The benchmarks worth separating

SWE-Bench Pro: Claude Opus 4.7 leads GPT-5.5

In OpenAI’s GPT-5.5 evaluation table, Claude Opus 4.7 scores 64.3% on SWE-Bench Pro, ahead of GPT-5.5 at 58.6%. This is one of the strongest comparisons in the evidence set because both numbers appear in the same OpenAI table.

Kilo AI’s third-party article also reports Kimi K2.6 at 58.6% on SWE-Bench Pro. That makes Kimi look competitive with GPT-5.5 on this metric, but the score does not appear alongside GPT-5.5 and Claude Opus 4.7 in the same citable cross-model table, so it is best treated as a useful signal rather than a strict head-to-head result.

Terminal-Bench 2.0: GPT-5.5 leads Claude Opus 4.7

The same OpenAI table gives GPT-5.5 an 82.7% score on Terminal-Bench 2.0, versus 69.4% for Claude Opus 4.7. If your product relies heavily on command-line-style workflows or terminal-based coding agents, GPT-5.5 should be high on the test list.

The provided sources do not include Terminal-Bench 2.0 scores for Kimi K2.6 or DeepSeek V4, so this benchmark cannot support a four-model ranking.

SWE-Bench Verified: useful, but not the same evidence tier

Claude Opus 4.7’s 87.6% SWE-Bench Verified score appears in third-party benchmark summaries; Verdent identifies it as Anthropic-conducted and says memorization screens were applied. Kimi K2.6’s 80.2% SWE-Bench Verified score comes from a Kilo AI third-party article.

Both figures are worth noting. They are not as clean for direct comparison as the GPT-5.5-versus-Claude numbers that appear together in OpenAI’s table.

Expert-SWE: do not use it as the ranking anchor

OpenAI reports GPT-5.5 at 73.1% on Expert-SWE (Internal), but the same table labels it as an internal evaluation and notes that labs have seen evidence of memorization on this eval. That makes it a useful OpenAI-side signal, not a solid basis for a four-model leaderboard.

Practical model selection

For GitHub-issue-style software engineering, test Claude Opus 4.7 early. On the cleanest shared SWE-Bench Pro data here, Claude Opus 4.7 scores 64.3% against GPT-5.5’s 58.6%. Vellum frames this comparison around real GitHub issue resolution.

For terminal-heavy coding agents, test GPT-5.5 early. GPT-5.5’s cited Terminal-Bench 2.0 score is 82.7%, above Claude Opus 4.7’s 69.4%. That does not mean GPT-5.5 is better for every coding task. It means it has a clear advantage on this specific benchmark.

For open-model evaluation, include Kimi K2.6 but rerun your own tests. Hugging Face describes Kimi K2.6 as an open-source, native multimodal agentic model, and Kilo AI reports 58.6% on SWE-Bench Pro and 80.2% on SWE-Bench Verified. Because those numbers are not shown in the same official comparison table as GPT-5.5 and Claude Opus 4.7, use them to build a shortlist, not to declare a winner.

For DeepSeek V4, wait for citable benchmarks or run your own suite. The provided sources do not contain verifiable DeepSeek V4 benchmark scores. Leaving the cells blank is more accurate than inventing a rank.

Why there is no overall champion

LLM benchmark comparisons often go wrong when scores from different sources, harnesses and task types are mixed into one headline. A safer reading uses three evidence tiers:

1. Shared benchmark table: GPT-5.5 and Claude Opus 4.7 on OpenAI’s SWE-Bench Pro and Terminal-Bench 2.0 table. This is the best apples-to-apples subset in the provided evidence.
2. Third-party summaries of vendor or partner evaluations: Claude Opus 4.7’s SWE-Bench Verified and CursorBench numbers are useful, but the source and eval setup matter.
3. Limited or missing cross-verification: Kimi K2.6 currently relies mainly on third-party SWE-Bench numbers in the provided evidence, while DeepSeek V4 lacks enough citable benchmark data here.

The defensible conclusion is narrower than a single ranking: Claude Opus 4.7 leads GPT-5.5 on SWE-Bench Pro, GPT-5.5 leads Claude Opus 4.7 on Terminal-Bench 2.0, Kimi K2.6’s SWE-Bench results look competitive but have a lower evidence level, and DeepSeek V4 should be marked as insufficient data for now.

For production decisions, treat this table as a starting shortlist. The next step is to rerun the contenders on your own repositories, languages, tests, tool calls, latency targets, cost limits and recovery paths. That will tell you more than a blended four-model ranking.

