Independent AI Research

“

“Enabling tools on SimpleQA Verified results in near perfect performance.”

— Haas et al., SimpleQA Verified (Google DeepMind, 2025) · arXiv:2509.07968

EMPIRICAL REALITY CHECK · 2026

They have tools.
They don’t use them.

31%

GPT-5 search trigger rate
Nectiv, N=8,500+ prompts [3]

<50%

Gemini grounding rate
DEJAN, N=10,000 prompts [12]

88–93%

Hallucination rate when wrong
AA-Omniscience, N=6,000 [4]

DECOMPOSING THE LIE

OpenAI reports GPT-5’s hallucination rate as 9.6%[1] with browsing enabled. But this is a blended average. GPT-5 only searches 31%[3] of the time. With browsing disabled, the rate is 47%[1].

For the 69% of queries where GPT-5 doesn’t search[3], users get the high-error regime — silently, without warning, disguised as the same confident output.

Sources: OpenAI GPT-5 System Card · Nectiv/SearchEngineLand · OpenAI GPT-5.2 Update

Capability ≠ Propensity.

Having tools is not the same as using them. The models can search. They choose not to. And when they choose wrong, they fabricate with confidence.

The Billion Dollar Lie

Tool Availability vs. Tool Usage

Metric	Google / OpenAI	VERITAS
Philosophy	“Let the model decide”	“Force the model”
Architecture	Optional Retrieval (RAG)	Mandatory Pipeline (C1–C6)
Search Trigger Rate	31% (GPT-5)[3] / <50% (Gemini)[12]	100% (hardcoded)
Failure Mode	Hallucination (fabrication)	Refusal (silence)
Cost Incentive	Minimize search (save $)	Maximize truth (scrape everything)
When It Doesn’t Know	Invents an answer (88–93%)[4]	Says “I don’t know” (9% refusals)

The economic incentive is hallucination. Google charges $14–$35 per 1,000 search grounding queries[35]. Every search adds latency and compute cost. RLHF training rewards confident answers over honest refusals[4]. The model’s parametric confidence — even when misplaced — is the cheaper path. Hallucination is not a bug. It’s a cost optimization.

↓ Read the Full Paper (PDF) ↗ 100% Open Source · GitHub

Architecture > Intelligence

Xu et al. (2024) proved hallucination is mathematically inevitable in autoregressive LLMs[26]. You can’t solve it inside the model. So we solve it around it.

STANDARD LLM (GPT-5 / Gemini)

01 User Query

02 “Do I feel confident?” → 69% YES[3]

03 Parametric Memory Recall (47% error)[1]

04 Output (no warning, same confidence)

> Failure: Model decides. Model is wrong. User never knows.

VERITAS (MANDATORY RETRIEVAL)

C1 Primary Scrape (10 sources via Camoufox)

C2–C3 Secondary Scrape + Query Expansion

C4 Synthesis FROM EVIDENCE ONLY

C5–C6 Independent Verification (10 new sources)

> Invariant: NO parametric memory. NO guessing. Evidence or silence.

Intent

Parse Query

C2+C3

Query

Expand + Search

SCRAPE

20 Sources

Camoufox

Synthesis

Evidence Only

C5+C6

Verify

10 New Sources

Model: Gemini 2.5 Flash Lite Cost: ~€0.003/query Latency: ~115s (Ask)

SimpleQA Verified Leaderboard

Google DeepMind · N=1,000 · 47 models evaluated · Kaggle 2025/26

arXiv → Kaggle →

F-SCORE

89.1%

#1 of 47 models

FABRICATION

0.0%

Unmatched on benchmark

ACCURACY

85.0%

85 / 100

ACC | ATTEMPTED

93.4%

85 / 91 attempted

#	Model	F-Score	Fabrication	Cost / 1k Queries

* Fabrication estimates based on AA-Omniscience [5] hallucination rates and browse-off error rates [1]. Veritas fabrication rate: empirically measured (0/100). All competitors evaluated on SimpleQA Verified [10] without forced tool-use (standard parametric API). Veritas enforces architectural verification. This is not cheating — this is architecture.

Cost & Latency: Transparent Numbers

No made-up figures. Here is exactly what Veritas costs, how it’s calculated, and how it compares.

VERITAS BENCHMARK COST (MEASURED)

~€0.003

per Ask query

€1 / ~400 queries

€0.05–0.20

per Deep Research

~500k tokens in, ~50k out

$0.20

100 SimpleQA queries total

Empirically measured

Basis: Gemini 2.5 Flash Lite [35] — $0.075/1M input, $0.30/1M output (Google AI pricing, Feb 2026). 6 LLM calls per query (C1–C6) + 20 web scrapes via Camoufox (free, no API fee).

ASK MODE (SINGLE QUERY)

VERITAS Ask scrapes 20 sources

~115s

GPT-5 (parametric) no search 69%[3]

3–10s[27]

GPT-5 (with Bing) when triggered

15–30s[28]

Gemini (parametric) no grounding >50%[12]

2–8s[30]

o3 (reasoning)

30s–3min[29]

Veritas is slower because it actually searches. The 115s is the cost of 0% fabrication. Fast responses = no verification = hallucination risk.

DEEP RESEARCH MODE

VERITAS Deep Research 15–40 min

ChatGPT Deep Research 5–30 min[32]

Gemini Deep Research 3–15 min

Perplexity Deep Research 2–4 min[33]

All deep research tools take minutes, not seconds. This is expected. The difference: Veritas produces 203k+ chars of sourced academic output for €0.05–0.20.

API TOKEN PRICING (PUBLIC, FEB 2026)

Model	Input $/1M	Output $/1M	+ Search Fee
Gemini 2.5 Flash Lite [35]	$0.075	$0.30	—
GPT-5 [34]	$2.00	$8.00	incl.
Gemini 3 Pro [35]	$1.25	$5.00	+$14/1k grounding[16]
o3 [34]	$10.00	$40.00	—
Claude Opus 4.5 [36]	$15.00	$75.00	—

Sources: Google AI Studio [35], OpenAI API pricing [34], Anthropic API pricing [36]. Prices as of Feb 2026. Subject to change. Veritas uses 6 Flash Lite calls per Ask query (~2k tokens each) = total ~12k tokens. Competitors use 1 call per query (~500 in, ~200 out) but skip verification.

The speed difference IS the accuracy difference. Models that respond in 3 seconds don’t search. Models that search take time. Veritas always searches — that’s why it takes 115 seconds per Ask query and why it never fabricates. 89.1% F-Score with the cheapest model on the market. Architecture beats budget.

We Make Errors. We Don’t Fabricate.

Not all wrong answers are equal. The cause defines the consequence.

TYPE A — FABRICATION

×Model invents without any source data

×No citation, no evidence trail, no audit path

×Post-hoc verification impossible

LIABILITY: Product defect. Undefendable. — VERITAS: 0

TYPE B — MISINTERPRETATION

✓Real sources retrieved and cited

✓Wrong fact extracted from correct source

✓Fully auditable, traceable, fixable

LIABILITY: Data landscape error. Defensible. — VERITAS: 6

$ veritas --query "Last US President born in 18th century?"
> Scraping... 10+10 sources found
> C4 Answer: Millard Fillmore (1800)
> STATUS: WRONG — 1800 is 19th century. Correct: Buchanan (1791)
> TYPE B: Real data, logic error. Source URL traceable. Auditable.

$ veritas --query "Lukacs newspaper 'oppressing classes'?"
> Scraping... sources found: relevant=0
> STATUS: REFUSAL — Insufficient data. No answer generated.
> When we don't know, we say so.

They have tools. They don’t use them.