GPT-5.5 vs Claude Opus 4.7 vs DeepSeek V4 vs Kimi K2.6: Benchmark Winners by Category

Benchmark tables make this four-way matchup look easier than it is. The strongest shared table covers GPT-5.5, GPT-5.5 Pro where available, Claude Opus 4.7, and DeepSeek-V4-Pro-Max; Kimi K2.6 appears mainly in separate comparisons, so it is less cleanly comparable across every category . The right conclusion is category-specific: pick the benchmark that resembles your workload, then test the finalists on your own prompts.

Winners at a glance

Workload	Best-supported pick	Why
Science reasoning	Claude Opus 4.7	94.2% on GPQA Diamond, ahead of GPT-5.5 at 93.6% and DeepSeek-V4-Pro-Max at 90.1%
No-tools expert reasoning	Claude Opus 4.7	46.9% on Humanity’s Last Exam without tools, ahead of GPT-5.5 Pro at 43.1%, GPT-5.5 at 41.4%, and DeepSeek-V4-Pro-Max at 37.7%
Tool-assisted exam reasoning	GPT-5.5 Pro	57.2% on Humanity’s Last Exam with tools, ahead of Claude Opus 4.7 at 54.7%
Terminal and agentic computing	GPT-5.5	82.7% on Terminal-Bench 2.0, ahead of Claude Opus 4.7 at 69.4% and DeepSeek-V4-Pro-Max at 67.9%
OS operation	GPT-5.5	78.7% on OSWorld-Verified versus Claude Opus 4.7 at 78.0%
Frontier math	GPT-5.5	51.7% on FrontierMath Tiers 1–3 versus Claude Opus 4.7 at 43.8%
Software engineering in the shared table	Claude Opus 4.7	64.3% on SWE-Bench Pro / SWE Pro, ahead of GPT-5.5 at 58.6% and DeepSeek-V4-Pro-Max at 55.4%
Browsing	GPT-5.5 Pro	90.1% on BrowseComp, ahead of GPT-5.5 at 84.4%, DeepSeek-V4-Pro-Max at 83.4%, and Claude Opus 4.7 at 79.3%
MCP-style public tool workflow	Claude Opus 4.7	79.1% on MCP Atlas / MCPAtlas Public, ahead of GPT-5.5 at 75.3% and DeepSeek-V4-Pro-Max at 73.6%
Vision and document analysis	Claude Opus 4.7	Reported #1 in Vision & Document Arena, with wins in diagram, homework, and OCR subcategories
Cost-sensitive evaluation	DeepSeek V4	VentureBeat reports near state-of-the-art intelligence at about one-sixth the cost of Opus 4.7 and GPT-5.5, but that cost claim should be validated on your own workload
Least clean four-way comparison	Kimi K2.6	Kimi has useful reported scores, but the cited Kimi evidence is mostly separate from the GPT-5.5, Claude Opus 4.7, and DeepSeek-V4-Pro-Max table

Detailed benchmark table

Benchmark / capability	GPT-5.5	GPT-5.5 Pro	Claude Opus 4.7	DeepSeek V4 / V4 Pro Max	Kimi K2.6	Best-supported read
GPQA Diamond	93.6%	Not reported	94.2%	90.1% for DeepSeek-V4-Pro-Max	Not reported	Claude leads the shared table
Humanity’s Last Exam, no tools	41.4%	43.1%

Rows that mix sources should be read carefully. A Kimi score reported in a separate Kimi-focused comparison is useful, but it is not as strong as a result produced in the same shared table and harness as GPT-5.5, Claude Opus 4.7, and DeepSeek-V4-Pro-Max .

GPT-5.5: strongest on terminal, OS, math, and tool use

GPT-5.5’s clearest win is Terminal-Bench 2.0: 82.7% versus Claude Opus 4.7 at 69.4% and DeepSeek-V4-Pro-Max at 67.9% in the shared table . That is one of the largest gaps in the sourced benchmark set.

It also leads Claude Opus 4.7 on OSWorld-Verified, but by a narrow 78.7% to 78.0% margin . On FrontierMath Tiers 1–3, the GPT-5.5 lead is larger: 51.7% versus Claude’s 43.8% .

GPT-5.5 Pro changes the picture when tools or browsing are central. It leads Humanity’s Last Exam with tools at 57.2%, ahead of Claude Opus 4.7 at 54.7%, GPT-5.5 at 52.2%, and DeepSeek-V4-Pro-Max at 48.2% . It also leads BrowseComp at 90.1%, ahead of GPT-5.5 at 84.4%, DeepSeek-V4-Pro-Max at 83.4%, and Claude Opus 4.7 at 79.3% .

GPT-5.5 does not lead every reasoning test. Claude Opus 4.7 narrowly beats it on GPQA Diamond, 94.2% to 93.6%, in the shared table . A separate GPT-5.5 guide reports GPT-5.5-only domain results including 91.7% on Harvey BigLaw Bench, 88.5% on an internal investment-banking benchmark, and 80.5% on BixBench, but those should not be treated as four-way wins because the cited excerpt does not report the same scores for Claude Opus 4.7, DeepSeek V4, and Kimi K2.6 .

