← All benchmarks
Pasiflora Benchmark · Professional Work

Any model can pass a quiz. Can it do the job?

Professional-Bench grades real work products — an M&A memo, a strategy call, a legal review, a systems design — against expert rubrics, then puts the same models to work in a live database with tools. Eleven frontier models across five labs, judged on what they produce, not what they recall.

11 models · 5 labs 4 professional domains 110 rubric criteria agentic tier graded on ground truth

Published July 12, 2026 · frontier model versions change fast — read these scores as a snapshot as of this date.

RESULTS

The scoreboard

Mean coverage across four multi-part deliverables, 110 rubric criteria in all. Each criterion is scored met (1), partial (½) or missed (0) by an expert-instructed judge, deliberately strict: a criterion that calls for a computed number is missed if the model only discusses it. Multiple choice couldn't separate these models. Real work spreads them 41 points.

01
Grok 4.5
XAI
95.9%
02
Claude Opus 4.8
ANTHROPIC
94.1%
03
Claude Sonnet 4.6
ANTHROPIC
90.0%
04
GPT-5.2
OPENAI
87.7%
05
GPT-5 mini
OPENAI
85.9%
06
Claude Haiku 4.5
ANTHROPIC
81.4%
07
Gemini 2.5 Pro
GOOGLE
80.9%
08
Gemini 2.5 Flash
GOOGLE
79.1%
09
DeepSeek V3.1
DEEPSEEK
75.0%
10
Qwen3 235B
ALIBABA
72.3%
11
Llama 4 Maverick
META
55.0%
0%255075100%
41-point spread

On plain multiple choice, these eleven models cluster inside ten points. Grade them on real deliverables against a strict rubric and the field opens to forty-one — Grok and Opus clear 94%, Llama lands at 55%. The frontier is decided on the banking task, where a model has to build the leveraged-buyout returns, not describe them — Llama scores 41% there, and only discovers a deal misses its hurdle if it actually runs the numbers.

Where they break

The same runs, cut by deliverable. Overall scores hide the split personalities: Sonnet writes a flawless legal review (98) then face-plants on the strategy case (81); Gemini 2.5 Pro nails the law (95) but botches the banking math (64). The quantitative, must-be-built tasks do the sorting; the knowledge-heavy legal task barely separates anyone.

ModelBanking (LBO)ConsultingLawSystems
Grok 4.596969596
Claude Opus 4.893929893
Claude Sonnet 4.693819888
GPT-5.289839880
GPT-5 mini84799189
Claude Haiku 4.580779671
Gemini 2.5 Pro91739564
Gemini 2.5 Flash80679573
DeepSeek V3.157739673
Qwen3 235B45819371
Llama 4 Maverick41608039
≥95% 88–94% 80–87% <80% Banking 28 · Consulting 26 · Law 28 · Systems 28 criteria
§1

What's on the test

Not questions — the multi-part artifacts a junior professional is handed on day one. Every rubric mixes objective catches (a number that must be computed) with judgment catches (a risk that must be flagged, a trap that must be avoided).

Investment Banking28

Build the LBO model and the IC memo

Sources & uses, a five-year debt-paydown schedule, IRR and MOIC, a downside case, a recommendation.

Trap · at 9× in / 9× out the base case returns ~1.8× / ~12% IRR — below a PE hurdle. You only find that if you actually build the returns.
Management Consulting26

Take a meal-kit brand into corporate catering

Bottoms-up TAM/SAM/SOM, B2B unit economics, build-vs-partner, a phased go-to-market, a go / no-go.

Trap · the B2C acquisition math doesn't transfer — B2B is a sales-led, contract-based, different-operations business.
Startup Law28

Find all ten landmines in a founder's story

Ten distinct issues, ranked by severity, each with remediation and a first-week action list.

Trap · beyond the obvious — an H-1B founder working unauthorized, GPL code in a proprietary build, 1099s who are really employees. Few catch all ten.
Systems Engineering28

Serve an 8GB model at 50k rps, p99 < 100ms

Capacity math, batching, a zero-downtime weekly deploy, multi-region failover, an SLO / error budget.

Trap · size only for steady state and you silently breach SLA at peak; cold-starting 8GB of weights breaks the deploy if replicas take traffic too early.
§2

The agentic tier — and the harness that nearly lied

Then we stopped grading prose and made the models work: each is dropped into a SQLite database it has never seen, given one run_sql tool, and has to investigate — discover the schema, query, revise, and answer six questions with multi-hop joins and a planted data-quality trap. No judge: answers are checked against a reference computed live from the data. The board below is the number we trust — but getting there took building the harness twice.

01
Claude Opus 4.8
ANTHROPIC
100%text harness: 100
01
Claude Sonnet 4.6
ANTHROPIC
100%text harness: 23
01
Claude Haiku 4.5
ANTHROPIC
100%text harness: 7
01
GPT-5.2
OPENAI
100%text harness: 40
01
Gemini 2.5 Flash
GOOGLE
100%text harness: 60
01
Gemini 2.5 Pro
GOOGLE
100%text harness: 83
07
Grok 4.5
XAI
94%text harness: 83
08
Qwen3 235B
ALIBABA
94%text harness: 50
09
GPT-5 mini
OPENAI
72%text harness: 100
10
DeepSeek V3.1
DEEPSEEK
67%text harness: 27
11
Llama 4 Maverick
META
33%text harness: 27
0%255075100%
23 → 100

Our first harness — a text protocol — put Sonnet at 23% and Haiku at 7%. It was lying: the protocol let a model fabricate its own query results (Sonnet invented a region, "West," and answered from it). We rebuilt it on native tool-calling, where the provider's API enforces the turn boundary. That harness had its own bug — it never told thorough models to stop, zeroing Grok and DeepSeek. Only after fixing both did the board converge. The text harness had understated nine of eleven models; the small number under each score above is what it wrongly reported.

§3

Method & honest limits

A benchmark that overstates itself isn't worth publishing. Here is exactly what these runs do and don't establish.

An LLM graded the rubric

The deliverable scores are judge-assigned. A credible board needs expert-authored, expert-validated rubrics — the human-graded work that is Pasiflora's product.

The harness can outweigh the model

On the agentic tier, two harnesses disagreed by up to 93 points on the same model. Only native tool-calling, validated against its own artifacts, is trustworthy.

Too easy at the top

Six models tie at 100% on the fair agentic board. The SQL task separates the mid-tier but not the frontier — it needs harder, longer-horizon tasks.

Averaged, not single-shot

Agentic scores are the mean of multiple runs; GPT-5.2 alone swung 6/6, 2/6, 3/6 before the harness fix. A single agentic run is noise.

§4

The takeaway

The most important result on this page isn't a score. It's the score we almost got wrong.

For frontier models, the hard part of evaluation isn't the models — it's the measurement.

Multiple choice saturates. Strict rubrics separate a little, but the ceiling holds. And the moment you reach for the thing that should be hardest — agentic work with real tools — the harness starts deciding the outcome. We twice built a board that ranked a strong model dead last, and twice it was our wiring, not the model.

Catching that is the whole game. A trustworthy benchmark is native tool-calling per provider, multiple runs, ground-truth grading, and adversarial validation against your own artifacts — before you publish a number. Getting the harness right is unglamorous, and it's exactly where most public benchmarks quietly go wrong.

That's the work Pasiflora does — turning a vetted expert network into evaluations rigorous enough to find the real edge of what these models can do. This bench is the floor of that capability, built in-house to prove the harness. The ceiling is what our experts author.