The smartest AI model isn't always the right one.
Dervity weighs intelligence against cost and speed for each kind of work, then tells you which model to use and what you trade if you go cheaper or faster. Independent, sourced, and rechecked whenever a model or a price changes.
One question, three right answers.
A leaderboard gives you the smartest model. But if you pay for every task or wait on every answer, a different model is often the better call. Working that out used to mean benchmark sites, price lists and a spreadsheet. Here's how it plays out today.
- Coding
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- DeepSeek V4.1 Flash
- Research
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- Grok 4.7
- Email
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- DeepSeek V4.1 Flash
- Image analysis
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- DeepSeek V4.1 Flash
- Math and reasoning
- Smartest
- Claude Opus 5.5
- Best value
- DeepSeek V4.1 Flash
- Fastest
- Grok 4.7
- Creative writing
- Smartest
- GPT-6 Astra
- Best value
- GPT-6 Sol
- Fastest
- Grok 4.7
- AI agents
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- DeepSeek V4.1 Flash
- OCR
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- DeepSeek V4.1 Flash
- Document analysis
- Smartest
- Claude Opus 5.5
- Best value
- MiMo-V2.6-Pro
- Fastest
- DeepSeek V4.1 Flash
Every choice pulls three ways.
None of them always wins, and the balance moves every time a model ships or a price changes. Our mark is those three forces turning around one point: the right choice for the work in front of you.
- Intelligence
- How good the answers are, scored at the setting people actually run, not the lab's best-case demo.
- Cost
- What a real job costs, thinking tokens included, not just the price per million tokens.
- Speed
- How long you wait for a full answer, measured end to end.
How we decide.
- The job comes first.
- A model is only good or bad at something. The best coding model can be a poor pick for OCR, so every recommendation names the task.
- Cost is part of the answer.
- A model that's too expensive to run at your volume is the wrong pick, however well it scores.
- No “it depends” without the why.
- When there's no single answer, we name the deciding factor, like context length or price per task.
- We show what labs leave out.
- Launch posts lead with the scores they win. We add the ones they skipped, including when a competitor is ahead.
Every number has a source.
Scores and prices trace back to a named source. Anything we measured ourselves is marked as ours.
- Artificial AnalysisIntelligence scores from their Intelligence Index v4.3.2, coding scores from their Coding Agent Index v1.5, plus list prices and measured output speed.
- OpenRouter and provider pricingInput and output prices per million tokens, from OpenRouter and the providers' official APIs.
- Provider documentationContext windows and output limits, taken from release notes and API docs.
- Arena results and practitioner reportsHow models hold up in head-to-head use, and what developers report after shipping with them.
- Our own testsSelected hands-on runs. When a guide includes one, it's labelled as ours and kept separate from sourced numbers.
Written and built by
Daniel Orson
Daniel builds Dervity's tools and writes its guides, 21 so far, on model selection, benchmark context, cost, and agent workflows.
Find the right model for your next task.
Answer three questions and get a ranked pick with pricing.