AI models guide · verified Nov 18, 2026 · 66 models · 6 providers · 14 model families · archive since 2026
Which AI should you use? Every model in PolyCog: what each is good at, how much it filters, and which few fit your project.
PolyCog runs 66 models from 6 providers. None is best at everything, and some refuse whole subjects. Tell us what you’re working on, and we’ll find the best models for it. The rest of the page shows the reasoning: strengths, filters, refusals, and retired models.
Runs in your browser. Nothing you type leaves this page. 1041 project types in the dictionary.
When is a debate worth it?
Run the full panel for
- A choice between options: two offers, three names, build or buy
- A draft, plan or piece of code you want picked apart
- Money, health, legal and career calls, where a second opinion is cheap insurance
- Anything contested, recent, or easy to get confidently wrong
One model is enough for
- A fact or a number with one right answer
- A quick how-to, a recipe or a conversion
- Translating, summarizing or reformatting text
- A first draft you plan to rewrite anyway
A debate earns its cost when the models disagree. If you don’t expect them to, one good model is faster and cheaper, and the box says so when it can tell.
The 6 seats at a glance
Each provider is one seat in a PolyCog debate. The Wildcard seat is one OpenRouter key that opens 8 more labs.
Guardrails level, ranked
How much each model filters, from least to most. Five general dimensions make up the level. A sixth, PRC-sensitive topics, is shown on its own because it moves independently of the rest. Scores are PolyCog’s dated read of each provider’s usage policy and observed behaviour in the app; they describe, they don’t referee, and nothing here is a guide to getting around a policy.
| VeniceCognitive Computations · Venice · Wildcard seat · 1 model | Minimal · 1/5 | 1 | 1 | 1 | 1 | 1 | PRC topics: none |
| NousNous Research · Wildcard seat · 1 model | Minimal · 1/5 | 1 | 1 | 1 | 1 | 2 | PRC topics: none |
| GrokxAI · 4 models | Minimal · 1/5 | 1 | 2 | 1 | 2 | 1 | PRC topics: none |
| KimiMoonshot · Wildcard seat · 3 models | Light · 2/5 | 3 | 3 | 2 | 2 | 1 | PRC topics: strict |
| DeepSeekDeepSeek · 2 models | Light · 2/5 | 3 | 3 | 3 | 2 | 1 | PRC topics: strict |
| LlamaMeta · Wildcard seat · 1 model | Light · 2/5 | 2 | 3 | 2 | 3 | 2 | PRC topics: none |
| MistralMistral AI · Wildcard seat · 3 models | Light · 2/5 | 2 | 3 | 2 | 4 | 1 | PRC topics: none |
| MuseMeta · Wildcard seat · 3 models | Light · 2/5 | 2 | 3 | 2 | 3 | 2 | PRC topics: none |
| QwenAlibaba · Wildcard seat · 3 models | Light · 2/5 | 3 | 3 | 3 | 2 | 1 | PRC topics: strict |
| MiniMaxMiniMax · Wildcard seat · 2 models | Moderate · 3/5 | 3 | 3 | 4 | 2 | 1 | PRC topics: strict |
| ChatGPTOpenAI · 17 models | Moderate · 3/5 | 3 | 4 | 3 | 3 | 2 | PRC topics: none |
| GLMZ.ai · Wildcard seat · 2 models | Moderate · 3/5 | 4 | 3 | 4 | 3 | 1 | PRC topics: strict |
| GeminiGoogle · 11 models | Moderate · 3/5 | 3 | 3 | 4 | 4 | 2 | PRC topics: none |
| ClaudeAnthropic · 13 models | Moderate · 3/5 | 4 | 4 | 4 | 3 | 2 | PRC topics: none |
What each will and won’t do
By subject, per model family. The “what are you working on?” box checks this table before it ranks anything. ✓ answers · ~ softens, hedges or partly refuses · ✗ refuses. Hover or tap a cell for the policy behind it.
| Model family | Sexual content | Romance & roleplay | Dark & violent fiction | Fan fiction & known characters | Drugs & substances | Weapons & firearms | Security research | Medical detail | Legal advice | Contested politics | PRC-sensitive topics | Edgy or offensive humour | Gambling & betting | Sex education & health | Self-harm & crisis |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ChatGPT | ✗ | ✓ | ~ | ✓ | ~ | ~ | ~ | ~ | ~ | ~ | ✓ | ~ | ~ | ✓ | ~ |
| Claude | ✗ | ~ | ~ | ✓ | ~ | ~ | ~ | ~ | ~ | ~ | ✓ | ~ | ✓ | ✓ | ~ |
| Gemini | ✗ | ✓ | ~ | ✓ | ~ | ~ | ~ | ~ | ~ | ~ | ✓ | ~ | ✓ | ✓ | ~ |
| DeepSeek | ✗ | ✓ | ~ | ✓ | ~ | ~ | ~ | ✓ | ✓ | ~ | ✗ | ✗ | ~ | ✓ | ~ |
| Grok | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ~ |
| Qwen | ✗ | ✓ | ~ | ✓ | ~ | ~ | ~ | ✓ | ✓ | ~ | ✗ | ✗ | ✗ | ✓ | ~ |
| Kimi | ✗ | ✓ | ~ | ✓ | ~ | ~ | ~ | ✓ | ✓ | ~ | ✗ | ~ | ✓ | ✓ | ~ |
| GLM | ✗ | ~ | ✗ | ✓ | ~ | ~ | ~ | ~ | ~ | ~ | ✗ | ✗ | ✓ | ✓ | ~ |
| Mistral | ✓ | ✓ | ~ | ✓ | ✗ | ~ | ~ | ~ | ~ | ~ | ✓ | ~ | ✓ | ✓ | ~ |
| Llama | ~ | ✓ | ✓ | ✓ | ~ | ~ | ~ | ~ | ~ | ~ | ✓ | ~ | ✓ | ✓ | ~ |
| Muse | ~ | ✓ | ✓ | ✓ | ~ | ~ | ~ | ~ | ~ | ~ | ✓ | ~ | ✓ | ✓ | ~ |
| MiniMax | ✗ | ✓ | ✗ | ✓ | ~ | ~ | ~ | ✓ | ✓ | ~ | ✗ | ✗ | ✓ | ✓ | ~ |
| Venice | ~ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | ✓ | ✓ | ~ |
| Nous | ~ | ✓ | ✓ | ✓ | ~ | ~ | ~ | ~ | ~ | ✓ | ✓ | ~ | ✓ | ✓ | ~ |
⛔ Every provider refuses, and so does PolyCog: functional malware · sexual content involving minors · weapons of mass harm, explosives & illicit synthesis · targeting a private person · fraud, scams & forgery · hate & extremism. Sources: each provider’s usage policy, read on the verified date; hover a name for the policy behind its row. Rows show a lab’s policy; the host serving an open-weight model through OpenRouter can be stricter.