The benchmarks worth separating

SWE-Bench Pro: Claude Opus 4.7 leads GPT-5.5

In OpenAI’s GPT-5.5 evaluation table, Claude Opus 4.7 scores 64.3% on SWE-Bench Pro, ahead of GPT-5.5 at 58.6%. This is one of the strongest comparisons in the evidence set because both numbers appear in the same OpenAI table.

Kilo AI’s third-party article also reports Kimi K2.6 at 58.6% on SWE-Bench Pro. That makes Kimi look competitive with GPT-5.5 on this metric, but the score does not appear alongside GPT-5.5 and Claude Opus 4.7 in the same citable cross-model table, so it is best treated as a useful signal rather than a strict head-to-head result.

Terminal-Bench 2.0: GPT-5.5 leads Claude Opus 4.7

The same OpenAI table gives GPT-5.5 an 82.7% score on Terminal-Bench 2.0, versus 69.4% for Claude Opus 4.7. If your product relies heavily on command-line-style workflows or terminal-based coding agents, GPT-5.5 should be high on the test list.

The provided sources do not include Terminal-Bench 2.0 scores for Kimi K2.6 or DeepSeek V4, so this benchmark cannot support a four-model ranking.

SWE-Bench Verified: useful, but not the same evidence tier

Claude Opus 4.7’s 87.6% SWE-Bench Verified score appears in third-party benchmark summaries; Verdent identifies it as Anthropic-conducted and says memorization screens were applied. Kimi K2.6’s 80.2% SWE-Bench Verified score comes from a Kilo AI third-party article.

Both figures are worth noting. They are not as clean for direct comparison as the GPT-5.5-versus-Claude numbers that appear together in OpenAI’s table.

Expert-SWE: do not use it as the ranking anchor

OpenAI reports GPT-5.5 at 73.1% on Expert-SWE (Internal), but the same table labels it as an internal evaluation and notes that labs have seen evidence of memorization on this eval. That makes it a useful OpenAI-side signal, not a solid basis for a four-model leaderboard.

Practical model selection

For GitHub-issue-style software engineering, test Claude Opus 4.7 early. On the cleanest shared SWE-Bench Pro data here, Claude Opus 4.7 scores 64.3% against GPT-5.5’s 58.6%. Vellum frames this comparison around real GitHub issue resolution.

For terminal-heavy coding agents, test GPT-5.5 early. GPT-5.5’s cited Terminal-Bench 2.0 score is 82.7%, above Claude Opus 4.7’s 69.4%. That does not mean GPT-5.5 is better for every coding task. It means it has a clear advantage on this specific benchmark.

For open-model evaluation, include Kimi K2.6 but rerun your own tests. Hugging Face describes Kimi K2.6 as an open-source, native multimodal agentic model, and Kilo AI reports 58.6% on SWE-Bench Pro and 80.2% on SWE-Bench Verified. Because those numbers are not shown in the same official comparison table as GPT-5.5 and Claude Opus 4.7, use them to build a shortlist, not to declare a winner.

For DeepSeek V4, wait for citable benchmarks or run your own suite. The provided sources do not contain verifiable DeepSeek V4 benchmark scores. Leaving the cells blank is more accurate than inventing a rank.

Why there is no overall champion

LLM benchmark comparisons often go wrong when scores from different sources, harnesses and task types are mixed into one headline. A safer reading uses three evidence tiers:

1. Shared benchmark table: GPT-5.5 and Claude Opus 4.7 on OpenAI’s SWE-Bench Pro and Terminal-Bench 2.0 table. This is the best apples-to-apples subset in the provided evidence.
2. Third-party summaries of vendor or partner evaluations: Claude Opus 4.7’s SWE-Bench Verified and CursorBench numbers are useful, but the source and eval setup matter.
3. Limited or missing cross-verification: Kimi K2.6 currently relies mainly on third-party SWE-Bench numbers in the provided evidence, while DeepSeek V4 lacks enough citable benchmark data here.

The defensible conclusion is narrower than a single ranking: Claude Opus 4.7 leads GPT-5.5 on SWE-Bench Pro, GPT-5.5 leads Claude Opus 4.7 on Terminal-Bench 2.0, Kimi K2.6’s SWE-Bench results look competitive but have a lower evidence level, and DeepSeek V4 should be marked as insufficient data for now.

For production decisions, treat this table as a starting shortlist. The next step is to rerun the contenders on your own repositories, languages, tests, tool calls, latency targets, cost limits and recovery paths. That will tell you more than a blended four-model ranking.