Thinking models compared: who has it, and what it costs across providers

Thinking is close to universal here

25 of the 31 catalog entries list thinking as a capability. All 12 Claude ids except stable-claude-haiku-4-5 have it; all 8 OpenAI ids except gpt-image-2; grok-4.6 and grok-4.5 on the xAI side; and all 5 Gemini ids.

Among text models the exceptions are just two: stable-claude-haiku-4-5, which lists text and tools at $0.40 per million input and $2.00 output, and grok-composer-2.5-fast, which lists text and tools at $0.15 and $0.75. The other four entries without thinking are the image and video models, where the capability would not mean anything.

So "does it think" is not a filter that narrows the catalog. Almost everything does. The useful questions are what it costs when the model uses it, and which of the two ids that cannot spend reasoning tokens you might actually want.

There is no separate rate for thinking

Every catalog entry publishes exactly two per-token rates: input and output. There is no third rate for reasoning. Whatever a model produces while working through a problem is billed under the rates you already know, which means the output rate is the one that decides what extended thinking costs you.

That rate is not close to the input rate. Across the catalog, official output is 3× to 6× official input, and the multiplier does not change the ratio because it applies to both sides equally. grok-4.6 and grok-4.5 have the narrowest gap at exactly 3× ($0.20 input against $0.60 output after the 0.1× multiplier). The six gpt-5.x ids, gemini-3-flash and gemini-3.1-pro-low sit at 6×. The remaining 17 text ids — every Claude id, stable-gpt-6-astra, grok-composer-2.5-fast and the other three Gemini ids — are at 5×.

The practical read: on a model with a 6× ratio, a response that runs long is six times more expensive per token than the prompt that triggered it. Anything that lengthens responses — thinking included — lands on the expensive side of the bill, and it lands there regardless of which lineup you picked.

Read the levels off the catalog, not off a name

Models that support thinking do not all offer the same set of effort levels, and the catalog is where that is recorded. GET /v1/models returns a thinking_levels array per model alongside its context_window, max_output_tokens, capabilities and endpoint, so a client can read the real set instead of hard-coding one.

This matters because ids that look related are not always configured alike, and an id whose name suggests a fixed effort level may still expose a range. Query the catalog when you integrate, and again when you add a model, rather than transcribing a table into your code.

The listing is also the cheapest way to check the three things that actually gate a switch: the endpoint the model answers on, its output ceiling, and whether it lists thinking at all.

curl -sS https://token-share.app/v1/models \
  -H "Authorization: Bearer $TOKEN_SHARE_KEY" \
| jq '[.data[]
      | select(.capabilities and (.capabilities | index("thinking")))
      | {id, endpoint, context_window, max_output_tokens, thinking_levels}]'

Deciding whether to spend the tokens

The catalog cannot tell you whether thinking improves an answer on your task — it records the capability, not the outcome. What it can tell you is the price of finding out, and on the Flash-tier ids that price is small enough that guessing is the more expensive option.

A workable test: take tasks where you can check the answer objectively, run each one twice on the same id — once with your normal request and once with extended thinking configured — and compare correctness against the output token counts the usage block reports. Multiply the extra output tokens by that model’s output rate to get the real premium, rather than assuming it.

Two results are common enough to plan for. If accuracy does not move, you have found a place to leave thinking off, and the two text ids without it — stable-claude-haiku-4-5 and grok-composer-2.5-fast — become worth a look, since they cannot run up a reasoning bill at all. If accuracy does move, compare that against the alternative of a stronger model without thinking, which is a different spend on the same problem. Both comparisons need your tasks; neither is answerable from a specification table.

Common questions