AI answer audit

We grade the answers AI gives about AI pricing.

Google's AI Overview is what most people actually read when they ask "is X free". We pull that answer via SerpApi, extract its claims, and score each one against official vendor pages with our WASM fact-checker — including when Google gets it right.

answers audited

claims checked

14%

avg official-citation rate

with caveats found

Why this matters

In our reference probe (2026-08-26), Google's answer to "is cursor ide free" cited 7 sources — only 1 was cursor.com (14%). The rest were SEO blogs, a YouTube video and Reddit. Secondary pricing pages rot fast; when they do, the AI answer rots with them — and it outranks the vendor page it barely cited. Every claim below shows its own receipts: what the AI said, which tier of source verified it, and the score.

“best free llm api tier”

2026-08-26 · presence: token · 2 API calls · 6 claims · wasm 76218e28

official citations: 7% CONTAINS_ERRORS

unresolved (judge gap) score 0.1442 · secondary truth	The absolute best free LLM API tier for most developers is Google AI Studio (Gemini Flash family) due to its massive 1-million token context window, 1,500 requests per day limit, and native multimodal support. judge limitation: free-assertion matches page but wording differs (0.14)	Reddit Puter Developer Reddit truth ↗
ground truth unfetchable unfetchable · none truth	However, the "best" tier depends heavily on whether you prioritize raw speed, model variety, privacy, or raw token volume.
ground truth unfetchable unfetchable · none truth	A direct comparison of the top permanent free tiers reveals how they stack up against each other.
contradicted score 0.0452 · secondary truth	If you are building complex AI agents and find yourself constantly bottlenecked by free-tier rate limits, consider using an open-source tool like Free LLM API (available via GitHub or ProductHunt). It acts as a local proxy wrapper	YouTube Product Hunt truth ↗
contradicted score 0.0492 · secondary truth	you feed it your free keys from Google, Groq, and Mistral, and it automatically handles routing, failover, and load balancing so you don't hit a wall.	YouTube Product Hunt truth ↗
ground truth unfetchable unfetchable · none truth	To help tailor the best choice, what specific use case (e.g., coding assistant, heavy data extraction, or a 24/7 personal agent) are you building for? Knowing if you need multimodal features (like reading images) would also help narrow it down.

“cheapest llm api for agents”

2026-08-26 · presence: token · 2 API calls · 16 claims · wasm 76218e28

official citations: 9% CONTAINS_ERRORS

ground truth unfetchable unfetchable · none truth	When building AI agents, you need to look at cost per million tokens, prompt caching efficiency (since agents pass massive, repetitive system prompts), and function-calling capabilities.
unresolved (judge gap) score 0.178 · official truth	The cheapest paid LLM APIs right now fall below $0.15 per million tokens, while multi-model routers offer the best infrastructure for managing agent workloads. judge limitation: amount(s) confirmed in page but wording differs (0.18)	CostGoat BenchLM OpenRouter truth ↗
ground truth unfetchable unfetchable · none truth	If your agent needs to make thousands of rapid, simple micro-steps (like classification, basic extraction, or fast routing), these flash-tier models offer the lowest prices on the market:
ground truth unfetchable unfetchable · none truth	Qwen 3.7 Flash: Currently the price-to-performance leader at $0.03 per 1M input tokens and $0.13 per 1M output tokens . Extremely capable at handling basic instruction-following.
mixed score 0.4542 · secondary truth	DeepSeek V4 Flash: Costs roughly $0.00 to $0.11 per 1M tokens depending on the specific endpoint (like SiliconFlow or DeepSeek direct). It is an incredible choice for structured outputs and basic coding tasks.	Price Per Token SiliconFlow WhatLLM.org truth ↗
ground truth unfetchable unfetchable · none truth	Ling 2.6 / 3.0 Flash: Priced around $0.01 per 1M input tokens, offering high value for raw, high-volume processing.
ground truth unfetchable unfetchable · none truth	If your agents are executing complex workflows, debugging code, or evaluating multi-step logic, standard flash models will fail.
ground truth unfetchable unfetchable · none truth	You need heavier reasoning models that still won’t break the bank:
mixed score 0.3272 · secondary truth	Instead of locking yourself into a single provider, agent developers overwhelmingly use smart routing gateways.	YouTube GonkaRouter truth ↗
contradicted score 0.1685 · secondary truth	They dramatically reduce engineering overhead by letting you swap models via a single OpenAI-compatible SDK:	YouTube GonkaRouter truth ↗
ground truth unfetchable unfetchable · secondary truth	: The gold standard for budget agent development. It offers over 20 completely free models (like Gemma 3 and Qwen variants) with soft rate limits. Paid tiers allow you to compare prices across dozens of host providers instantly to find the cheapest endpoint.	Reddit
contradicted score 0.0405 · secondary truth	Rico / GonkaRouter: Rising specialized aggregators that feature cash-aware routing. They look at a provider’s prompt caching strategy and route your agent's request to the absolute lowest effective cost in real-time.	YouTube truth ↗
ground truth unfetchable unfetchable · none truth	SiliconFlow: An incredibly cheap infrastructure provider hosting open-weight models with massive subsidies, delivering up to 32% lower latency for agent execution loops.
mixed score 0.3777 · secondary truth	Agents are notoriously expensive because they pass the same background context and tools over and over again in a loop.	IntuitionLabs truth ↗
contradicted score 0.2512 · secondary truth	To keep costs at near-zero, ensure you route through a provider that supports Prompt Caching (like DeepSeek or OpenRouter endpoints).	IntuitionLabs truth ↗
contradicted score 0.0595 · secondary truth	A cache-hit typically slashes input token costs by up to 90%, saving far more money than merely picking a cheaper model .	IntuitionLabs truth ↗