Model sizes
Parameter counts, where the lab publishes one. Size is one clue to what a model can do, not a ranking: training data and tuning matter as much, and OpenAI, Anthropic, Google and xAI don’t publish sizes at all. Click a column to sort.
| Design | |||||
|---|---|---|---|---|---|
| Kimi K3Moonshot · Wildcard | 2.8T | 104B | MoE | 1M | $3.00 / $15.00 |
| Qwen3.8 MaxAlibaba · Wildcard | 2.4T | 95B | MoE | 1M | $2.00 / $6.00 |
| DeepSeek V4 ProDeepSeek | 1.6T | 49B | MoE | 1M | $1.32 / $3.96 |
| Mistral Large 4Mistral AI · Wildcard | 1.05T | 52B | MoE | 1M | $0.70 / $2.10 |
| Kimi K2.6Moonshot · Wildcard | 1T | 32B | MoE | 256K | $0.95 / $4.00 |
| GLM 5.3Z.ai · Wildcard | 744B | 40B | MoE | 1M | $1.40 / $4.40 |
| DeepSeek V4.1 FlashDeepSeek | 552B | 8–16B | MoE | 1M | $0.30 / $1.20 |
| MiniMax M3MiniMax · Wildcard | ~428B | ~23B | MoE | 1M | $0.30 / $1.20 |
| MiniMax M3.1 FlashMiniMax · Wildcard | ~428B | ~23B | MoE | 1M | $0.15 / $0.60 |
| Hermes 4 405BNous Research · Wildcard | 405B | all | Dense | 131K | $1.00 / $3.00 |
| Llama 4 MaverickMeta · Wildcard | 400B | 17B | MoE | 1M | $0.35 / $1.15 |
| GLM 5.3 FlashZ.ai · Wildcard | 320B | 18B | MoE | 1M | $0.15 / $0.50 |
| Mistral Medium 3.5Mistral AI · Wildcard | 128B | all | Dense | 256K | $1.50 / $7.50 |
| Qwen3.8 FlashAlibaba · Wildcard | 125B | 6B | MoE | 1M | $0.15 / $0.47 |
| Mistral Small 4Mistral AI · Wildcard | 119B | 6.5B | MoE | 256K | $0.15 / $0.60 |
| Venice UncensoredCognitive Computations · Venice · Wildcard | 24B | all | Dense | 128K | $0.20 / $0.90 |
| GPT-6.2 AstraOpenAI | Not published | — | — | 1.05M | $8.00 / $40.00 |
| GPT-6.2 SolOpenAI | Not published | — | — | 1.05M | $2.00 / $10.00 |
| GPT-6.1 TerraOpenAI | Not published | — | — | 1.05M | $1.00 / $5.00 |
| GPT-6 LunaOpenAI | Not published | — | — | 1.05M | $0.10 / $0.50 |
| GPT-6 AstraOpenAI | Not published | — | — | 1.05M | $10.00 / $50.00 |
| GPT-6.1 SolOpenAI | Not published | — | — | 1.05M | $2.00 / $10.00 |
| GPT-5.6 SolOpenAI | Not published | — | — | 1.05M | $4.00 / $20.00 |
| GPT-6 SolOpenAI | Not published | — | — | 1.05M | $2.00 / $10.00 |
| GPT-5.6 TerraOpenAI | Not published | — | — | 1.05M | $2.00 / $12.00 |
| GPT-5.6 LunaOpenAI | Not published | — | — | 1.05M | $0.20 / $1.20 |
| GPT-5.5OpenAI | Not published | — | — | 1.05M | $5.00 / $30.00 |
| GPT-5.4OpenAI | Not published | — | — | 1.05M | $2.50 / $15.00 |
| GPT-5.1OpenAI · legacy | Not published | — | — | 400K | $1.25 / $10.00 |
| GPT-4.1OpenAI · legacy | Not published | — | — | 1M | $2.00 / $8.00 |
| GPT-5.4 miniOpenAI | Not published | — | — | 400K | $0.75 / $4.50 |
| GPT-4.1 miniOpenAI · legacy | Not published | — | — | 1M | $0.40 / $1.60 |
| GPT-4o miniOpenAI · legacy | Not published | — | — | 128K | $0.15 / $0.60 |
| Claude Fable 5.2Anthropic | Not published | — | — | 1M | $8.00 / $40.00 |
| Claude Opus 5.5Anthropic | Not published | — | — | 1M | $4.00 / $20.00 |
| Claude Sonnet 5.5Anthropic | Not published | — | — | 1M | $2.00 / $10.00 |
| Claude Haiku 5.5Anthropic | Not published | — | — | 1M | $0.10 / $0.50 |
| Claude Fable 5.1Anthropic | Not published | — | — | 1M | $10.00 / $50.00 |
| Claude Haiku 4.5Anthropic | Not published | — | — | 200K | $1.00 / $5.00 |
| Claude Sonnet 5Anthropic | Not published | — | — | 1M | $2.00 / $10.00 |
| Claude Fable 5Anthropic | Not published | — | — | 1M | $10.00 / $50.00 |
| Claude Opus 5Anthropic | Not published | — | — | 1M | $5.00 / $25.00 |
| Claude Opus 4.8Anthropic | Not published | — | — | 1M | $5.00 / $25.00 |
| Claude Opus 4.7Anthropic | Not published | — | — | 1M | $5.00 / $25.00 |
| Claude Opus 4.6Anthropic | Not published | — | — | 1M | $5.00 / $25.00 |
| Claude Sonnet 4.6Anthropic | Not published | — | — | 1M | $3.00 / $15.00 |
| Gemini 4 CarbonGoogle | Not published | — | — | 1M | $2.00 / $10.00 |
| Gemini 4 NeonGoogle | Not published | — | — | 1M | $3.00 / $15.00 |
| Gemini 4 HydrogenGoogle | Not published | — | — | 1M | $1.50 / $7.50 |
