Best Model for Windsurf (Devin Desktop)

Windsurf now ships as Cognition's Devin Desktop, and the model list carries over. Claude Sonnet 5.5 is the best everyday model there: it scores 61.8% on Terminal-Bench 4.0 and is one of the models marked as recommended. Self-serve usage is billed at the token cost of the model you select, so Sonnet 5.5 ($2/$10 per million tokens) costs half as much as Claude Opus 5.5 ($4/$20), which scores 64.9%. GLM 5.3, also marked recommended, is the strongest open-weight model on Terminal-Bench 4.0 (41.8%) at $1.40/$4.40.

Read the full Windsurf review → · updated 2026-10-17

Sources: Windsurf / Devin Desktop models · Terminal-Bench 4.0 leaderboard · Scale SWE-Bench Pro V2 · DeepSeek V4.1 Flash release

#1Claude Sonnet 5.5Anthropic · $4.40/1M · 1M ctx

Best overall. 61.8% on Terminal-Bench 4.0 at $2/$10 per million tokens, and the Claude model Devin Desktop recommends for everyday agent work.

Overall 96/100 · Value 59Full breakdown →

#2Claude Opus 5.5Anthropic · $8.80/1M · 1M ctx

Best for long agent runs. #1 on Terminal-Bench 4.0 at 64.9%, for the hardest multi-step agent tasks. At $4/$20 it uses quota twice as fast as Sonnet 5.5.

Overall 98/100 · Value 51Full breakdown →

#3GLM 5.3Z.AI · $1.50/1M · 1.0M ctx

Best value. The top open-weight model on Terminal-Bench 4.0 (41.8%) and 95.6 on Scale's SWE-Bench Pro V2, listed at $1.40/$4.40 per million tokens and marked recommended in Windsurf.

Overall 83/100 · Value 69Full breakdown →

#4GPT-6.1 SolOpenAI · $4.40/1M · 1.1M ctx

OpenAI pick. 58.2% on Terminal-Bench 4.0 at $2/$10 per million tokens. Windsurf also lists the older GPT-6 Sol at the same price, which scores 49.4%.

Overall 92/100 · Value 58Full breakdown →

#5Gemini 3.8 FlashGoogle · $1.65/1M · 1.0M ctx

Fast, large-context pick. 94.9 on Scale's SWE-Bench Pro V2 at an introductory $0.75/$3.75 per million tokens through December 31, 2026. Weak on long agent runs (19.1% on Terminal-Bench 4.0).

Overall 76/100 · Value 63Full breakdown →

#6DeepSeek V4.1 FlashDeepSeek · $0.22/1M · 1.0M ctx

Budget pick. The cheapest of these in Windsurf's price table, with open weights. DeepSeek reports 42.3% on Terminal-Bench 4.0 for its October 17 update, up from 31.2%, so use it for routine edits.

Overall 77/100 · Value 93Full breakdown →

FAQ

Which model is best for Windsurf?

Claude Sonnet 5.5 for everyday agent work (61.8% on Terminal-Bench 4.0), Claude Opus 5.5 for the hardest tasks (64.9%), and GLM 5.3 when you want strong results at a lower token price.

How does Windsurf charge for models?

Self-serve quota and extra usage are billed at the token cost of the model you select, so a model's input and output price is what matters. Enterprise plans bill in ACUs, and legacy enterprise plans still use credit multipliers.

Is SWE-2 worth using?

SWE-2 is Cognition's own coding model and is free on self-serve plans through October 15, 2026, then $3/$15 per million tokens. It is not on the public Terminal-Bench 4.0 or SWE-Bench Pro V2 boards yet, so test it on your own tasks.

Are Claude Opus 4.7 or GPT-5.5 still worth picking in Windsurf?

Windsurf lists Claude Opus 4.7 at $5/$25 per million tokens and GPT-5.5 at $5/$30. Their successors cost less per token: Claude Opus 5.5 at $4/$20 (64.9% on Terminal-Bench 4.0) and GPT-6.1 Sol at $2/$10 (58.2%).

Every model on this page runs on Standard Compute

Standard Compute

Current frontier models Claude Fable 5.1 · GPT-6 Astra