Benchmark snapshot: the comparable numbers we can cite

A dash means the provided sources do not contain a directly comparable score for that model and row. It does not mean the model is weak at that task. More importantly, these figures do not all come from the same official harness, so they are best treated as a shortlist builder, not a final leaderboard.

Why one overall ranking would be misleading

First, the source layers differ. GPT-5.5’s Terminal-Bench 2.0 and SWE-Bench Pro numbers come from Yahoo Finance / Investing.com reporting OpenAI-provided benchmark results; Claude Opus 4.7’s SWE-Bench Pro, SWE-Bench Verified and Terminal-Bench 2.0 figures come from a secondary roundup that says AWS cites those scores; Kimi K2.6 and DeepSeek V4-Pro figures come from Hugging Face model-card data or related model reporting.

Second, tool access can change the ordering. Mashable’s HLE example is not a GPT-5.5 datapoint, but it shows the methodological problem clearly: without tools, Claude Opus 4.7 scored 46.9% and GPT-5.4 Pro scored 42.7%; with tools, GPT-5.4 Pro scored 58.7% and Claude Opus 4.7 scored 54.7%. Mixing with-tools and without-tools results in one table is a fast way to get the wrong conclusion.

Third, variants and effort settings matter. DeepSeek V4 is split into V4-Pro and V4-Flash, with Yahoo Finance reporting that V4-Flash is positioned as the more efficient and economical version; the detailed scores cited here are for DeepSeek-V4-Pro. Artificial Analysis also separates GPT-5.5 into effort variants and reports that GPT-5.5 xhigh costs about 20% more than its predecessor to run its Index, while coming in 30% cheaper than Claude Opus 4.7 max.

GPT-5.5: strongest signal for terminal workflows

GPT-5.5’s clearest benchmark advantage in this source set is Terminal-Bench 2.0. Yahoo Finance / Investing.com reports OpenAI-provided results showing GPT-5.5 at 82.7% on Terminal-Bench 2.0, a benchmark for command-line workflows, and 58.6% on SWE-Bench Pro, which evaluates GitHub issue resolution.

OpenAI’s long-context table gives more granular evidence. GPT-5.5 scores 73.7 on Graphwalks BFS at 256k and 45.4 at 1M; on Graphwalks parents, it scores 90.1 at 256k and 58.5 at 1M. The same table shows GPT-5.4 at 9.4 on Graphwalks BFS 1M, compared with GPT-5.5’s 45.4.

Third-party analysis also points to efficiency as part of the GPT-5.5 story. Artificial Analysis calls GPT-5.5 a new leading AI model, says OpenAI leads five of its headline evaluations and places second to Gemini 3.1 Pro Preview on three, and reports that GPT-5.5 xhigh uses about 40% fewer output tokens than its predecessor when running its Index.

Test GPT-5.5 first for: CLI automation, terminal agents, long-context retrieval and agentic coding workflows where output-token use matters.

Claude Opus 4.7: coding scores plus the clearest vision evidence

Claude Opus 4.7 has the strongest official evidence for vision-heavy and computer-use work. Anthropic’s API documentation says the update should unlock performance gains on vision-heavy workloads, especially computer use, screenshots, artifacts and document-understanding workflows. The same documentation says coordinate mapping is simpler because model coordinates map 1:1 to actual pixels.

Anthropic’s launch page cites XBOW’s visual-acuity benchmark, where Claude Opus 4.7 is reported at 98.5% versus 54.5% for Opus 4.6. That gives Claude the clearest direct support among these four models for screenshot understanding, document layout analysis, desktop UI navigation and computer-use agents.

On coding benchmarks, a secondary roundup says AWS cites Claude Opus 4.7 at 64.3% on SWE-Bench Pro, 87.6% on SWE-Bench Verified and 69.4% on Terminal-Bench 2.0. Those figures put Claude ahead in the cited SWE-Bench columns, but because the source is not a shared official benchmark table, serious adoption decisions should still be validated on your own repositories.

There is also a production caveat: Anthropic warns that high-resolution images use more tokens, and recommends downsampling images before sending them to Claude when the extra fidelity is unnecessary.

Test Claude Opus 4.7 first for: GitHub issue repair, coding agents, screenshot and document understanding, computer-use agents and UI tasks that depend on precise pixel coordinates.

Kimi K2.6: a Workers AI candidate for multimodal agents

Cloudflare’s changelog says Moonshot AI’s Kimi K2.6 became available on Workers AI on April 20, 2026, with the model ID @cf/moonshotai/kimi-k2.6, and describes the launch as Day 0 support in partnership with Moonshot AI.

