Windsurf Deploys Opus 4.7 Fast Mode With 81 TPS Output

On May 12, 2026, Windsurf made Claude Opus 4.7 (fast mode) immediately available in its AI-native code editor. The integration brings Anthropic’s flagship model to the IDE with a high-speed configuration that outputs approximately 81 tokens per second. This 2.5x speed increase over standard throughput removes a major latency bottleneck for developers using heavy AI coding assistants.

Performance and Capabilities

The fast mode configuration maintains the full intelligence baseline of standard Claude Opus 4.7. The model scores 70% on CursorBench, representing a 12-point increase over its predecessor, and achieves 87.6% on SWE-bench Verified.

Vision capabilities received a substantial upgrade in this generation. The model supports high-resolution image processing up to 3.75 megapixels, capping at 2,576 pixels on the longest edge. This 3x resolution increase allows the editor to parse complex application state diagrams, wireframes, and densely packed UI screenshots without downscaling artifacts.

Metric	Opus 4.6	Opus 4.7
CursorBench Score	58%	70%
Agentic Tool Calls	16.3	7.1

Despite these performance gains, the restricted Mythos model retains the top spot for autonomous engineering tasks, leading Opus 4.7 on SWE-bench Pro 77.8% to 64.3%.

Architecture and Context Updates

The integration leverages core Opus 4.7 framework updates to reduce round-trip latency. Anthropic improved the model’s agentic persistence, halving the number of model calls required to solve complex problems from an average of 16.3 down to 7.1.

A redesigned tokenizer shifts the underlying cost calculation. The new architecture improves encoding efficiency for non-Latin scripts by 20 to 35 percent. English workloads experience a strict tradeoff, with token counts increasing by roughly 12 to 18 percent for the exact same plaintext input. The model maintains a 1 million token input context window alongside a 128,000 token output limit.

Credit Consumption and Tuning

Standard API pricing for Opus 4.7 is $5 per million input tokens and $25 per million output tokens. In standalone environments like the Claude Code CLI, fast mode is priced at $30 per 150 million tokens. Windsurf abstracts this into a proprietary credit system. Fast mode requests consume 10x standard credits per prompt. Enabling reasoning steps pushes this consumption to 12x credits.

To manage this cost, Windsurf introduced a new xhigh effort level. Situated precisely between the existing “high” and “max” settings, xhigh provides granular control over reasoning depth, allowing developers to balance latency and credit burn for intermediate refactoring tasks.

If you manage complex, multi-step editing tasks in Windsurf, default to the xhigh effort level in fast mode. The 2x reduction in model calls offsets the 12x credit premium, keeping overall usage efficient while significantly reducing idle wait times during large refactors.