DeepSeek V4 vs. Kimi K2.6 vs. Claude Opus 4.7 vs. GPT-5.5: which model wins where?

The short version

• Coding and GitHub issue fixes: Claude Opus 4.7 looks strongest in the cited SWE-Bench, SWE-Bench Verified and CursorBench figures . OpenAI's own SWE-Bench Pro number for GPT-5.5 is 58.6% .
• Terminal agents and computer use: GPT-5.5 is the clearest pick on documented numbers, including 82.7% on Terminal-Bench 2.0 in OpenAI's release .
• Budget-conscious coding agents: Kimi K2.6 is described by CodeRouter as the cost/quality winner at $0.60 input and $4.00 output per million tokens .
• DeepSeek V4: V4-Pro and V4-Flash are available through the DeepSeek API, but the sources used here do not contain an official four-model benchmark table against Kimi K2.6, Claude Opus 4.7 and GPT-5.5 .

What can be compared reliably

OpenAI describes Terminal-Bench 2.0 as a benchmark for complex command-line workflows that require planning, iteration and tool coordination; GPT-5.5 reaches 82.7% there, according to OpenAI . On SWE-Bench Pro, which evaluates real-world GitHub issue resolution, OpenAI lists GPT-5.5 at 58.6% .

For DeepSeek V4, the official evidence is about availability rather than benchmark dominance. DeepSeek says V4-Pro and V4-Flash can be used through both the OpenAI ChatCompletions interface and the Anthropic interface, with the model parameters deepseek-v4-pro and deepseek-v4-flash .

For Claude Opus 4.7 and Kimi K2.6, the most useful direct comparisons in this source set are less official: LushBinary provides concrete Claude-vs.-GPT benchmark figures, while CodeRouter provides pricing and positioning claims for Kimi K2.6 and DeepSeek V4 .

Benchmark table: the numbers in the available sources

N/A means the provided sources do not include a sufficiently comparable figure for that model-benchmark combination.

Coding: Claude has the stronger cited scores; Kimi is the budget test

If coding benchmarks are your north star, Claude Opus 4.7 has the strongest cited case. LushBinary gives Claude Opus 4.7 64.3% on SWE-Bench Pro versus GPT-5.5's 58.6%, and OpenAI independently reports the GPT-5.5 SWE-Bench Pro figure as 58.6% . The same third-party table puts Claude ahead on SWE-Bench Verified and CursorBench .

Kimi K2.6 is still worth a look for engineering teams running many agent attempts, drafts or retries. CodeRouter says Kimi K2.6 performs at GPT-5.5 level on SWE-Bench Pro while listing much lower token prices for Kimi . That is not a substitute for your own evaluation, but it is a strong reason to include Kimi in a cost-per-accepted-fix test.

For DeepSeek V4, the official DeepSeek evidence in this source set supports availability, not a coding score: V4-Pro and V4-Flash can be called through the current API surfaces .

Terminal agents and computer use: GPT-5.5 is best documented

For terminal-heavy agents, GPT-5.5 has the clearest public support. OpenAI says Terminal-Bench 2.0 tests complex command-line workflows requiring planning, iteration and tool coordination; GPT-5.5 reaches 82.7% . LushBinary places Claude Opus 4.7 in the same benchmark at about 72% .

On the cited knowledge-work and computer-use figures, GPT-5.5 also leads Claude: 84.9% versus about 78% on GDPval, and 78.7% versus about 65% on OSWorld-Verified . That makes GPT-5.5 the strongest documented starting point for shell-heavy tool orchestration and GUI-adjacent workflows.

Vision and documents: Claude has the clearest positive signal

The available sources do not provide a complete four-model vision table. The strongest positive signal is for Claude Opus 4.7: Latent Space/AINews cites an Arena report placing Claude Opus 4.7 at No. 1 in the Vision & Document Arena .

LLM Stats also reports that Claude Opus 4.7 can process images up to 2,576 pixels on the long edge, or roughly 3.75 megapixels, while GPT-5.5 supports image input and is listed with MMMU Pro scores of 81.2% without tools and 83.2% with tools . Useful context, yes; a full four-way comparison with Kimi K2.6 and DeepSeek V4, no.

Price-performance: include Kimi K2.6 and DeepSeek V4 Flash in your own eval

Kimi K2.6 has the clearest price-performance argument in the provided sources. CodeRouter describes it as the cost/quality winner and lists $0.60 input and $4.00 output per million tokens .

DeepSeek V4 Flash is even cheaper in that same source, listed at $0.14 input and $0.28 output per million tokens with a 1M context window . DeepSeek's changelog separately confirms V4-Pro and V4-Flash support via the OpenAI ChatCompletions interface and the Anthropic interface, using deepseek-v4-pro and deepseek-v4-flash .

But price-performance is not the same as benchmark leadership. A cheaper model may be excellent for low-risk drafts, batch attempts and fallback routes; in production, the meaningful metric is cost per accepted result after retries, review time and failure handling.

How to test the four models fairly

Do not make a production decision from a single public leaderboard. Build a compact eval set from your real repositories, documents, support tickets or agent workflows. Measure not only the first answer, but also:

• accepted-result rate;
• retries needed;
• latency;
• total token cost;
• severity of wrong answers;
• human review time.

Also keep official and secondary data in separate columns. In this comparison, GPT-5.5 has official OpenAI numbers for Terminal-Bench 2.0 and SWE-Bench Pro . DeepSeek V4 has official API availability evidence . The strongest direct Claude and Kimi comparison claims here are from third-party sources .

Bottom line

There is no universal benchmark winner. Claude Opus 4.7 leads the cited coding-side numbers, GPT-5.5 is best documented for terminal and computer-use workflows, Kimi K2.6 has the strongest cost/quality narrative, and DeepSeek V4 is an available API candidate that still needs workload-specific evaluation before you treat it as a benchmark leader .