| Gemini 3.8 FlashGoogle | Not published | — | — | 1M | $0.75 / $3.75 |
| Gemini 2.5 Flash-LiteGoogle · legacy | Not published | — | — | 1M | $0.10 / $0.40 |
| Gemini 4 ArgonGoogle | Not published | — | — | 1M | $2.00 / $10.00 |
| Gemini 3.6 FlashGoogle | Not published | — | — | 1M | $0.75 / $3.75 |
| Gemini 3.1 Pro (preview)Google | Not published | — | — | 1M | $2.00 / $12.00 |
| Gemini 3 Flash (preview)Google · legacy | Not published | — | — | 1M | $0.50 / $3.00 |
| Gemini 3.5 Flash-LiteGoogle | Not published | — | — | 1M | $0.30 / $2.50 |
| Gemini 3.1 Flash-LiteGoogle · legacy | Not published | — | — | 1M | $0.25 / $1.50 |
| Grok 4.7xAI | Not published | — | — | 500K | $2.00 / $6.00 |
| Grok 4.6xAI | Not published | — | — | 500K | $2.00 / $6.00 |
| Grok 4.5xAI | Not published | — | — | 500K | $2.00 / $6.00 |
| Grok 4.3xAI | Not published | — | — | 1M | $1.25 / $2.50 |
| Qwen4 MaxAlibaba · Wildcard | Not published | — | — | 1M | $1.425 / $4.275 |
| Kimi K3.1 ProMoonshot · Wildcard | Not published | — | — | 1M | $2.00 / $10.00 |
| Muse WonderMeta · Wildcard | Not published | — | — | 1M | $4.00 / $20.00 |
| Muse Spark 1.3Meta · Wildcard | Not published | — | — | 1M | $1.25 / $4.25 |
| Muse EmberMeta · Wildcard | Not published | — | — | 1M | $0.10 / $0.50 |
ChatGPT · OpenAI
OpenAI’s line runs from a frontier flagship (GPT-6.2 Astra) through GPT-6.2 Sol for complex coding and agents and GPT-6.1 Terra for computer use and everyday work to GPT-6 Luna for high-volume work, with web search across the GPT-6 and 5.x models. OpenAI describes Astra as built for the hardest projects from start to finish. Filters sit mid-field: sexual content is refused, edgy humour and gore are softened, and its usage policies rule out tailored legal or medical advice without a professional and real-money gambling.
GPT-6.2 Astragpt-6.2-astrasynthesisOpenAI’s frontier flagship since Nov 17, 2026, with a million-token window and web search. OpenAI builds it for the hardest projects from start to finish and reports higher scores than GPT-6 Astra on professional work at a lower cost per task. Frontier-priced, a fifth below GPT-6 Astra and on the same rate card as Claude Fable 5.2.$8.00 / $40.00 per MTokModerate
gpt-6.2-astrasynthesisReach for it when
- hard reasoning
- large codebases
- agentic pipelines
- research with citations
Look elsewhere for
- bulk or high-volume jobs
- quick lookups
GPT-6.2 Solgpt-6.2-solThe GPT-6.2 everyday tier, with a million-token window and web search. OpenAI calls it its go-to model for everyday work and reports it close to GPT-6.2 Astra on a coding benchmark at about a fifth of the cost. It replaces GPT-6.1 Sol at the same rate, so it’s the OpenAI pick for most prompts without the Astra bill.$2.00 / $10.00 per MTokModerate
gpt-6.2-solReach for it when
- complex coding
- multi-step agents
- long documents
Look elsewhere for
- throwaway prompts
GPT-6.1 Terragpt-6.1-terraThe GPT-6.1 balanced tier, at half GPT-6.2 Sol’s rate. OpenAI builds it for computer use and visual work (reading screens, invoices and technical drawings) and reports large gains in coding and factual accuracy over GPT-5.6 Terra, while still pointing its hardest coding at Sol and Astra.$1.00 / $5.00 per MTokModerate
gpt-6.1-terraReach for it when
- computer use and screens
- visual documents
- everyday coding
Look elsewhere for
- the hardest coding (pick 6.2 Sol)
GPT-6 Lunagpt-6-lunadefault seatFast and very cheap, with a million-token window and search. The OpenAI default seat, for drafts, extraction and high-volume work.$0.10 / $0.50 per MTokModerate
gpt-6-lunadefault seatReach for it when
- bulk extraction
- drafts
- classification
Look elsewhere for
- nuanced writing
- hard math
More ChatGPT models (13 behind the picker’s expander)
GPT-6 Astragpt-6-astraThe prior frontier flagship, on a higher rate card than GPT-6.2 Astra. Pick 6.2 Astra unless a session started here.$10.00 / $50.00 per MTokModerate
gpt-6-astraReach for it when
