Skip to content

14 reasoning models

Reasoning models

Nearly every frontier model now thinks before it answers. That changes the cost model: hidden reasoning tokens bill at the output rate, so the model with the lower sticker price is not always the cheaper one per finished task.

Pricing checked against provider documentation on . How we verify

The cost trap

A model at $20 per million output tokens that solves your task in 2,000 tokens is cheaper than one at $6 that burns 15,000 tokens thinking. This is the single most common mistake in model selection — always benchmark cost per completed task on your own workload before locking in a provider.

14 of 14 models

Model Provider Input / 1M Output / 1M Context Best for Status
DeepSeek V4-Flash reasoning open DeepSeek $0.22 $0.66 1M The cheapest capable open-weight option stable
GPT-5.6 Luna reasoning OpenAI $0.20 $1.20 1.05M High-volume work on a tight budget stable
DeepSeek V4-Pro reasoning open DeepSeek $0.66 $1.98 1M Frontier-adjacent reasoning at open-weight prices stable
Claude Haiku 4.5 reasoning Anthropic $1.00 $5.00 200K Lowest-latency Claude responses stable
GLM-5.2 reasoning open Z.ai $1.40 $4.40 1M MIT-licensed coding agents stable
Gemini 3.5 Flash reasoning Google $1.50 $9.00 1.05M Agentic loops and sub-agent fleets stable
Claude Sonnet 5 reasoning Anthropic $2.00 $10.00 1M Fast Claude-quality reasoning at scale stable
GPT-5.6 Terra reasoning OpenAI $2.00 $12.00 1.05M Balanced production workloads stable
Gemini 3.1 Pro reasoning Google $2.00 $12.00 1.05M Best price-to-reasoning ratio preview
Kimi K3 reasoning open Moonshot AI $3.00 $15.00 1.05M Natively multimodal long-horizon agents stable
GPT-5.6 Sol reasoning OpenAI $4.00 $20.00 1.05M Frontier coding and long-horizon agents stable
Claude Opus 5 reasoning Anthropic $5.00 $25.00 1M Most teams' best Claude starting point stable
Claude Fable 5 reasoning Anthropic $10.00 $50.00 1M Long-horizon agentic engineering stable
Claude Mythos 5 reasoning Anthropic $10.00 $50.00 1M Vetted cyber-defence work only restricted

Effort settings are the main dial

OpenAI exposes effort levels from none through max, and Anthropic ships adaptive thinking that scales with problem difficulty. Tuning this per request typically saves more than switching providers.

Always-on thinking has a downside

Some models cannot turn reasoning off. That guarantees quality on hard problems but makes them poor fits for latency-sensitive paths like autocomplete or live chat suggestions.

Route, don't standardise

Production systems increasingly send easy requests to a cheap tier and escalate only when a check fails. See the agent guide

Frequently asked

What is a reasoning model?

A reasoning model generates an internal chain of thought before producing its visible answer. That hidden reasoning is billed at the output token rate, which is why a reasoning model's real cost per task can be several times its sticker price suggests.

Are reasoning models always better?

No. They help on multi-step problems — maths, debugging, planning, long-horizon agent work — and hurt on simple extraction or classification, where they add latency and cost without improving accuracy. Route by task difficulty rather than defaulting to the strongest model.

How do I control reasoning cost?

Most providers now expose an effort or thinking-budget parameter. Lowering effort on easy requests is the single biggest lever; prompt caching and batch processing are the next two. Measure cost per completed task, not cost per token.

Which reasoning model is cheapest?

DeepSeek V4-Flash at $0.22 input and $0.66 output per million tokens, which makes it the usual choice for high-volume agent subtasks.