Cloudflare describes Kimi K2.6 as a native multimodal agentic model aimed at long-horizon coding, coding-driven design, proactive autonomous execution and swarm-based task orchestration. It also says the model uses a Mixture-of-Experts architecture with 1T total parameters and 32B active parameters per token.

For public benchmark figures, Kimi K2.6’s Hugging Face model card lists 66.7 on Terminal-Bench 2.0, 58.6 on SWE-Bench Pro and 76.7 on SWE-Bench Multilingual. MarkTechPost separately reports 80.2 on SWE-Bench Verified.

Test Kimi K2.6 first for: teams already deploying on Cloudflare Workers AI, especially those building long-horizon coding agents, coding-driven design tools, multimodal workflows or multi-agent orchestration systems.

DeepSeek V4-Pro: the fullest knowledge and math model card

DeepSeek V4 appears in these sources as V4-Pro and V4-Flash. Yahoo Finance reports DeepSeek’s claim that V4-Pro significantly leads other open-source models on world-knowledge benchmarks and is only slightly behind Google’s Gemini-Pro-3.1; the same report says V4-Flash is positioned as the more efficient and economical option.

DeepSeek-V4-Pro’s Hugging Face model card provides the most complete knowledge, math, coding and terminal score set in this comparison: GPQA Diamond 90.1, GSM8K 92.6, HLE 37.7, MMLU-Pro 87.5, SWE-Bench Pro 55.4, SWE-Bench Verified/Resolved 80.6 and TerminalBench 2.0 67.9.

CNBC reports that DeepSeek says V4 has been optimized for agent tools such as Claude Code and OpenClaw. The same article cites Counterpoint principal AI analyst Wei Sun saying V4’s benchmark profile suggests it could offer excellent agent capability at significantly lower cost.

Test DeepSeek V4-Pro first for: open-model evaluation, knowledge and math benchmarks, agent-tooling cost studies, or local and self-managed deployment analysis where downloadable models matter.

Which model should you shortlist by task?

• Terminal automation and command-line agents: start with GPT-5.5. In the cited data, GPT-5.5 reaches 82.7% on Terminal-Bench 2.0, ahead of Claude Opus 4.7 at 69.4%, DeepSeek V4-Pro at 67.9 and Kimi K2.6 at 66.7.
• Software-engineering repair and SWE-Bench Pro-like work: start with Claude Opus 4.7, then rerun on your own repos. The cited SWE-Bench Pro figures put Claude at 64.3%, GPT-5.5 and Kimi K2.6 both at 58.6, and DeepSeek V4-Pro at 55.4, though Claude’s number comes through a lower-level secondary source.
• Screenshots, document understanding and computer-use: prioritize Claude Opus 4.7. Anthropic directly documents vision-heavy workflow gains, computer-use relevance and 1:1 pixel coordinates, and its launch page cites a 98.5% XBOW visual-acuity result.
• Knowledge, math and the open-model route: put DeepSeek V4-Pro on the shortlist. Its Hugging Face model card lists GPQA Diamond, GSM8K, HLE, MMLU-Pro, SWE-Bench and TerminalBench 2.0 results in one place.
• Multimodal agent workflows on Cloudflare Workers AI: test Kimi K2.6. Cloudflare already exposes Kimi K2.6 on Workers AI and frames it around long-horizon coding, coding-driven design and swarm-based task orchestration.

Before adoption, rerun everything under one set of rules

For an internal decision you can defend, use the same model version or API model ID, context length, tool permissions, reasoning effort, temperature, token budget and scoring harness. Tool access is especially important: the HLE example above shows that with-tools and without-tools results can change the relative ordering.

Cost belongs in the same evaluation as accuracy. Artificial Analysis reports that GPT-5.5 xhigh costs about 20% more than its predecessor on its Index, about 30% less than Claude Opus 4.7 max, and uses about 40% fewer output tokens than its predecessor. Anthropic also warns that high-resolution images increase token usage for Claude. For production agents, latency, tool-call success rate, token consumption and failure recovery often matter as much as a single benchmark percentage.

Bottom line

The most reliable reading is task-based, not rank-based. GPT-5.5 has the strongest cited Terminal-Bench signal; Claude Opus 4.7 has the strongest cited SWE-Bench and vision/computer-use case; DeepSeek V4-Pro has the richest public model-card evidence for knowledge and math; and Kimi K2.6 is a serious candidate for Workers AI multimodal agent workflows. A true overall leaderboard would require all four models to be tested in the same harness, with the same tools, versions and effort settings.