- same as 6.2 Astra
Look elsewhere for
- new sessions (pick GPT-6.2 Astra)
GPT-6.1 Solgpt-6.1-solThe prior Sol. GPT-6.2 Sol now does the same work at the same rate, so keep this one for restored sessions or a side-by-side.$2.00 / $10.00 per MTokModerate
gpt-6.1-solReach for it when
- restored sessions
- complex coding
Look elsewhere for
- new work (pick GPT-6.2 Sol)
GPT-5.6 Solgpt-5.6-solThe GPT-5.6 flagship on a promotional price, for deep reasoning and long agentic runs. GPT-6.2 Sol now does the same work for half.$4.00 / $20.00 per MTokModerate
gpt-5.6-solReach for it when
- complex coding
- multi-step agents
- analysis
Look elsewhere for
- throwaway prompts
GPT-6 Solgpt-6-solThe first GPT-6 mid tier, built for complex coding and agentic workflows. GPT-6.2 Sol now does the same work at the same rate, so keep this one for restored sessions or a side-by-side.$2.00 / $10.00 per MTokModerate
gpt-6-solReach for it when
- restored sessions
- complex coding
Look elsewhere for
- new work (pick GPT-6.2 Sol)
GPT-5.6 Terragpt-5.6-terraThe previous Terra, capable enough for real development and business work. GPT-6.1 Terra now does more of it for less.$2.00 / $12.00 per MTokModerate
gpt-5.6-terraReach for it when
- everyday coding
- business writing
- planning
Look elsewhere for
- frontier-hard reasoning
GPT-5.6 Lunagpt-5.6-lunaFast and cheap with search. The former OpenAI default seat; GPT-6 Luna now costs less.$0.20 / $1.20 per MTokModerate
gpt-5.6-lunaReach for it when
- bulk extraction
- drafts
- classification
Look elsewhere for
- nuanced writing
- hard math
- new sessions (pick GPT-6 Luna)
GPT-5.5gpt-5.5An older flagship still priced like one; capable, but Sol does the same work for less.$5.00 / $30.00 per MTokModerate
gpt-5.5Reach for it when
- reasoning
Look elsewhere for
- anything Sol can do
GPT-5.4gpt-5.4A solid all-rounder from the prior line, now overlapping Terra.$2.50 / $15.00 per MTokModerate
gpt-5.4Reach for it when
- general work
Look elsewhere for
- new projects (pick Terra)
GPT-5.1gpt-5.1legacyOlder flagship line at a budget price. OpenAI shuts it down on April 1, 2027 and names GPT-6 Sol as the replacement.$1.25 / $10.00 per MTokModerate
gpt-5.1legacyReach for it when
- general work
Look elsewhere for
- new sessions (pick 6 Sol)
GPT-4.1gpt-4.1legacyLegacy million-token generalist without search. OpenAI shuts it down on January 14, 2027 and names GPT-6 Luna as the replacement.$2.00 / $8.00 per MTokModerate
gpt-4.1legacyReach for it when
- very long inputs
Look elsewhere for
- current events
- agents
- new sessions (pick 6 Luna)
GPT-5.4 minigpt-5.4-miniA fast daily driver with search, pitched by OpenAI at coding, computer use and subagents. Now pricier than either Luna.$0.75 / $4.50 per MTokModerate
gpt-5.4-miniReach for it when
- volume work
Look elsewhere for
- hard reasoning
GPT-4.1 minigpt-4.1-minilegacyLegacy small model with a long window and no search. OpenAI shuts it down on January 14, 2027 and names GPT-6 Luna as the replacement.$0.40 / $1.60 per MTokModerate
gpt-4.1-minilegacyReach for it when
- cheap long-context extraction
Look elsewhere for
- anything current
GPT-4o minigpt-4o-minilegacyThe oldest OpenAI row. OpenAI shuts it down on January 14, 2027 and names GPT-6 Luna, which costs less and beats it on everything, as the replacement.$0.15 / $0.60 per MTokModerate
gpt-4o-minilegacyReach for it when
- very cheap bulk
Look elsewhere for
- quality work
Claude · Anthropic
Anthropic’s line runs from a frontier flagship (Fable 5.2) through Opus and Sonnet to Haiku, which also runs PolyCog’s free preview. Anthropic recommends Opus 5.5 as the starting point for most work and Fable 5.2 for the hardest technical work. Filters are firm on content: its usage policy refuses sexual content, erotic chat and graphic violence, and treats legal and healthcare uses as high-risk; Claude’s published constitution asks it not to moralise or be over-cautious on medical and legal questions.
