Frontier AI Benchmarks Compared: GPT‑5.5, Claude Opus 4.7, Gemini 3.5 Flash, Grok 4.3, and DeepSeek V4

The model also performs strongly on knowledge‑work benchmarks: GDPval evaluates tasks across dozens of professions including law, finance, and product management, where GPT‑5.5 matches or beats professionals in roughly 84.9% of comparisons.

Together, these numbers suggest GPT‑5.5 is particularly strong at autonomous multi‑step tasks and agentic workflows.

Claude Opus 4.7: standout model for coding benchmarks

Anthropic’s Claude Opus 4.7 is widely regarded as one of the strongest models for software engineering tasks.

Its most prominent benchmark results include:

• SWE‑bench Verified: 87.6%
• SWE‑bench Pro: 64.3%
• GPQA Diamond: 94.2%
• Terminal‑Bench 2.0: 69.4%

SWE‑bench evaluates whether a model can fix real bugs in open‑source repositories. Opus 4.7 resolving 87.6% of SWE‑bench Verified tasks represents a significant improvement over its predecessor and places it among the best models for coding agents.

While its Terminal‑Bench score trails GPT‑5.5, its coding‑centric benchmarks remain among the strongest reported in public comparisons.

Gemini 3.5 Flash: high performance at much higher speed

Google’s Gemini 3.5 Flash is unusual because it is positioned as a fast, cost‑efficient model rather than a flagship — yet it still posts competitive results on several agentic benchmarks.

Reported results include:

• Terminal‑Bench 2.1: 76.2%
• MCP Atlas: 83.6%
• GDPval‑AA: 1656 Elo
• CharXiv reasoning: 84.2%

Google says the model runs roughly four times faster than comparable frontier models while outperforming the earlier Gemini 3.1 Pro on several agent and coding benchmarks.

In practice, Gemini 3.5 Flash’s main strength is the speed‑to‑capability trade‑off: it delivers near‑flagship benchmark results while targeting low latency and production workloads.

DeepSeek V4: powerful open‑weight challenger

DeepSeek V4 is notable because it is one of the most capable open‑weight frontier models released so far.

The model family includes two variants:

• V4‑Pro (large mixture‑of‑experts model)
• V4‑Flash (lighter and more efficient variant)

According to the model’s technical reporting and summaries of its benchmarks, V4‑Pro in maximum reasoning mode achieves:

• SWE‑bench Verified: 80.6%
• GPQA Diamond: 90.1%
• LiveCodeBench: 93.5%

These results place it close to leading proprietary models in some coding benchmarks.

However, an independent evaluation by the U.S. National Institute of Standards and Technology’s CAISI program found the model’s capabilities lag the frontier by roughly eight months, highlighting a gap between self‑reported and independent results.

Grok 4.3: strong improvements but not a clear frontier leader

xAI’s Grok 4.3 represents a large improvement over earlier Grok models, especially on agentic task benchmarks.

Published figures include:

• Artificial Analysis Intelligence Index: ~53
• GDPval‑AA: ~1500 Elo
• τ²‑Bench Telecom: 98%
• IFBench: 81%

The jump of more than 300 Elo on GDPval‑AA compared with earlier Grok versions suggests substantial progress on real‑world task automation.

Still, third‑party analyses generally place the model below the newest OpenAI and Anthropic systems on overall capability benchmarks.

What the benchmark evidence suggests overall

Looking across these evaluations, a consistent pattern emerges:

• Best overall agentic workflows: GPT‑5.5
• Best coding‑focused model: Claude Opus 4.7
• Best speed‑to‑capability balance: Gemini 3.5 Flash
• Best open‑weight contender: DeepSeek V4 Pro
• Improved but mid‑tier frontier model: Grok 4.3

However, these conclusions should be treated as directional rather than definitive because each vendor highlights different evaluation suites.

Why frontier model rankings remain unstable

Benchmark comparisons in modern AI are increasingly complicated for several reasons:

1. Different benchmark versions. For example, Terminal‑Bench 2.0 and 2.1 measure similar tasks but produce slightly different results.
2. Reasoning effort settings. Many models now have configurable "thinking" modes that change scores.
3. Vendor‑reported numbers. Many benchmark tables originate from model launch posts rather than independent evaluations.
4. Task specialization. Coding, agentic work, and knowledge‑work benchmarks reward different capabilities.

Because of these factors, the true relative ranking of frontier models often becomes clearer only after months of independent testing.

The bottom line

The latest benchmark evidence does not show a single model dominating every domain.

Instead, the current frontier appears specialized:

• GPT‑5.5 leads many agentic workflow benchmarks.
• Claude Opus 4.7 excels in real‑world coding tasks.
• Gemini 3.5 Flash pushes the speed‑performance frontier.
• DeepSeek V4 demonstrates that open‑weight models can approach proprietary capabilities.
• Grok 4.3 shows significant improvement but still trails the newest leaders in most broad evaluations.

As independent benchmarks converge and more apples‑to‑apples testing emerges, the exact ordering of these systems will likely continue to evolve.