Benchmark snapshot: the comparable numbers we can cite

A dash means the provided sources do not contain a directly comparable score for that model and row. It does not mean the model is weak at that task. More importantly, these figures do not all come from the same official harness, so they are best treated as a shortlist builder, not a final leaderboard.

Why one overall ranking would be misleading

First, the source layers differ. GPT-5.5’s Terminal-Bench 2.0 and SWE-Bench Pro numbers come from Yahoo Finance / Investing.com reporting OpenAI-provided benchmark results; Claude Opus 4.7’s SWE-Bench Pro, SWE-Bench Verified and Terminal-Bench 2.0 figures come from a secondary roundup that says AWS cites those scores; Kimi K2.6 and DeepSeek V4-Pro figures come from Hugging Face model-card data or related model reporting.

Second, tool access can change the ordering. Mashable’s HLE example is not a GPT-5.5 datapoint, but it shows the methodological problem clearly: without tools, Claude Opus 4.7 scored 46.9% and GPT-5.4 Pro scored 42.7%; with tools, GPT-5.4 Pro scored 58.7% and Claude Opus 4.7 scored 54.7%. Mixing with-tools and without-tools results in one table is a fast way to get the wrong conclusion.

Third, variants and effort settings matter. DeepSeek V4 is split into V4-Pro and V4-Flash, with Yahoo Finance reporting that V4-Flash is positioned as the more efficient and economical version; the detailed scores cited here are for DeepSeek-V4-Pro. Artificial Analysis also separates GPT-5.5 into effort variants and reports that GPT-5.5 xhigh costs about 20% more than its predecessor to run its Index, while coming in 30% cheaper than Claude Opus 4.7 max.

GPT-5.5: strongest signal for terminal workflows

GPT-5.5’s clearest benchmark advantage in this source set is Terminal-Bench 2.0. Yahoo Finance / Investing.com reports OpenAI-provided results showing GPT-5.5 at 82.7% on Terminal-Bench 2.0, a benchmark for command-line workflows, and 58.6% on SWE-Bench Pro, which evaluates GitHub issue resolution.

OpenAI’s long-context table gives more granular evidence. GPT-5.5 scores 73.7 on Graphwalks BFS at 256k and 45.4 at 1M; on Graphwalks parents, it scores 90.1 at 256k and 58.5 at 1M. The same table shows GPT-5.4 at 9.4 on Graphwalks BFS 1M, compared with GPT-5.5’s 45.4.

Third-party analysis also points to efficiency as part of the GPT-5.5 story. Artificial Analysis calls GPT-5.5 a new leading AI model, says OpenAI leads five of its headline evaluations and places second to Gemini 3.1 Pro Preview on three, and reports that GPT-5.5 xhigh uses about 40% fewer output tokens than its predecessor when running its Index.

Test GPT-5.5 first for: CLI automation, terminal agents, long-context retrieval and agentic coding workflows where output-token use matters.

Claude Opus 4.7: coding scores plus the clearest vision evidence

Claude Opus 4.7 has the strongest official evidence for vision-heavy and computer-use work. Anthropic’s API documentation says the update should unlock performance gains on vision-heavy workloads, especially computer use, screenshots, artifacts and document-understanding workflows. The same documentation says coordinate mapping is simpler because model coordinates map 1:1 to actual pixels.

Anthropic’s launch page cites XBOW’s visual-acuity benchmark, where Claude Opus 4.7 is reported at 98.5% versus 54.5% for Opus 4.6. That gives Claude the clearest direct support among these four models for screenshot understanding, document layout analysis, desktop UI navigation and computer-use agents.

On coding benchmarks, a secondary roundup says AWS cites Claude Opus 4.7 at 64.3% on SWE-Bench Pro, 87.6% on SWE-Bench Verified and 69.4% on Terminal-Bench 2.0. Those figures put Claude ahead in the cited SWE-Bench columns, but because the source is not a shared official benchmark table, serious adoption decisions should still be validated on your own repositories.

There is also a production caveat: Anthropic warns that high-resolution images use more tokens, and recommends downsampling images before sending them to Claude when the extra fidelity is unnecessary.

Test Claude Opus 4.7 first for: GitHub issue repair, coding agents, screenshot and document understanding, computer-use agents and UI tasks that depend on precise pixel coordinates.

Kimi K2.6: a Workers AI candidate for multimodal agents

Cloudflare’s changelog says Moonshot AI’s Kimi K2.6 became available on Workers AI on April 20, 2026, with the model ID @cf/moonshotai/kimi-k2.6, and describes the launch as Day 0 support in partnership with Moonshot AI.