Claude Fable 5.2claude-fable-5-2synthesisAnthropic’s frontier flagship, with a million-token window, web search and adaptive thinking that is always on. Anthropic aims it at the hardest technical work: demanding reasoning, long-horizon agentic coding and multi-step research. Frontier-priced, a fifth below Fable 5.1.$8.00 / $40.00 per MTokModerate
claude-fable-5-2synthesisReach for it when
- hard reasoning
- long-horizon agentic coding
- multi-step research
- long-document analysis
Look elsewhere for
- bulk work
- subjects Anthropic’s policy refuses (see the table)
Claude Opus 5.5claude-opus-5-5Anthropic’s recommended starting point for most work: long-running agentic coding and knowledge work, with adaptive thinking always on. It costs well under Fable’s rate, so it’s the Claude pick for most prompts without the frontier bill.$4.00 / $20.00 per MTokModerate
claude-opus-5-5Reach for it when
- coding
- writing
- analysis
Look elsewhere for
- volume jobs
Claude Sonnet 5.5claude-sonnet-5-5default seatThe mid-tier workhorse and PolyCog’s Claude default. Anthropic places it between Opus and Haiku on speed and cost, with a million-token window and adaptive thinking on by default. It costs half Opus 5.5’s rate.$2.00 / $10.00 per MTokModerate
claude-sonnet-5-5default seatReach for it when
- everyday coding
- editing
- client-facing writing
Look elsewhere for
- frontier-hard problems
Claude Haiku 5.5claude-haiku-5-5free previewAnthropic’s fast tier, which it pitches at high-volume work such as classification, extraction and routing. It has a million-token window and adaptive thinking on by default, and the free preview runs on it. Prompts over 100K tokens bill at a higher rate.$0.10 / $0.50 per MTokModerate
claude-haiku-5-5free previewReach for it when
- quick edits
- summaries
- the free preview
Look elsewhere for
- hard reasoning
- long agentic runs
More Claude models (9 behind the picker’s expander)
Claude Fable 5.1claude-fable-5-1The prior frontier flagship, on a higher rate card than 5.2. Pick 5.2 unless a session started here.$10.00 / $50.00 per MTokModerate
claude-fable-5-1Reach for it when
- same as 5.2
Look elsewhere for
- new sessions (pick 5.2)
Claude Haiku 4.5claude-haiku-4-5-20251001The previous fast Claude, still served. Light tasks, quick edits, classification; Haiku 5.5 replaced it as the free preview’s Claude.$1.00 / $5.00 per MTokModerate
claude-haiku-4-5-20251001Reach for it when
- quick edits
- summaries
- classification
Look elsewhere for
- hard reasoning
- long agentic runs
Claude Sonnet 5claude-sonnet-5The previous Sonnet, at the same rate as Sonnet 5.5, which replaces it as PolyCog’s Claude default. Keep it for restored sessions or a side-by-side.$2.00 / $10.00 per MTokModerate
claude-sonnet-5Reach for it when
- restored sessions
- everyday coding
Look elsewhere for
- new work (pick Sonnet 5.5)
Claude Fable 5claude-fable-5An earlier frontier flagship, on the same rate card as 5.1. Pick 5.2 unless a session started here.$10.00 / $50.00 per MTokModerate
claude-fable-5Reach for it when
- same as 5.2
Look elsewhere for
- new sessions (pick 5.2)
Claude Opus 5claude-opus-5The prior everyday flagship. Opus 5.5 does the same work for less.$5.00 / $25.00 per MTokModerate
claude-opus-5Reach for it when
- coding
- writing
- analysis
Look elsewhere for
- volume jobs
- new sessions (pick Opus 5.5)
Claude Opus 4.8claude-opus-4-8The prior Opus, for heavy-duty reasoning and long agentic work. Opus 5 supersedes it at the same price.$5.00 / $25.00 per MTokModerate
claude-opus-4-8Reach for it when
- agentic coding
Look elsewhere for
- new sessions (pick Opus 5)
Claude Opus 4.7claude-opus-4-7An older Opus snapshot, still active at Anthropic; served for continuity.$5.00 / $25.00 per MTokModerate
claude-opus-4-7Reach for it when
- restored sessions
Look elsewhere for
- new work
Claude Opus 4.6claude-opus-4-6An older Opus snapshot, still active at Anthropic; served for continuity.$5.00 / $25.00 per MTokModerate
claude-opus-4-6Reach for it when
- restored sessions
Look elsewhere for
- new work
Claude Sonnet 4.6claude-sonnet-4-6Prior-generation Sonnet, now pricier than Sonnet 5.5 for less.$3.00 / $15.00 per MTokModerate
claude-sonnet-4-6Reach for it when
- restored sessions
Look elsewhere for
- new work (pick Sonnet 5.5)
Gemini · Google
Google’s line leads with Gemini 4 Carbon, its strongest reasoning model and the one that writes every Gemini synthesis on PolyCog, then Gemini 4 Neon, the middle model Google pitches at complex coding and analysis, and Gemini 4 Hydrogen, its everyday model; Gemini 4 Argon, the previous Pro (3.1, still a preview id) and a deep Flash bench follow, and one Flash-Lite runs PolyCog’s free preview. Google describes Carbon around hard reasoning, complex coding, science and security work, and says it spends more thinking per answer than Argon. Every current model takes text, images, video, audio and PDFs, with a million-token window and Search grounding. Filters are moderate, and firm on medical detail: Google’s API terms rule out using it to give medical advice. Sexual content is refused and gratuitous violence softened.
