Gemini leads on ARC-AGI-2 and GPQA Diamond. Claude dominates GDPval-AA. GPT-5.3-Codex sweeps Terminal-Bench. Every major lab has a benchmark they can claim to "win." But in 2026, the most important number in AI might be a 48-percentage-point gap between what one vendor reported on the HLE and what independent evaluators found.
If you follow AI news in June 2026, you'll see a pattern. Google DeepMind publishes a press release: "Gemini 3.1 Pro Sets New State-of-the-Art on ARC-AGI-2." Anthropic fires back days later: "Claude Opus 4.6 Achieves Breakthrough Reasoning on GDPval-AA." OpenAI, quiet for a moment, drops GPT-5.3-Codex — and suddenly it owns every terminal-benchmark leaderboard.
This is the Benchmark Wars of 2026, and every lab is winning. But here's the uncomfortable truth: they can't all be right — not simultaneously, not in a way that lets you make an informed buying decision.
The problem is not that benchmarks are useless. It's that they're optimized for selectively. Every lab designs evaluation protocols, picks test splits, and — in some cases — reports results under conditions that independent evaluators cannot reproduce. The result is a landscape where every model looks like a champion, until you zoom out and see the whole picture.
Let's lay out the data — all of it, in one place. This is the most current snapshot of the major public benchmarks across the Big Three labs (Google DeepMind, Anthropic, and OpenAI) as of late May / early June 2026.
| Benchmark | Gemini 3.1 Pro | Claude Opus 4.6 | GPT-5.3-Codex | Leader |
|---|---|---|---|---|
| ARC-AGI-2 | 77.1% | 68.8% | — | Gemini (+8.3pp) |
| GPQA Diamond | 94.3% | 91.3% | — | Gemini (+3.0pp) |
| SWE-bench | 80.8% | 80.6% | — | Statistical Tie |
| HLE* | — | 66.6% (claimed) 18.6% (actual) |
— | ⚠️ Controversy |
| Terminal-Bench 2.0 | — | — | 77.3% | GPT-5.3-Codex |
| CyberSec CTF | — | — | 77.6% | GPT-5.3-Codex |
| GDPval-AA | 1,317 | 1,606 | — | Claude (+289) |
| MCP Atlas | 69.2% | 59.5% | — | Gemini (+9.7pp) |
The first thing you notice: OpenAI doesn't even publish on several of the benchmarks Gemini and Claude compete on — and vice versa. GPT-5.3-Codex appears only on Terminal-Bench 2.0 and CyberSec CTF, where it dominates. Gemini appears only where it leads. Claude shows up everywhere, but with mixed results.
This is cherry-picking by absence. When a lab knows it doesn't lead on a benchmark, it simply doesn't publish. The public never sees the full cross-comparison — and that's by design.
ARC-AGI-2 is the sequel to François Chollet's Abstraction and Reasoning Corpus, specifically designed to resist memorization. Unlike its predecessor, which models eventually saturated, ARC-AGI-2 uses entirely new grid-puzzle formats that reward genuine abstraction rather than pattern matching from training data.
On paper, ARC-AGI-2 is the cleanest test of general intelligence available. Gemini 3.1 Pro's 77.1% against Claude's 68.8% looks decisive — a near 10-point gap.
But even ARC-AGI-2 shows architecture-specific advantages. Gemini's mixture-of-experts (MoE) routing and its massive vision encoder give it a structural edge on the visual pattern-matching tasks that dominate ARC-AGI-2. Claude's architecture, optimized for long-context text reasoning, struggles with the visual abstraction format even when its underlying reasoning capability is comparable.
ARC-AGI-2 was supposed to be the benchmark that couldn't be gamed. And it's perhaps the most resistant to overt manipulation. But it still rewards one architecture over another — and once a lab realizes this, they can tune their training pipeline to favor ARC-AGI-2's specific puzzle types. The benchmark arms race is never truly over; it just gets more sophisticated.
Google DeepMind's 94.3% on GPQA Diamond — a graduate-level biology, physics, and chemistry QA benchmark — is genuinely impressive. Claude's 91.3% is also strong, but the 3-point gap is statistically significant at this level of difficulty.
However, GPQA Diamond measures knowledge recall + multi-step reasoning — which plays directly to Gemini's strengths. Its training data includes an enormous corpus of scientific papers (Google owns DeepMind, which has access to Google Scholar, PubMed, and a massive indexing of the scientific literature). Claude has strong training data too, but it doesn't have Google's corpus.
