Datasets:

Salesteq
/

sara-agent-benchmark

	# We benchmarked 18 models on our sales agent. 16 tied with the one we pay a premium for.

	*A field report from Salesteq — the team that runs Sara, a bilingual (English + Arabic)
	car-sales agent, live for dealerships in Saudi Arabia.*

	---

	We pay real money, every day, for GPT-5.2 to power our production sales agent. So we did
	the uncomfortable experiment: we swapped in 18 other models — the same agent, the same
	prompt, the same tools — and scored every one of them on the actual job.

	Sixteen of them landed in a statistical dead heat with GPT-5.2. One of them costs
	one-sixtieth as much.

	Not because GPT-5.2 is bad. Because the field has quietly gotten very good — good enough
	that, for a well-built agent, the model you pick is no longer a quality decision. It's a
	cost decision. And almost nobody is measuring that on anything that looks like real work.

	## The gap nobody's testing

	If you're about to put a model inside an agent, the leaderboards don't help you. MMLU tells
	you a model can pass exams. BFCL and τ-bench test tool-calling — but in English, on
	synthetic academic or single-industry tasks. None of them tell you what happens when a real
	customer opens your chat widget, in Arabic, on a Friday, and asks whether the BMW X5 is in
	stock and can they book a test drive.

	That last sentence is our production system. So instead of trusting a proxy, we benchmarked
	the models on the real thing.

	## What we actually did

	Sara is our customer-facing agent. Under the hood she runs a "dual-layer" prompt
	assembler — a static, cache-friendly prefix plus one dynamic block — and a real tool set
	(inventory search, lead capture, OTP, booking, human handoff). We took **62 hand-built
	car-sales scenarios** — 49 English, 13 Arabic — drawn from the kinds of conversations Sara
	handles in production, and ran 18 models through the exact same agent.

	Scoring is deliberately boring:

	- A model passes a scenario when an independent judge (Claude-Haiku — a different
	vendor from every model on the board) rates the response complete enough, and the
	agent actually calls the right tools. Talking about booking a test drive doesn't
	count; calling the booking tool does.
	- Cost is computed from real token usage × live OpenRouter pricing, so "quality per
	dollar" is a real number, not a vibe.

	Same agent, same prompt, same tools. Only the model changes.

	## What we found

	Quality is a statistical tie. Run the numbers properly — a two-proportion z-test, not
	eyeballed ranks — and 16 of the 18 models fall into one statistical tier, GPT-5.2 among
	them. DeepSeek-V4-Flash and Qwen3.5-122B lead by point estimate (90.3%), but at 62 scenarios
	you can't say they're significantly better than our GPT-5.2 — or than an 8B open model. The
	raw ranking is inside the noise. The tie is the finding. And remember: the prompt was tuned
	for GPT-5.2, so it walked in with every advantage and still couldn't pull clear.

	So cost is the whole decision. gpt-oss-20B — OpenAI's own open-weight model, small enough
	to self-host — sits in that same top tier at roughly 1/60th the cost per conversation. Not
	60% cheaper. Sixty times. When quality ties, you're only choosing what to pay.

	Arabic separates the field like nothing else. Several models handle Arabic sales
	conversations as well as English (DeepSeek, Qwen3.5-35B, and our own gpt-5.6-luna all hit
	100% on the Arabic set). Others collapse — Llama 3.3 70B dropped to 54% in Arabic while
	scoring 78% in English. If you serve Arabic-speaking customers, the "best model" is a
	different model. Almost no public leaderboard will ever tell you that.

	Security mostly held — and that's about our agent as much as the models. On a
	90-scenario adversarial probe (jailbreak, PII-exfiltration, prompt-injection), most models
	resisted every attack; only the smallest open models leaked a little. When 15 different
	models all refuse the same attacks, that's the security layer doing its job, not luck.

	And we ran one model off the cloud entirely. Qwen3.5-4B on a MacBook (Apple M5) — no
	API, no bill — scored 77.4% with balanced English and Arabic. A capable sales agent, a
	4-billion-parameter model, on a laptop. That's the self-hosting future in one data point.

	## What it means if you're building an agent

	- You probably don't need a frontier model. For a well-scoped agent with a good prompt,
	a self-hostable open model can match a frontier closed one — and the cost difference is
	large enough to change your business model, not just your bill.
	- Benchmark on your own task. A model's MMLU score won't predict whether it calls your
	booking tool correctly. Ours didn't.
	- If you're bilingual, test bilingual. The English ranking and the Arabic ranking are
	not the same ranking.

	## Scope, and what's next

	Every number here comes from a live production agent — real tools, tool-call-verified
	scoring, two languages, cost from live market pricing. We're deliberate about what it covers:

	- We score the decision turn. Each scenario is the highest-signal moment — the
	tool-grounded response, in full context — where a car sale is actually won or lost.
	Multi-turn booking rollouts are the next axis we're adding.
	- Curated, not crawled. 62 scenarios hand-built by the people who run Sara in production —
	search, booking, OTP, handoff, financing, cash-price, identity — in English and Arabic.
	Depth of real coverage over synthetic volume. We report statistical tiers (two-proportion
	z-test) with Wilson CIs, not exact ranks; the Arabic set is growing, so treat it as an early
	signal.
	- Neutral judging, checked. A third-vendor judge (Claude-Haiku) scores every model — no
	home-team advantage. And we validated it: an independent stronger judge (Claude-Sonnet)
	agreed with it on 80% of a balanced sample (Cohen's κ=0.58). Real judge noise exists —
	another reason to read tiers, not ranks — and a multi-judge panel is on the roadmap.
	- Measured on the real agent, not a proxy. That's the entire point, and why it's a domain
	benchmark rather than a universal capability score.
	- Cost is uncached list price — real deployments with prompt caching pay less, which is
	exactly why the dual-layer design exists.

	The security column samples a 5,000-scenario adversarial corpus, run across every model.
	(Two dedicated guard models — Llama-Guard-4, gpt-oss-safeguard — scored a perfect 100%, but
	guards are classifiers, not agents, so that's expected, not comparable.) Next on the roadmap:
	multi-turn rollouts, a larger Arabic set, and more self-hosted models — LFM2.5-2.6B is queued
	(its laptop runtime isn't stable yet).

	## Open method, private data, living board

	The methodology and a representative fixture sample are public. The full 62-fixture set stays
	private, on purpose — an open test set is a test set that gets gamed. This board will keep
	moving as new models ship.

	Thanks to the teams shipping open weights that made this interesting — Qwen, DeepSeek, Zhipu,
	Mistral, Google, Meta, Moonshot, and OpenAI's open-weight line. You made our own model work
	to keep its seat.

	→ Explore the interactive leaderboard:
	[Salesteq/sara-agent-benchmark-leaderboard](https://huggingface.co/spaces/Salesteq/sara-agent-benchmark-leaderboard)
	· Data: [Salesteq/sara-agent-benchmark](https://huggingface.co/datasets/Salesteq/sara-agent-benchmark)

	*Built by [Salesteq](https://salesteq.com). Numbers are a snapshot; see the dataset for the
	current board.*