Gemini 4 Carbongemini-4-carbonsynthesisGoogle’s strongest reasoning model and its flagship since Nov 10, 2026, built for hard reasoning, complex coding, science and security work. Google says it spends more thinking per answer than Gemini 4 Argon, so it is slower and costs more per answer on the same per-token price. A million-token window shared by input, thinking and output, native image, video, audio and PDF input, and Search grounding. On an introductory price through Nov 30, 2026, and not on Google’s free tier.$2.00 / $10.00 per MTokModerate
gemini-4-carbonsynthesisReach for it when
- hard reasoning
- complex coding
- security review
- synthesis
Look elsewhere for
- quick lookups
- free-tier keys (paid tier only)
Gemini 4 Neongemini-4-neonGoogle’s middle Gemini 4 model, released Nov 18, 2026, between Hydrogen and Carbon, aimed at complex coding, analysis and enterprise work. Google says it comes close to Carbon on coding and reasoning at a faster pace, and costs about a third less per task than Carbon at the default setting. A million-token window shared by input, thinking and output, native image, video, audio and PDF input, and Search grounding. It thinks at Google’s high setting by default. One flat price, and not on Google’s free tier.$3.00 / $15.00 per MTokModerate
gemini-4-neonReach for it when
- complex coding
- data and spreadsheet work
- agents
- long documents
Look elsewhere for
- quick lookups
- free-tier keys (paid tier only)
Gemini 4 Hydrogengemini-4-hydrogenGoogle’s everyday Gemini 4 model, released Nov 12, 2026, and the step up from Gemini 3.8 Flash: Google pitches it at everyday coding and agentic work. A million-token window shared by input, thinking and output, native image, video, audio and PDF input, and Search grounding. It thinks at Google’s high setting by default. One flat price, twice Gemini 3.8 Flash’s introductory rate.$1.50 / $7.50 per MTokModerate
gemini-4-hydrogenReach for it when
- everyday coding
- agents
- image understanding
- long documents
Look elsewhere for
- the very hardest reasoning
- free-tier keys in a multi-model run (rate limits)
Gemini 3.8 Flashgemini-3.8-flashdefault seatGoogle’s recommended Flash and PolyCog’s Gemini default: Google’s most intelligent Flash, pitched at long-horizon software engineering and autonomous agents, on an introductory price.$0.75 / $3.75 per MTokModerate
gemini-3.8-flashdefault seatReach for it when
- coding
- agents
- image understanding
- value
Look elsewhere for
- the very hardest reasoning
Gemini 2.5 Flash-Litegemini-2.5-flash-litefree previewlegacyThe sample-tier shared model. Google now limits the 2.5 models to projects that already used them; it works on those keys and in the free preview.$0.10 / $0.40 per MTokModerate
gemini-2.5-flash-litefree previewlegacyReach for it when
- the free preview
- very cheap bulk
Look elsewhere for
- new keys (use 3.5 Flash-Lite)
More Gemini models (6 behind the picker’s expander)
Gemini 4 Argongemini-4-argonThe lighter Gemini 4 model, which Google recommends for most advanced coding and knowledge work: it thinks less per answer than Gemini 4 Carbon, so it answers sooner and costs less per answer on the same per-token price. A million-token window shared by input, thinking and output, native image, video, audio and PDF input, and Search grounding. On an introductory price through Nov 30, 2026, and not on Google’s free tier. Google’s own Gemini app now offers Gemini 4 Neon in its place.$2.00 / $10.00 per MTokModerate
gemini-4-argonReach for it when
- complex coding
- agentic work
- image- and video-heavy work
- faster hard answers
Look elsewhere for
- quick lookups
- free-tier keys (paid tier only)
Gemini 3.6 Flashgemini-3.6-flashA previous-generation Flash that Google pitches at balancing speed and multimodal input across everyday and agentic tasks; 3.8 supersedes it at the same price.$0.75 / $3.75 per MTokModerate
gemini-3.6-flashReach for it when
- everyday tasks
Look elsewhere for
- new sessions (pick 3.8)
Gemini 3.1 Pro (preview)gemini-3.1-pro-previewGoogle’s previous flagship, still a preview id: a million-token window, native image, video, audio and PDF input, grounded search, and Google’s emphasis on complex problem-solving and agentic coding. Gemini 4 Argon took the flagship seat on Oct 13, 2026; this one thinks at Google’s high setting by default and keeps its own long-prompt price tier.$2.00 / $12.00 per MTokModerate
gemini-3.1-pro-previewReach for it when
- complex problem-solving
- image- and video-heavy work
- long documents
- synthesis
Look elsewhere for
- latency-sensitive chat
Gemini 3 Flash (preview)gemini-3-flash-previewlegacyThe Gemini 3 Flash preview: fast and cheap, still served with no shutdown date announced.$0.50 / $3.00 per MTokModerate
gemini-3-flash-previewlegacyReach for it when
- cheap volume with images
Look elsewhere for
- quality-critical work
Gemini 3.5 Flash-Litegemini-3.5-flash-liteThe lite tier for new Google keys and the model PolyCog’s own utility jobs (titles, scans) run on.$0.30 / $2.50 per MTokModerate
gemini-3.5-flash-liteReach for it when
- high-volume extraction
Look elsewhere for
- nuance
Gemini 3.1 Flash-Litegemini-3.1-flash-litelegacyLight and cheap, with an earliest shutdown date announced by Google.$0.25 / $1.50 per MTokModerate
gemini-3.1-flash-litelegacyReach for it when
- cheap bulk
Look elsewhere for
- new sessions (pick 3.5 Flash-Lite)
DeepSeek · DeepSeek
Two models: V4 Pro and V4.1 Flash, which runs PolyCog’s free preview; both have thinking and non-thinking modes and a million-token window. DeepSeek positions the pair around reasoning and code at very low prices. Web search runs on your own key through DeepSeek’s own search tool; Flash reads images natively, while V4 Pro gets them through PolyCog’s bridge. DeepSeek’s terms forbid sexually explicit, offensive and vulgar content, and PRC-sensitive topics are declined or deflected.
DeepSeek V4 Prodeepseek-v4-prosynthesisDeepSeek’s flagship: thinking and non-thinking modes and a million-token window at roughly a tenth of a Western flagship’s price. Web search on your own key through DeepSeek’s own search tool. No native images; PolyCog’s bridge describes them.$1.32 / $3.96 per MTokLight
deepseek-v4-prosynthesisReach for it when
- math
- algorithms
- code review
- second opinions
Look elsewhere for
- image input
- PRC-sensitive topics (see the table)
DeepSeek V4.1 Flashdeepseek-flashdefault seatfree previewA million tokens of context for cents, both thinking modes, native image input, and the sample-tier shared model; web search runs on your own key through DeepSeek’s own search tool.$0.30 / $1.20 per MTokLight
deepseek-flashdefault seatfree previewReach for it when
- cheap long-context coding
- the free preview
Look elsewhere for
- current events on the free preview (no search there)
- PRC-sensitive topics
Grok · xAI
Four models, all with live web and X search. xAI positions Grok 4.7 around code and agentic tool calling. Its acceptable use policy does not restrict adult content between adults (it bars sexual depictions of real people and anything involving minors), and xAI’s published Grok prompts tell the model not to moralise, so filters on general subjects are light. A full debate participant, but not a synthesis candidate in PolyCog (owner call, 2026-08).
