In one line: a model's overall score is a weighted average of task performance, cost efficiency, context window, speed, safety, privacy, integration and adoption ease. Weights are locked at v1.0 and version-controlled. Every individual score links to its source.
The 8 scoring factors
| Factor | Weight | What it measures |
|---|---|---|
| Task Performance | 25% | Benchmark scores by task: coding (SWE-bench), reasoning (ARC-AGI-2), writing, analysis, multimodal |
| Cost Efficiency | 20% | Input/output token price, context caching discounts, agentic multiplier (5–20x), volume economics |
| Context Window | 15% | Token context length — critical for document processing, long workflows and large codebases |
| Speed / Latency | 10% | Tokens per second, time-to-first-token — critical for live interactions and real-time agents |
| Safety & Reliability | 10% | Hallucination rates by task, refusal consistency, uptime, context faithfulness |
| Data Privacy | 10% | Training opt-out, enterprise DPA, GDPR/HIPAA/SOC2 compliance, data residency |
| Integration Quality | 5% | API quality, SDK availability, MCP support, rate limits, enterprise SLA |
| Adoption Ease | 5% | No-code access, fine-tuning availability, documentation quality, community size |
The overall weighted score reflects all-round business value. It is not a verdict on which model is "best" — the best model for a specific task is often different, which is exactly why the match engine and per-task pages exist.
What our scores are — and are not
They are an editorial synthesis of published evidence: provider pricing pages, official model documentation, public benchmark leaderboards (SWE-bench, ARC-AGI-2, Scale SEAL, Artificial Analysis) and independent test reports.
They are not first-person lab tests. We never claim to have personally benchmarked a model's speed or run our own evaluations. Where a vendor-reported number differs materially from an independent one, we show both and flag the gap.
Data sources
- Provider pricing and documentation pages (Anthropic, OpenAI, Google, DeepSeek, Meta, xAI, Microsoft, Perplexity)
- SWE-bench Verified — swebench.com
- Scale SEAL leaderboard — scale.com/leaderboard
- ARC Prize / ARC-AGI — arcprize.org
- Artificial Analysis — artificialanalysis.ai
Update cadence
AI pricing has dropped roughly 80% in the past year and benchmark leaderboards change monthly. This is the product: a stale comparison site loses authority fast in this market.
- Monthly: token pricing, data-verified date, new flagship model scoring within two weeks of general availability.
- Quarterly: full re-score of all models, schema validation, cross-link audit.
Scoring weight version log
| Version | Date | Change |
|---|---|---|
| v1.0 | June 2026 | Initial weighting across 8 factors. Locked. |
Weights will not change without incrementing the version here and recording the rationale on the ethics page.
Model profiles and decision pages
Entity profiles use the same published scores as ai-comparison-2026.json. The model index lists the first three so they are not orphans. Open a profile when you need one model's task, cost and privacy story. Use the match engine when the job should choose the model.
- GPT-5.4 — task fit, API cost and privacy
- Claude Sonnet 4.6 — truth, privacy and the workhorse price
- Gemini 3 Flash — speed, long context and support volume
Hallucination and governance questions below are separate decisions from which model hallucinates most and from how to reduce hallucinations.
- Can you trust AI citations without sources?
- Do legal and medical questions raise hallucination risk?
- How do you detect hallucinations in customer support?
- When should a business block AI from final decisions?
- Which controls matter before the first pilot?
Further head-to-heads: Gemini vs Claude and ChatGPT vs DeepSeek, alongside ChatGPT vs Claude.