The lesson: benchmarks that reward training data breadth will always favor the lab with the biggest corpus. That's not "intelligence" — that's infrastructure.
If there's one story from the 2026 benchmark wars that changes how you should read every other number in this article, it's the HLE (Humanity's Last Exam) controversy.
What Anthropic reported: Claude Opus 4.6 scored 66.6% on the HLE benchmark.
What independent evaluators found: When they ran the identical benchmark under controlled conditions, Claude scored just 18.6%.
That's a 48-percentage-point gap between vendor-reported and independently verified performance.
The discrepancy is too large to be explained by "implementation differences" or "prompt variation." It suggests fundamental differences in evaluation methodology, test set handling, scoring criteria, or disclosure practices.
The HLE is a particularly important benchmark because it was designed as a crowd-sourced collection of extremely difficult questions spanning mathematics, science, and reasoning. A score of 66.6% would put Claude near the top of the leaderboard, competitive with expert human performance. A score of 18.6% tells a very different story — one of a model that is struggling with genuinely hard problems.
Who is right? Without full transparency into both evaluation protocols, we cannot know. And that's exactly the point: in 2026, vendor-reported AI benchmark scores are only as trustworthy as the methodology behind them, and there is no independent auditing body ensuring consistency.
"The HLE discrepancy is not a minor error bar. It's a 48-percentage-point canyon. If one lab's internal evaluation can differ from independent evaluation by this magnitude, then every self-reported benchmark score should carry a massive asterisk." — Apick Analysis
SWE-bench (Software Engineering Benchmark) evaluates models on their ability to resolve real GitHub issues by generating patches. Gemini's 80.8% and Claude's 80.6% — a difference of 0.2 percentage points — is a statistical dead heat. Any responsible analysis would call this a tie.
Yet you can bet that whichever lab publishes a press release next week will frame "their" number (depending on which decimal place they round to) as a victory. In PR terms, 80.8% beats 80.6% — even though the margin is smaller than the benchmark's own inter-run variance.
This is the Nielsen effect in AI benchmarking: when scores are within the noise floor, whichever lab has the better PR team "wins." The data doesn't support a winner, but the headlines will create one anyway.
MCP (Model Context Protocol) Atlas tests models on their ability to navigate and reason across tool-use scenarios — API calls, file operations, multi-step agent loops. Gemini's 69.2% comfortably beats Claude's 59.5%.
But MCP Atlas is a Google-developed benchmark, testing abilities that are uniquely important to Google's product ecosystem (Agentic tool use across Google Workspace, APIs, and cloud services). That doesn't invalidate the result — but it means the benchmark measures a skillset that Google has specifically optimized for. Claude wasn't trained for the same tool-use paradigm, and its lower score reflects an architectural and training priority difference, not a gap in "general ability."
GDPval-AA (a benchmark for general decision-process validation) is where Claude Opus 4.6 shines. A score of 1,606 against Gemini's 1,317 — a 289-point advantage — is Claude's most decisive win in this comparison.
GDPval-AA measures long-context reasoning with complex constraint satisfaction — a domain where Claude's architecture (designed around lengthy deliberative chains) naturally excels. If your use case involves multi-hour analysis sessions, complex document reasoning, or planning with many interlocking constraints, Claude's GDPval-AA lead is directly relevant.
But notice: OpenAI doesn't publish on GDPval-AA at all. Neither do several other labs. The benchmark lacks cross-model coverage, which means we're comparing a subset of the field — and comparing a model (Claude) on a benchmark it was arguably designed to dominate.
While the Gemini-vs-Claude narrative dominates mainstream AI coverage, OpenAI's GPT-5.3-Codex has quietly taken command of the code execution and cybersecurity benchmark category:
These results are significant because they measure agentic execution — not just text generation. GPT-5.3-Codex appears to be genuinely superior at navigating live environments, running commands, and adapting to dynamic outcomes. For developers building AI-powered automation pipelines, these benchmarks matter more than ARC-AGI-2 ever will.
But again: OpenAI doesn't publish on ARC-AGI-2 or GPQA Diamond. Is GPT-5.3-Codex worse than Gemini and Claude on those general reasoning benchmarks? We don't know — because OpenAI doesn't release the numbers. The silence tells you everything.