Grok 4.7grok-4.7xAI’s flagship: live web and X search, a 500K window, and xAI’s own emphasis on code, agentic tool calling and minimal hallucinations. Same rate card as 4.6. Not a synthesis candidate in PolyCog.$2.00 / $6.00 per MTokMinimal
grok-4.7Reach for it when
- today’s news
- X/social research
- coding
- subjects other labs refuse (see the table)
Look elsewhere for
- writing the final synthesis (not eligible in PolyCog)
Grok 4.6grok-4.6The prior flagship at the same price, with the same live-web and X reach. A step behind 4.7 on coding and agents.$2.00 / $6.00 per MTokMinimal
grok-4.6Reach for it when
- today’s news
- X/social research
- coding
- subjects other labs refuse (see the table)
Look elsewhere for
- new sessions (pick 4.7)
Grok 4.5grok-4.5A prior model at the same price, with the same live-web reach. xAI pitches it at agentic software engineering.$2.00 / $6.00 per MTokMinimal
grok-4.5Reach for it when
- current events
- coding
Look elsewhere for
- new sessions (pick 4.7)
Grok 4.3grok-4.3default seatThe Grok default seat: xAI’s fast tool-calling model and the cheapest Grok. Candid, quick, with live web.$1.25 / $2.50 per MTokMinimal
grok-4.3default seatReach for it when
- news checks
- a cheap live-web seat
Look elsewhere for
- long careful writing
Wildcard · OpenRouter
The sixth seat. One OpenRouter key opens Qwen, Kimi, GLM, Mistral, Meta’s Llama and Muse, MiniMax, Venice and Nous: mostly open-weight and non-US models, most of them very cheap, with OpenRouter’s web plugin for search. OpenRouter passes each lab’s terms (and the serving host’s) through, so guardrails vary by lab and by host. The Chinese labs’ terms are strict on paper about sexual, violent and insulting content, and their models decline PRC-sensitive topics. Mistral’s policy leaves adult content alone but restricts drug-use information and professional advice. Llama’s licence restricts drugs, weapons and professional practice, and Meta’s Muse terms do the same and bar adult sexual imagery. Venice (uncensored by its model card) and Nous (trained for a low refusal rate) sit at the minimal end of the scale, though their hosts’ terms bar pornographic or obscene content. BYOK only, and never the synthesis.
Qwen · Alibaba Light PRC topics: strict
Qwen4 Maxqwen/qwen4-maxAlibabaAlibaba’s flagship since Nov 16, 2026, closed-weight, with a million-token window, image and video input and reasoning on by default. Alibaba aims it at coding, knowledge work and long agent tasks and says it finishes the same work with fewer tokens than Qwen3.8 Max, at a lower rate. A non-US second opinion.$1.425 / $4.275 per MTokLight
qwen/qwen4-maxAlibabaReach for it when
- reasoning
- long inputs
- a non-US second opinion
Look elsewhere for
- PRC-sensitive topics
Qwen3.8 Maxqwen/qwen3.8-max-0902AlibabaAlibaba’s prior flagship: a 2.4-trillion-parameter MoE with a million-token window and a reasoning mode. Qwen4 Max is its newer, cheaper successor. A non-US second opinion.$2.00 / $6.00 per MTokLight
qwen/qwen3.8-max-0902AlibabaReach for it when
- reasoning
- long inputs
- a non-US second opinion
Look elsewhere for
- PRC-sensitive topics
Qwen3.8 Flashqwen/qwen3.8-flashAlibabadefault seatFast, cheap Qwen with reasoning, and the Wildcard default.$0.15 / $0.47 per MTokLight
qwen/qwen3.8-flashAlibabadefault seatReach for it when
- cheap multilingual volume
Look elsewhere for
- PRC-sensitive topics
Kimi · Moonshot Light PRC topics: strict
Kimi K3.1 Promoonshotai/kimi-k3.1-proMoonshotMoonshot’s flagship: a post-trained Kimi K3 with a million-token window, priced below K3. Moonshot aims it at long-horizon coding, 3D and CAD work and financial modelling, and says the open weights follow by November 4.$2.00 / $10.00 per MTokLight
moonshotai/kimi-k3.1-proMoonshotReach for it when
- long-horizon coding
- 3D and CAD work
- financial modelling
Look elsewhere for
- PRC-sensitive topics
- bulk
Kimi K3moonshotai/kimi-k3MoonshotMoonshot’s flagship: a 2.8-trillion-parameter open-weight model with a million-token window. Moonshot pitches it at complex coding, knowledge work and long-running agents. Kimi K3.1 Pro is its newer, cheaper successor.$3.00 / $15.00 per MTokLight
moonshotai/kimi-k3MoonshotReach for it when
- complex coding
- knowledge work
- long-horizon agents
Look elsewhere for
- PRC-sensitive topics
- bulk
Kimi K2.6moonshotai/kimi-k2.6MoonshotMid-priced Kimi for agentic and coding work at a fraction of K3.$0.95 / $4.00 per MTokLight
moonshotai/kimi-k2.6MoonshotReach for it when
- agentic coding
Look elsewhere for
- PRC-sensitive topics
GLM · Z.ai Moderate PRC topics: strict
GLM 5.3z-ai/glm-5.3Z.aiZ.ai’s flagship: a million-token reasoning model that Z.ai aims at complex software engineering and long-running agents.$1.40 / $4.40 per MTokModerate
z-ai/glm-5.3Z.aiReach for it when
- agentic coding
- long inputs
Look elsewhere for
- PRC-sensitive topics
GLM 5.3 Flashz-ai/glm-5.3-flashZ.aiZ.ai’s efficient tier, for coding and agent work at a very low price.$0.15 / $0.50 per MTokModerate
z-ai/glm-5.3-flashZ.aiReach for it when
- bulk agent steps
- cheap coding
Look elsewhere for
