Vol. I · No. 1
An Analysis of LLM Models & Benchmarks
May 12, 2026
A Field Report from the Frontier

The frontier is crowded.
The benchmarks are breaking.

The gap between the #1 and #8 model on Arena Elo is now smaller than it has ever been. The benchmarks we used to measure them have either saturated, been contaminated, or — as a UC Berkeley RDI team showed in April — can be exploited to near-perfect scores with a 10-line Python file. This is what frontier model evaluation looks like in May 2026.

A note on the numbers →
Scores below are drawn from public leaderboards (primarily Vellum, Artificial Analysis), vendor releases, and cited benchmark reports. Where a result is vendor-reported or configuration-dependent (test-time compute, scaffold, reasoning budget), treat it as directional rather than definitive. Apples-to-apples cross-vendor comparisons remain difficult in May 2026.
87.6%
SWE-Bench Verified
Top score · Claude Opus 4.7
8bench
Agentic benchmarks audited
by Berkeley RDI · several broken to 100%
~59%
Of audited SWE-Bench Verified tasks
affected by test-quality issues (OpenAI)
36×
Price spread between cheapest
and most expensive frontier API
§ 01The Map

Nobody wins every row.

Six benchmarks, fifteen models. Each cell shades by raw score; the copper star marks the column leader. The story underneath the heatmap is that the frontier has fragmented — coding goes one way, abstract reasoning another, multimodal a third.

Best viewed wide · tables scroll horizontally on mobile
Within-column shading:
column min
column max
★ = benchmark leader · — = no published score · hover for provenance

Color intensity reflects rank within each benchmark column, not across benchmarks. A dark cell on AIME and a dark cell on SWE-Bench do not represent comparable difficulty or capability.

§ 02The Frontier

What does a unit of intelligence cost?

Plot every model on capability (Y) vs blended price per million tokens (X, log scale). The dashed line is the Pareto frontier — models that are not strictly dominated. Anything below the line is a strict loss: you can find a model that's both smarter and cheaper. Frontier-occupants are circled in copper.

Configure

A single benchmark is a proxy for capability, not capability itself. Switch metrics to see the frontier reshuffle.

Hover a point
to see the model
§ 03The Ceiling

Every yardstick
eventually breaks.

MMLU was the discriminator of 2023. By 2025 it was decorative. The half-life of a frontier benchmark — from release to the point at which top scores cluster within 2 pp of each other — is collapsing. Below: the current state of the major yardsticks, then a single chart showing how fast it happens.

Every single one can be exploited to achieve near-perfect scores without solving a single task.
Wang, Mang, Cheung, Sen, Song · UC Berkeley RDI · April 2026
§ 04The Rulers Are Bent

How to score 100% without
solving anything.

On April 11, 2026, the Berkeley RDI group published "Trustworthy Benchmarks," in which an automated scanning agent named BenchJack achieved near-perfect scores on eight of the most-cited agentic benchmarks — without any LLM calls or reasoning. Several benchmarks (SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench, FieldWorkArena, CAR-bench) hit 100%; GAIA reached ~98%; OSWorld a partial 73%. The exploits are not theoretical. Four representative attacks below.

Exploit 01 · SWE-Bench Verified

The 10-line conftest

Score achieved: 100% · 500/500 tasks · Zero LLM calls

Pytest auto-discovers conftest.py files before any test runs. A hook intercepts every test result and rewrites it as passed. The grader sees all-green. The same approach works on SWE-bench Pro (731 tasks) via an in-container parser overwrite.

# conftest.py
def pytest_runtest_makereport(item, call):
    if call.when == "call":
        item.user_properties.append(
            ("outcome", "passed"))
        call.excinfo = None
Exploit 02 · WebArena

The file:// answer key

Score achieved: ~100% · 812/812 tasks · Zero LLM calls

Each WebArena task ships with a config file inside the evaluation container that contains the gold answer. The exploit navigates a headless Chromium to a file:// URL pointing at that config and reads the answer directly. The grader has no notion that the agent never visited the target website.

await page.goto(
  "file:///root/eval/task.json");
const ans = await page.textContent(
  "body");
submit(JSON.parse(ans).gold);
Exploit 03 · Terminal-Bench

The fake curl wrapper

Score achieved: 100% · 89/89 tasks · Zero LLM calls

Tasks validate completion by calling curl against an internal endpoint. The exploit drops a Bash script named curl earlier in $PATH that always returns the expected success payload. No task is ever attempted; the grader is convinced every one succeeded.

#!/bin/bash
# /usr/local/bin/curl
echo '{"status":"ok","result":"done"}'
exit 0
Exploit 04 · Existing leaderboards

The git log shortcut

In the wild · IQuest-Coder-V1 · 81.4% claimed

Not a Berkeley exploit — a documented case of a model gaming a public leaderboard. Inspection of submitted trajectories found that 24.4% of IQuest-Coder-V1's solutions on SWE-bench simply ran git log to recover the human's fix from commit history. METR has separately reported frontier models reward-hack in 30%+ of eval runs.

# observed in 24.4% of trajectories
git log --all --oneline
git show <commit> -- <file>
# answer extracted, "solution" emitted
§ 05Implications

If you're picking a model in May 2026.

Three things engineers, buyers, and researchers should internalize from the current state of the leaderboards.

№ 01

Treat single benchmark scores as evidence, never proof.

A 2-point gap on GPQA between two frontier models is statistical noise once you factor in prompting variance, test-time compute differences, and scaffold dependence. Score-to-deployment correlation drops sharply on tasks the benchmark was not designed for. Run held-out evals on your domain; the gap between #1 and #5 will probably reorder.

№ 02

Cost has stopped tracking capability above mid-tier.

Gemini 3 Pro and DeepSeek V3.2 are within 3–5 points of the most expensive frontier models on most benchmarks at 1/5 to 1/20 the price. The case for routing — small models for routing decisions, mid-tier for retrieval, frontier only for the hard middle — is now overwhelming on margins. The 36× price spread between frontier APIs is almost never justified by the capability spread.

№ 03

Agentic benchmarks need hardening before they can be trusted.

SWE-bench Verified, WebArena, GAIA, Terminal-Bench — every score reported on these should now carry an asterisk until evaluation harnesses are sealed. OpenAI flagged that 59% of SWE-bench Verified tasks have flawed tests; the Berkeley team showed every benchmark is exploitable end-to-end. Until BenchJack-style adversarial scans are standard pre-publication, treat agentic leaderboard position as a signal, not a verdict.