Safety Center & Disclaimers

Effective Date: March 27, 2026
Last Reviewed: March 27, 2026

InferenceBench is committed to responsible, transparent, and safe operation. This Safety Center outlines our practices, limitations, disclaimers, and guidelines for responsible use of the platform.

1. Data Accuracy & Limitations

1.1 Pricing Data

• GPU and inference API pricing is sourced from provider websites and APIs. Prices change without notice and may differ from what you see on this platform at any given time.
• Pricing does not include taxes, egress fees, support tiers, committed-use discounts, spot/preemptible pricing variations, or region-specific surcharges unless explicitly stated.
• Currency conversions, where applicable, use approximate exchange rates and may not reflect real-time market rates.
• Data is refreshed periodically (up to every 6 hours when auto-sync is enabled), but staleness between refresh cycles is possible.

1.2 Performance Estimates

• Throughput estimates (tokens/second) are based on roofline models, CUDA kernel modeling, and publicly available benchmark data. Actual throughput depends on batch size, sequence length, quantization, serving framework, network latency, and hardware utilization.
• Memory (VRAM) estimates use theoretical calculations for model weights, KV-cache, and activations. Real-world VRAM usage varies with framework overhead, memory fragmentation, and concurrent workloads.
• Training cost and time estimates assume idealized conditions (100% GPU utilization, no failures, no checkpointing overhead). Real training runs typically take 20-50% longer due to operational overhead.

1.3 Benchmark Data

• Benchmark results are sourced from HuggingFace LLM Perf Leaderboard, official vendor publications, and community submissions. We do not independently verify all benchmark claims.
• Benchmark conditions (hardware, software, prompts, datasets) vary across sources. Direct comparisons between benchmarks from different sources may not be meaningful.
• Quality scores (MMLU, HumanEval, GSM8K, etc.) are from published evaluations and may not reflect performance on your specific use case.

1.4 Model & GPU Specifications

• Model specifications (parameter counts, architecture, context lengths) are sourced from official model cards and publications. Specifications may be updated by model creators without notice.
• GPU specifications (VRAM, bandwidth, FLOPS, TDP) are sourced from official vendor documentation. Actual performance may vary between individual units, firmware versions, and cooling configurations.

2. Responsible Use Guidelines

InferenceBench is designed to help organizations make informed decisions about AI infrastructure. To use the platform responsibly:

2.1 Do

• Use estimates as starting points for planning and budgeting, not as final procurement specifications
• Validate cost estimates against actual provider quotes before committing budget
• Run pilot benchmarks on your specific workloads before large-scale deployment
• Consider total cost of ownership (TCO) including engineering time, operational overhead, and opportunity costs beyond what this calculator covers
• Report pricing or data inaccuracies through our community reporting tools to improve data quality for all users

2.2 Do Not

• Present InferenceBench estimates as guaranteed pricing, binding quotes, or contractual commitments
• Make irreversible procurement or financial decisions based solely on calculator outputs without independent verification
• Use leaderboard rankings as the sole criterion for model or provider selection — rankings are based on limited metrics and may not reflect fitness for your specific use case
• Redistribute or republish InferenceBench data as your own proprietary analysis without attribution

3. No Professional Advice

Nothing on InferenceBench constitutes:

• Financial advice: Cost estimates are not financial projections. Consult a qualified financial advisor for investment decisions related to AI infrastructure.
• Legal advice: Licensing information is provided for informational purposes. Consult legal counsel for compliance and contractual matters.
• Engineering advice: Hardware recommendations are algorithmic suggestions based on specifications. Consult qualified engineers for production architecture decisions.
• Procurement advice: Provider comparisons are based on published data. Engage directly with providers for enterprise pricing, SLAs, and contractual terms.

4. Limitation of Liability

TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW:

• InferenceBench, its owners, contributors, and affiliates shall not be liable for any direct, indirect, incidental, special, consequential, punitive, or exemplary damages arising from your use of or reliance on the platform.
• This includes, without limitation: financial losses from procurement decisions, opportunity costs from suboptimal hardware selection, damages from inaccurate pricing or performance data, business interruption, or loss of profits.
• Our total aggregate liability shall not exceed USD $100.00.

5. No Warranty

THE PLATFORM IS PROVIDED "AS IS" AND "AS AVAILABLE" WITHOUT WARRANTIES OF ANY KIND, WHETHER EXPRESS, IMPLIED, STATUTORY, OR OTHERWISE. WE SPECIFICALLY DISCLAIM ALL IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, ACCURACY, COMPLETENESS, RELIABILITY, AND NON-INFRINGEMENT.

We do not warrant that:

• Data will be accurate, complete, current, or error-free at any given time
• The platform will be available without interruption
• Calculations will produce correct results for all edge cases and configurations
• Rankings, scores, or recommendations will remain stable over time

6. Community-Submitted Data

InferenceBench allows community contributions including pricing reports, benchmark results, and experience reports. Regarding this data:

• Community submissions are accepted in good faith but are not independently verified by InferenceBench
• We make reasonable efforts to detect and remove spam, fraudulent, or malicious submissions
• Users should exercise judgment when relying on community-sourced data and cross-reference with official sources
• InferenceBench is not liable for inaccuracies in community-submitted data
• By submitting data, contributors affirm that their submissions are accurate to the best of their knowledge and do not violate any confidentiality obligations

7. Third-Party Content & Links

• The platform references third-party products, services, and organizations. Such references do not constitute endorsement, sponsorship, or recommendation.
• Provider logos and brand assets are used for identification purposes under fair use. All trademarks belong to their respective owners.
• External links are provided for convenience. We are not responsible for the content, privacy practices, or availability of third-party websites.

8. Reporting Inaccuracies & Safety Issues

If you discover inaccurate data, safety concerns, or content that could lead to harmful decisions, please report it immediately:

We commit to acknowledging all safety-related reports within 48 hours and providing a substantive response within 5 business days.

9. Data Integrity Incident Response

If we discover or are notified of significant data inaccuracies that could lead to material financial harm:

• Affected data will be flagged or removed within 24 hours of confirmed discovery
• A correction notice will be published on the affected pages
• Root cause analysis will be conducted and preventive measures implemented
• Users who reported the issue will be notified of the resolution

10. Related Policies

This Safety Center should be read in conjunction with:

• Terms of Service — Legal terms governing use of the platform
• Privacy Policy — How we handle data (GDPR/CCPA compliant)
• Cookie Policy — Browser storage and cookie usage
• Compliance & Security — GDPR, CCPA, SOC 2, OWASP, open source compliance
• Open Source Licenses — License information and attribution

11. Contact

For safety-related inquiries:

• Safety: safety@inferencebench.io
• Data accuracy: data@inferencebench.io
• Legal: legal@inferencebench.io
• Website: inferencebench.io