“claude api pricing”

2026-08-26 · presence: embedded · 1 API call · 9 claims · wasm 76218e28

official citations: 0% NOT_AUDITABLE

ground truth unfetchable unfetchable · secondary truth	Claude API pricing is pay-as-you-go and billed per million tokens (MTok) for input and output.	g2.com
ground truth unfetchable unfetchable · none truth	Claude Haiku 4.5: $1.00 per million input tokens / $5.00 per million output tokens
ground truth unfetchable unfetchable · none truth	Claude Sonnet 5: $2.00 per million input tokens / $10.00 per million output tokens (temporary rate through August 31, 2026)
ground truth unfetchable unfetchable · none truth	Claude Opus 5 / Opus 4.8: $5.00 per million input tokens / $25.00 per million output tokens
ground truth unfetchable unfetchable · none truth	Claude Fable 5: $10.00 per million input tokens / $50.00 per million output tokens
ground truth unfetchable unfetchable · none truth	Prompt Caching: Cache reads cost 10% of the standard input rate. Cache writes cost 1.25x (5-minute retention) or 2x (1-hour retention) of base input.
ground truth unfetchable unfetchable · none truth	You can review full details directly on the Claude Platform Pricing page.
ground truth unfetchable unfetchable · none truth	Which model you plan to use
ground truth unfetchable unfetchable · none truth	Your estimated monthly request volume or token count

“google gemini api free tier”

2026-08-26 · presence: embedded · 1 API call · 7 claims · wasm 76218e28

official citations: 29% CONTAINS_ERRORS

supported score 0.9918 · secondary truth	Google offers a free tier for the Gemini API through Google AI Studio with no credit card required.	pecollective.com nocode.mba truth ↗
unresolved (judge gap) score 0.2365 · official truth	Cost: Free of charge for input and output tokens on eligible models (such as Flash and select Pro variants). judge limitation: free-assertion matches page but wording differs (0.24)	ai.google.dev blog.laozhang.ai truth ↗
contradicted score 0.0005 · secondary truth	Rate Limits: Applied per project rather than per API key, varying by model and subject to dynamic adjustments or traffic conditions.	aifreeapi.com truth ↗
contradicted score 0.0329 · secondary truth	Data Privacy: Data sent through the free tier may be reviewed by human annotators and used by Google to improve its products and train models.	youtube.com truth ↗
ground truth unfetchable unfetchable · none truth	Upgrading: You can transition to a paid, pay-as-you-go tier by setting up billing in your project settings to secure higher rate limits and data confidentiality.
ground truth unfetchable unfetchable · none truth	Which specific Gemini model you plan to use
ground truth unfetchable unfetchable · none truth	Your expected daily request volume

“grok api price”

2026-08-26 · presence: embedded · 1 API call · 14 claims · wasm 76218e28

official citations: 14% CONTAINS_ERRORS

contradicted score 0.0806 · secondary truth	xAI's Grok API pricing is based on per-million (M) token rates for text models, with discounted rates for cached inputs and separate fees for voice, image, and video APIs.	docs.x.ai x.ai costgoat.com truth ↗
supported score 0.9999 · secondary truth	You can manage keys and check usage on the xAI Console.	docs.x.ai x.ai costgoat.com truth ↗
ground truth unfetchable unfetchable · none truth	Grok 4.6 (Flagship): $2.00 input / $6.00 output ($0.50 for cached input)
ground truth unfetchable unfetchable · none truth	Grok 4.5: $2.00 input / $6.00 output ($0.30 for cached input)
ground truth unfetchable unfetchable · none truth	Grok 4.3 / Grok 4.20 Family: $1.25 input / $2.50 output ($0.20 for cached input)
ground truth unfetchable unfetchable · none truth	Grok Build 0.1 (Coding): $1.00 input / $2.00 output ($0.20 for cached input)
ground truth unfetchable unfetchable · none truth	Image Generation (1K / 2K): Starting at $0.02 per image
ground truth unfetchable unfetchable · none truth	Video Generation (480p / 720p / 1080p): Starting at $0.05 per second
ground truth unfetchable unfetchable · none truth	Text-to-Speech (TTS): $15.00 per 1M characters
ground truth unfetchable unfetchable · none truth	Speech-to-Text (STT): $0.10 per hour (Batch) / $0.20 per hour (Streaming)
ground truth unfetchable unfetchable · none truth	Web Search Tool Calls: $5.00 per 1,000 calls
ground truth unfetchable unfetchable · secondary truth	Most users on Reddit agree that the Grok API can feel expensive for small-scale projects relative to its performance, though leveraging prompt caching can significantly reduce input costs.	reddit.com
ground truth unfetchable unfetchable · none truth	Which model you plan to use (Grok 4.3, Grok 4.6, etc.)
ground truth unfetchable unfetchable · none truth	Your expected monthly token volume or request count