Let's step back and identify the pattern. Looking across all eight benchmarks in 2026:
| Benchmark | What It Measures | Who It Favors | Why |
|---|---|---|---|
| ARC-AGI-2 | Visual abstraction, pattern reasoning | Gemini | MoE + vision encoder advantage on visual puzzle formats |
| GPQA Diamond | Graduate-level science QA | Gemini | DeepMind's massive scientific corpus advantage |
| SWE-bench | Software engineering patch generation | Tie | Both optimized for code; noise-floor difference |
| GDPval-AA | Long-context reasoning, constraint satisfaction | Claude | Deliberative architecture, long-context optimization |
| MCP Atlas | Tool-use, agentic API calls | Gemini | Google-developed, tied to Workspace tool ecosystem |
| Terminal-Bench 2.0 | CLI execution, agentic workflows | GPT-5.3-Codex | Codex lineage, code-execution focused training |
| CyberSec CTF | Cybersecurity real-world challenges | GPT-5.3-Codex | Safety/security tuning provided domain expertise |
The pattern is stark: every benchmark has a winner, and the winner is almost always the lab whose architecture and training priorities align most closely with that benchmark's format.
This is not fraud. This is expected behavior in an unregulated evaluation landscape. If your job depends on showing leadership, you will naturally:
The HLE scandal proves that the gap between vendor-reported and independently verified results can be 48 percentage points. If that level of discrepancy is possible, it casts doubt on every self-reported number — not because labs are lying, but because there is no shared standard for what "administering a benchmark" means.
Stop buying models based on headline benchmark scores. The single number in a press release has been optimized, curated, and selectively presented to make one lab look good. Instead:
Don't trust the leaderboards. The model that wins SWE-bench may perform worse on your actual codebase. The model that wins ARC-AGI-2 may fail at your data pipeline. The only reliable test is the one you run yourself.
Watch for benchmark-specific overfitting. As training pipelines increasingly optimize for specific benchmarks, the risk grows that models will memorize benchmark patterns rather than develop general capabilities. The gap between benchmark performance and real-world performance will widen.
Beware the benchmark narrative trap. A lab that leads ARC-AGI-2 is not "winning AI" — it's winning on a specific metric that its architecture was designed to favor. Before investing, demand to see:
The 2026 benchmark wars have exposed structural problems that won't fix themselves:
1. We need an independent benchmarking body. Just as the EIAudit or MLCommons provides standardized MLPerf results for training and inference, we need an independent organization that administers evaluation benchmarks under controlled, transparent conditions across all major labs. The HLE scandal is a direct consequence of the absence of such a body.
2. We need methodology transparency as a requirement. Every published benchmark result should come with: exact prompt templates, sampling parameters, test set provenance, inter-run variance, and scoring rubric. Without this, the number is a press release, not a scientific result.
3. We need cross-benchmark reporting. Labs should be encouraged (or required) to publish on a standardized basket of benchmarks — not just the ones they win. The selective absence of results is as informative as the results themselves.
4. We need to separate "benchmark performance" from "capability." A model can score 94.3% on GPQA Diamond and still fail at simple tasks in production. Benchmarks measure specific, narrow abilities under ideal conditions. They are not holistic evaluations of AI capability.
"ARC-AGI-2 was supposed to resist memorization. GPQA Diamond was supposed to test graduate-level reasoning. Terminal-Bench 2.0 was supposed to measure real execution. They all do — but they also all favor one architecture or lab over another. The benchmark is never neutral." — Apick Analysis
| If You Care About... | Choose | Score to Beat |
|---|---|---|
| Visual abstraction / puzzle reasoning | Gemini 3.1 Pro | 77.1% ARC-AGI-2 |
| Graduate-level science QA | Gemini 3.1 Pro | 94.3% GPQA Diamond |
| Software engineering (code generation) | Tie | ~80.7% SWE-bench |
| Long-context constraint reasoning | Claude Opus 4.6 | 1,606 GDPval-AA |
| Agentic tool use / API workflows | Gemini 3.1 Pro | 69.2% MCP Atlas |
| Terminal / CLI automation | GPT-5.3-Codex | 77.3% Terminal-Bench 2.0 |
| Cybersecurity / CTF challenges | GPT-5.3-Codex | 77.6% CyberSec CTF |
The benchmark wars of 2026 are not going away. They will get worse before they get better — because as long as benchmark leadership translates into market share, funding, and talent acquisition, every lab has an incentive to game the system, not fix it.
The numbers don't mean what you think. But if you know why — which architecture advantages, which cherry-picked domains, which evaluation protocol choices produce each number — then you can use them as data, not as dogma.
And in 2026, that distinction might be the most important benchmark of all.