Previous edition: Reasoning-token share of provider output, April 2026
Reasoning-token share of provider output, October 2026
Reasoning-token share of provider output, October 2026
reasoning_tokens / output_tokens within each provider. Pure reasoning routes versus mixed routes.
In October 2026 reasoning tokens made up 78% of output on Fireworks, the highest share, and 6% on Mistral, the lowest among providers with meaningful reasoning volume.Providers with negligible reasoning output in October are omitted (ratio too noisy).
How much provider output is reasoning? In October 2026 reasoning tokens made up 78% of output on Fireworks, the highest share, and 6% on Mistral, the lowest among providers with meaningful reasoning volume.
Why it mattersReasoning tokens are billed as output but never shown to users. Providers with a high share serve mostly thinking models, so output cost per visible answer is higher there.
Period
Oct 2026
Updated
November 11, 2026
ID
reasoning-share-october-2026
§ 01
Key findings
- 01Fireworks: 78%.
- 02Melious: 69%.
- 03Vertex (Claude): 58%.
- 04TensorX: 56%.
- 05OpenAI: 48%.
§ 01b
Questions this answers
- Which providers serve the most reasoning tokens?
- What share of output tokens are reasoning tokens?
- How much do thinking models change output cost?
§ 02
Data
| Provider | Reasoning share(percent) |
|---|---|
| Vertex (Gemini) | 40.79% |
| Azure | 41.82% |
| TensorX | 55.94% |
| OpenAI | 47.69% |
| DeepInfra | 10.05% |
| Nebius | 11.86% |
| Bedrock | 26.93% |
| Anthropic | 45.19% |
| DeepSeek | 45.46% |
| Vertex (Claude) | 58.00% |
| Novita | 23.93% |
| Melious | 68.53% |
| Mistral | 6.46% |
| Lyceum | 11.74% |
| Fireworks | 78.49% |
§ 03
Cite as
APA
Click to copy
BibTeX
Click to copy
ID: reasoning-share-october-2026·Updated November 11, 2026·Period Oct 2026