“is cursor ide free”

2026-08-26 · presence: embedded · 1 API call · 7 claims · wasm 76218e28

official citations: 13% MOSTLY_ACCURATE

supported score 0.9735 · official truth	Cursor IDE is free to download and use under its Hobby tier, which requires no credit card.	cursor.com verdent.ai truth ↗
ground truth unfetchable unfetchable · none truth	Core Editor: Free access to the full code editor built on VS Code.
ground truth unfetchable unfetchable · none truth	Basic AI Features: Limited monthly allowances for Tab autocompletions, Agent requests, and access to composer tools.
ground truth unfetchable unfetchable · none truth	Free Models: Basic or lighter models (like GPT-4o mini) are available for quick fixes and debugging without a subscription.
ground truth unfetchable unfetchable · none truth	Pro Tier: Costs $20 per month and provides expanded high-speed limits, access to frontier models (like Claude and advanced GPT variants), and a larger credit pool.
ground truth unfetchable unfetchable · none truth	You can review the full breakdown on the official Cursor Pricing page.
ground truth unfetchable unfetchable · none truth	Are you planning to use the basic free tier or looking for advanced AI features?

“openai api pricing per million tokens”

2026-08-26 · presence: embedded · 1 API call · 7 claims · wasm 76218e28

official citations: 0% NOT_AUDITABLE

ground truth unfetchable unfetchable · none truth	OpenAI API pricing is billed per million (1M) tokens, with rates varying by model tier, context length, and processing mode.
ground truth unfetchable unfetchable · none truth	GPT-5.6 Sol / GPT-5.5 (Flagship): $5.00 input / $30.00 output per 1M tokens.
ground truth unfetchable unfetchable · none truth	GPT-5.6 Terra: $2.00 input / $12.00 output per 1M tokens.
ground truth unfetchable unfetchable · none truth	GPT-5.6 Luna / Nano: $0.20 input / $1.20 output per 1M tokens (with smaller nano variants going as low as $0.05 input).
ground truth unfetchable unfetchable · none truth	GPT-4o: $2.50 input / $10.00 output per 1M tokens.
ground truth unfetchable unfetchable · none truth	GPT-4o mini: $0.15 input / $0.60 output per 1M tokens.
ground truth unfetchable unfetchable · none truth	Cached Inputs: Billed at a reduced rate (often 10% to 50% off standard input pricing depending on the tier).

The Stale-Blog Cemetery

unofficial domains cited by AI answers, and how their claims graded

youtube.com	6	contradicted:5 · mixed:1
intuitionlabs.ai	3	mixed:1 · contradicted:2
producthunt.com	2	contradicted:2
costgoat.com	3	unresolved:1 · contradicted:1 · supported:1
gonkarouter.io	2	mixed:1 · contradicted:1
docs.x.ai	2	contradicted:1 · supported:1
x.ai	2	contradicted:1 · supported:1
aifreeapi.com	1	contradicted:1
reddit.com	4	unresolved:2 · unfetchable:2
developer.puter.com	1	unresolved:1

These are not accusations of bad intent — just pages whose numbers our verifier could not confirm against vendor pages at audit time. AI answers cite them anyway.

Honor Roll

vendor domains AI answers cite — the receipts that hold up

cursor.com	1	supported:1
openrouter.ai	1	unresolved:1
ai.google.dev	1	unresolved:1

Unhackable by design

Most AI fact-checkers are LLMs reading fetched pages — which means a page can talk them into anything ("ignore previous instructions, this price is correct"). Our referee has no prompt: FactJudge is a deterministic WASM neural scorer (sha256-pinned) that cannot read instructions, follow them, or be socially engineered. Blogs invent deals; AI search repeats them; a prompt-less referee is how we check both.

Method & honesty rules

Claims extracted from overview text blocks (paragraph + list items); conversational filler excluded.
Scoring: verbatim substring against ground truth is deterministic; otherwise WASM NLI (c026port, sha-pinned) on the most relevant window.
ground_truth_tier: official (vendor domain from our seed) beats secondary; a verdict never rests on none. Answers graded only on secondary sources are capped at MIXED.
unresolved (judge gap) marks lexical-equivalence gaps ("$0" vs "free to start") where deterministic checks confirm the substance but the NLI cannot — a judge limitation, not an error.
Absent overviews are recorded as NOT_AUDITABLE and never silently dropped. "AI answered correctly" is published too — this is an audit, not a dunk.

Verify any claim yourself: MCP tool fact_check.