Cloudflare describes Kimi K2.6 as a native multimodal agentic model aimed at long-horizon coding, coding-driven design, proactive autonomous execution and swarm-based task orchestration. It also says the model uses a Mixture-of-Experts architecture with 1T total parameters and 32B active parameters per token.

For public benchmark figures, Kimi K2.6’s Hugging Face model card lists 66.7 on Terminal-Bench 2.0, 58.6 on SWE-Bench Pro and 76.7 on SWE-Bench Multilingual. MarkTechPost separately reports 80.2 on SWE-Bench Verified.

Test Kimi K2.6 first for: teams already deploying on Cloudflare Workers AI, especially those building long-horizon coding agents, coding-driven design tools, multimodal workflows or multi-agent orchestration systems.

DeepSeek V4-Pro: the fullest knowledge and math model card

DeepSeek V4 appears in these sources as V4-Pro and V4-Flash. Yahoo Finance reports DeepSeek’s claim that V4-Pro significantly leads other open-source models on world-knowledge benchmarks and is only slightly behind Google’s Gemini-Pro-3.1; the same report says V4-Flash is positioned as the more efficient and economical option.

DeepSeek-V4-Pro’s Hugging Face model card provides the most complete knowledge, math, coding and terminal score set in this comparison: GPQA Diamond 90.1, GSM8K 92.6, HLE 37.7, MMLU-Pro 87.5, SWE-Bench Pro 55.4, SWE-Bench Verified/Resolved 80.6 and TerminalBench 2.0 67.9.

CNBC reports that DeepSeek says V4 has been optimized for agent tools such as Claude Code and OpenClaw. The same article cites Counterpoint principal AI analyst Wei Sun saying V4’s benchmark profile suggests it could offer excellent agent capability at significantly lower cost.

Test DeepSeek V4-Pro first for: open-model evaluation, knowledge and math benchmarks, agent-tooling cost studies, or local and self-managed deployment analysis where downloadable models matter.

Which model should you shortlist by task?

• Terminal automation and command-line agents: start with GPT-5.5. In the cited data, GPT-5.5 reaches 82.7% on Terminal-Bench 2.0, ahead of Claude Opus 4.7 at 69.4%, DeepSeek V4-Pro at 67.9 and Kimi K2.6 at 66.7.
• Software-engineering repair and SWE-Bench Pro-like work: start with Claude Opus 4.7, then rerun on your own repos. The cited SWE-Bench Pro figures put Claude at 64.3%, GPT-5.5 and Kimi K2.6 both at 58.6, and DeepSeek V4-Pro at 55.4, though Claude’s number comes through a lower-level secondary source.
• Screenshots, document understanding and computer-use: prioritize Claude Opus 4.7. Anthropic directly documents vision-heavy workflow gains, computer-use relevance and 1:1 pixel coordinates, and its launch page cites a 98.5% XBOW visual-acuity result.
• Knowledge, math and the open-model route: put DeepSeek V4-Pro on the shortlist. Its Hugging Face model card lists GPQA Diamond, GSM8K, HLE, MMLU-Pro, SWE-Bench and TerminalBench 2.0 results in one place.
• Multimodal agent workflows on Cloudflare Workers AI: test Kimi K2.6. Cloudflare already exposes Kimi K2.6 on Workers AI and frames it around long-horizon coding, coding-driven design and swarm-based task orchestration.

Before adoption, rerun everything under one set of rules

For an internal decision you can defend, use the same model version or API model ID, context length, tool permissions, reasoning effort, temperature, token budget and scoring harness. Tool access is especially important: the HLE example above shows that with-tools and without-tools results can change the relative ordering.

Cost belongs in the same evaluation as accuracy. Artificial Analysis reports that GPT-5.5 xhigh costs about 20% more than its predecessor on its Index, about 30% less than Claude Opus 4.7 max, and uses about 40% fewer output tokens than its predecessor. Anthropic also warns that high-resolution images increase token usage for Claude. For production agents, latency, tool-call success rate, token consumption and failure recovery often matter as much as a single benchmark percentage.

Bottom line

The most reliable reading is task-based, not rank-based. GPT-5.5 has the strongest cited Terminal-Bench signal; Claude Opus 4.7 has the strongest cited SWE-Bench and vision/computer-use case; DeepSeek V4-Pro has the richest public model-card evidence for knowledge and math; and Kimi K2.6 is a serious candidate for Workers AI multimodal agent workflows. A true overall leaderboard would require all four models to be tested in the same harness, with the same tools, versions and effort settings.