Claude Opus 4.7: strongest on no-tools reasoning and documents

Claude Opus 4.7 has the best no-tools reasoning profile in the main shared table. It leads GPQA Diamond at 94.2% and Humanity’s Last Exam without tools at 46.9% . It also leads SWE-Bench Pro / SWE Pro at 64.3% and MCP Atlas / MCPAtlas Public at 79.1% in that same table .

Claude’s weaker area in the cited data is terminal-style operation. GPT-5.5 leads Claude by more than 13 points on Terminal-Bench 2.0, 82.7% to 69.4%, and also leads Claude on OSWorld-Verified and FrontierMath Tiers 1–3 .

Claude has the strongest cited multimodal and document signal. One source reports Claude Opus 4.7 taking #1 in Vision & Document Arena, improving by 4 points over Opus 4.6 in Document Arena, and winning diagram, homework, and OCR subcategories . The same source does not provide comparable numeric Vision & Document Arena scores for GPT-5.5, DeepSeek V4, or Kimi K2.6, so this supports Claude’s document strength but not a complete four-way multimodal ranking .

DeepSeek V4: competitive, but the main cited edge is cost-performance

The sources use more than one DeepSeek label. The shared benchmark table reports DeepSeek-V4-Pro-Max, while the Artificial Analysis comparison reports DeepSeek V4 Pro with a 1,000k-token context window . Those labels should not be treated as automatically interchangeable.

In the main shared table, DeepSeek-V4-Pro-Max is competitive but does not lead any row. It scores 90.1% on GPQA Diamond, 37.7% on Humanity’s Last Exam without tools, 48.2% on Humanity’s Last Exam with tools, 67.9% on Terminal-Bench 2.0, 55.4% on SWE-Bench Pro / SWE Pro, 83.4% on BrowseComp, and 73.6% on MCP Atlas / MCPAtlas Public .

DeepSeek’s strongest cited product claim is cost-performance rather than a category win. VentureBeat describes DeepSeek V4 as delivering near state-of-the-art intelligence at about one-sixth the cost of Opus 4.7 and GPT-5.5 . That is a reason to test it for cost-sensitive workloads, not a reason to skip workload-level validation.

For long-context screening, one Artificial Analysis comparison lists both DeepSeek V4 Pro and Claude Opus 4.7 at 1,000k-token context windows . That supports parity for those listed configurations, not a broader claim about every DeepSeek or Claude mode .

Kimi K2.6: promising scores, weaker direct comparability

Kimi K2.6 is the hardest model to rank cleanly in this set because it is not included in the main shared table against GPT-5.5, Claude Opus 4.7, and DeepSeek-V4-Pro-Max . A Kimi-focused comparison reports K2.6 at 58.6% on SWE-Bench Pro, 80.2% on SWE-Bench Verified, 66.7% on Terminal-Bench 2.0, 54.0% on Humanity’s Last Exam with tools, and 89.6% on LiveCodeBench v6 . That source says the K2.6 numbers come from a Moonshot AI official model card, but the comparison set is mainly Claude Opus 4.6 and GPT-5.4 rather than the exact four-way lineup here .

A separate Kimi vs DeepSeek comparison reports Kimi K2.6 at 96.4% on AIME 2026 in Thinking mode, 27.9% on APEX Agents in Thinking mode, and 83.2% on BrowseComp with Thinking mode and context management . In that same source, DeepSeek-V4 Pro is listed at 83.4% on BrowseComp, while DeepSeek values are not available for AIME 2026 and APEX Agents .

That makes Kimi worth testing, especially for coding, agentic, and browsing workloads, but the sourced material does not support a clean overall ranking against GPT-5.5 and Claude Opus 4.7 across the same benchmark suite .

Which model should you test first?

• Test GPT-5.5 first for terminal-heavy agents, OS-operation tasks, and FrontierMath-like work; it leads the cited Terminal-Bench 2.0, OSWorld-Verified, and FrontierMath results .
• Test GPT-5.5 Pro first when tool-augmented reasoning or browsing is central; it leads Humanity’s Last Exam with tools and BrowseComp in the shared table .
• Test Claude Opus 4.7 first for GPQA-style science reasoning, no-tools expert Q&A, SWE-Bench Pro-style software engineering, MCP-style workflows, and document-heavy multimodal work .
• Test DeepSeek V4 first when cost-performance is the main constraint and you can run your own quality checks; the cited advantage is the reported near-frontier performance at about one-sixth the cost of Opus 4.7 and GPT-5.5 .

Benchmark caveats that matter

This is not a universal leaderboard. The sources mix base and Pro variants, including GPT-5.5, GPT-5.5 Pro, DeepSeek-V4-Pro-Max, DeepSeek V4 Pro, Claude Opus 4.7, and Kimi K2.6 . Some results are also vendor-reported, and OpenAI notes that its GPT evaluations for ARC were run with reasoning effort set to xhigh in a research environment that may differ from production ChatGPT .

Close margins should be treated as directional. Claude’s GPQA Diamond lead over GPT-5.5 is 0.6 points, and GPT-5.5’s OSWorld-Verified lead over Claude is 0.7 points . Larger gaps are more actionable: GPT-5.5’s Terminal-Bench 2.0 lead over Claude is more than 13 points, and its FrontierMath lead over Claude is 7.9 points .

The practical bottom line: there is no single winner across GPT-5.5, Claude Opus 4.7, DeepSeek V4, and Kimi K2.6. Use the benchmark category that maps to your real workload, then rerun the same evaluation across the models you can actually deploy.