode” and “non-thinking mode” within a single model

Mode toggle — in a sub-1B model

Qwen3-0.6B model card

SmolLM3-3B (SLM)

“It supports dual mode reasoning” — thinking toggled via /think and /no_think

Mode toggle

SmolLM3-3B model card

Read the matrix twice. The first pass gives you today’s facts. The second pass — the one that lasts — gives you a taxonomy of control styles: a continuous dial, discrete levels, a mode toggle inside one model, a default-on mode with a tunable, and a separately listed variant. Every reasoning control you will meet for the next several release cycles is likely to be one of these shapes wearing a new name. When a card documents something, your job is to classify it, because the classification tells you what you can do operationally: dials and per-request settings mean you can route — choose an effort level for each request — while separate variants mean routing happens at model-selection time instead.

Notice also what the fourth column is doing. Every control is attributed to a specific model card, because that is exactly how the underlying fact base records it — several of its rows carry the hedge “verified for that model only.” Family branding does not transfer: a hybrid think/no-think design documented on one Model card tells you nothing certain about a sibling model, an earlier version, or next quarter’s successor. The per-model discipline is not pedantry; it is what keeps your architecture decisions attached to things that are actually documented.

Key terms: Reasoning model, Chain-of-thought (CoT), Model family, Model card, Model release

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.