Salesteq/sara-agent-benchmark

🚗 Sara Open-Model Benchmark

We run a production AI sales agent. We pay a premium for GPT-5.2 to power it. So we tested 19 models on our own agent — and 17 of them — open and closed — are statistically tied with GPT-5.2 on quality (Tier 1), and one of them, gpt-oss-20B, costs ~1/60th as much. We're paying a premium for a lead we don't have.

Everyone is shipping AI agents. Almost nobody knows which model to actually put inside one — because the benchmarks that rank models test trivia (MMLU) or English synthetic tool-calls (BFCL, τ-bench), not a real agent talking to real customers.

We're not an eval lab. We're Salesteq — we run Sara, a bilingual (English + Arabic) car-sales agent live for dealerships in Saudi Arabia. So we benchmarked the models on the real thing: our production dual-layer prompt assembler, the real tool set, and 62 real car-sales scenarios in both languages. Same agent, same prompt, same tools — only the model swaps.

📊 What we found

• Quality is a statistical tie. 17 models — open and closed — land in one top tier (Tier 1); a two-proportion z-test can't separate GPT-5.2 from the open models above it (DeepSeek-V4-Flash leads by point estimate at 90.3%). At n=62 the raw ranking is noise — the tie is the finding. And the prompt was tuned for GPT-5.2, so it had every advantage.
• So cost decides — and it's lopsided. gpt-oss-20B sits in that same top tier at ~1/60th GPT-5.2's cost per conversation. When quality ties, price is the whole decision.
• Arabic separates the field. Some models handle Arabic sales conversations as well as English; others fall off a cliff (worst: Llama 3.3 70B). This is the gap most leaderboards never test.
• Security holds — mostly. Against a 90-scenario adversarial probe (jailbreak, PII-exfil, prompt-injection), most models resist every attack on our agent; the smallest open models leak a little. Sara's security layer travels across models.
• Self-hostable, proven. Qwen3.5-4B (self-host) running on a MacBook (Apple M5) — no cloud — scored 77.4% with balanced English/Arabic. A capable agent, on a laptop.

The full, sortable board — with the quality-vs-cost chart and the EN/AR view — is the interactive Space: Salesteq/sara-agent-benchmark-leaderboard.

🔬 Method (plain)

• One real turn through the live Sara agent (with prior-message context), scored on 62 hand-built car-sales fixtures (49 EN + 13 AR).
• Pass = an LLM judge scores completeness ≥ 0.5 and the agent calls the required tools (tool-recall ≥ 0.5). Judge = Claude-Haiku — a different vendor from every model on the board.
• Cost = real input/output tokens × live OpenRouter pricing. $/1k convos projects 6 turns/conversation, at uncached list price — real deployments with prompt caching pay less (that's the whole point of the dual-layer design).
• Tiers, not ranks. Models are grouped by a two-proportion z-test (α=0.05) with Wilson 95% CIs — read the tier, not the exact position. Latency is excluded (it was measured under concurrent load, so it isn't a clean per-model number).
• Judge, checked. An independent stronger judge (Claude-Sonnet) agreed with the primary Haiku judge on 80.0% of a balanced sample (Cohen's κ=0.58). Real judge noise exists — a multi-judge panel is next.
• Dual-layer only. We report the production prompt architecture, never the deprecated one.

📐 Scope & how to read it

20 models · 62 curated sales scenarios · a 5,000-scenario security corpus · 2 languages. Every model runs through a live production agent — real tools, tool-call-verified scoring, cost from live market pricing. Deliberate design choices, stated plainly:

• We score the decision turn. Each scenario tests the highest-signal moment — the tool-grounded response, with full conversation context — where a car sale is won or lost. Multi-turn booking rollouts are the next axis we're adding.
• Curated, not crawled. 62 scenarios hand-built by the team that runs Sara in production — inventory search, booking, OTP, human handoff, financing, cash-price and identity behavior — across English and Arabic. Depth of real coverage over synthetic volume. Read scores within a few points of each other as a tier, not an exact rank; the Arabic set is expanding, so treat AR as an early signal.
• Neutral judge. A third-vendor model (Claude-Haiku) scores every entry — no home-team advantage. A multi-judge panel is on the roadmap.
• Measured on the real agent, not a proxy. Scores reflect a deployed production system — which is the whole point, and why this is a domain benchmark rather than a universal capability score.
• Security is measured, not promised: the column samples a 5,000-scenario adversarial corpus (jailbreak, PII-exfiltration, prompt-injection). Two dedicated guard models (Llama-Guard-4, gpt-oss-safeguard) were also run and scored 100% — but note guards are classifiers, not agents, so their "perfect" score is expected, not comparable.

A living production eval we maintain and share. The method and a representative fixture sample (fixtures-sample.jsonl) are public; the full set stays private so the benchmark can't be gamed.

📁 Files

• leaderboard.csv / .json / .md — the ranked results.
• models.json, pricing.json — roster + the OpenRouter pricing snapshot used for cost.
• fixtures-sample.jsonl — 8 representative fixtures (of 62).

Read the story → BLOG.md. Built by Salesteq.