- nuance
- PRC-sensitive topics
Mistral · Mistral AI Light PRC topics: none
Mistral Large 4mistralai/mistral-large-4-0Mistral AIMistral’s open-weight flagship, released as a public preview, with a million-token window, image input and optional reasoning. Mistral aims it at reasoning, coding and agentic work. Its policy leaves adult content alone and is strict on professional advice.$0.70 / $2.10 per MTokLight
mistralai/mistral-large-4-0Mistral AIReach for it when
- agentic workflows
- coding
- a non-US voice
Look elsewhere for
- work that needs a settled release (it is a preview)
Mistral Medium 3.5mistralai/mistral-medium-3-5Mistral AIMistral’s frontier-class tier, from a European lab, with configurable reasoning effort. Mistral aims it at agentic workflows and coding. Its policy leaves adult content alone and is strict on professional advice.$1.50 / $7.50 per MTokLight
mistralai/mistral-medium-3-5Mistral AIReach for it when
- agentic workflows
- light-touch creative work
- a non-US voice
Look elsewhere for
- very long inputs
Mistral Small 4mistralai/mistral-small-2603Mistral AIMistral Small 4: quick and cheap, with configurable reasoning.$0.15 / $0.60 per MTokLight
mistralai/mistral-small-2603Mistral AIReach for it when
- cheap drafts
Look elsewhere for
- long inputs
Llama · Meta Light PRC topics: none
Llama 4 Maverickmeta-llama/llama-4-maverickMetaMeta’s open-weight MoE and the open-model voice in the debate, low-priced. The window (up to a million tokens) and the filtering depend on the host serving it.$0.35 / $1.15 per MTokLight
meta-llama/llama-4-maverickMetaReach for it when
- an open-weight opinion
- cheap long context
Look elsewhere for
- frontier reasoning
Muse · Meta Light PRC topics: none
Muse Wondermeta/muse-wonderMetaMeta’s frontier reasoning model from its Superintelligence Labs, closed-weight, with a million-token window and text, image, audio and video input. Meta aims it at long agentic tasks and professional work in law, finance and health. A hard question can take a while to start answering.$4.00 / $20.00 per MTokLight
meta/muse-wonderMetaReach for it when
- long agentic tasks
- professional documents
- a second US frontier opinion
Look elsewhere for
- bulk or high-volume jobs
- quick lookups
Muse Spark 1.3meta/muse-spark-1.3MetaMeta’s mid-tier Muse reasoning model, closed-weight, with a million-token window and image and video input. Meta aims it at long-running agentic and coding work, at about a third of Muse Wonder’s price.$1.25 / $4.25 per MTokLight
meta/muse-spark-1.3MetaReach for it when
- agentic coding
- long inputs
- a US mid-tier opinion
Look elsewhere for
- the hardest reasoning
Muse Embermeta/muse-emberMetaMeta’s smallest Muse model, closed-weight, with a million-token window and text, image, audio and video input. Meta aims it at the steps agents repeat all day: tool calls, coding subtasks, extraction and long-document parsing.$0.10 / $0.50 per MTokLight
meta/muse-emberMetaReach for it when
- cheap agent steps
- long inputs
- audio and video input
Look elsewhere for
- the hardest reasoning
- long-form writing
MiniMax · MiniMax Moderate PRC topics: strict
MiniMax M3minimax/minimax-m3MiniMaxMiniMax’s million-context generalist: cheap, and pitched at long-running agentic work and coding.$0.30 / $1.20 per MTokModerate
minimax/minimax-m3MiniMaxReach for it when
- cheap generalist
Look elsewhere for
- PRC-sensitive topics
MiniMax M3.1 Flashminimax/minimax-m3.1-flashMiniMaxMiniMax’s faster M3 at half its price: the same million-context mixture-of-experts with sparse attention, aimed at agentic work, coding and computer use.$0.15 / $0.60 per MTokModerate
minimax/minimax-m3.1-flashMiniMaxReach for it when
- cheap agent steps
- cheap coding
Look elsewhere for
- PRC-sensitive topics
Venice · Cognitive Computations · Venice Minimal PRC topics: none
Venice Uncensoredcognitivecomputations/dolphin-mistral-24b-venice-editionCognitive Computations · VeniceA 24B Dolphin fine-tune served by Venice and described as uncensored, built to answer without refusals (dark roleplay, unsoftened answers). Venice’s terms still ask that content not be pornographic, so the table marks sexual content may-soften. A small model with weak reasoning and no tools.$0.20 / $0.90 per MTokMinimal
cognitivecomputations/dolphin-mistral-24b-venice-editionCognitive Computations · VeniceReach for it when
- dark fiction
- roleplay without refusals
- unsoftened answers
Look elsewhere for
- reasoning
- code
- facts without the 🔍
Nous · Nous Research Minimal PRC topics: none
Hermes 4 405Bnousresearch/hermes-4-405bNous ResearchNous Research’s 405B Hermes on a Llama 3.1 base: a frontier-sized open model with a hybrid reasoning mode, trained for steerability and a low refusal rate. A light-handed debate voice. Turn on the 🔍 for anything recent.$1.00 / $3.00 per MTokMinimal
nousresearch/hermes-4-405bNous ResearchReach for it when
- an open-weight heavyweight
- steerable persona work
- subjects the US labs soften
Look elsewhere for
- current events without search
- vision
Archive: models that have left
Every model PolyCog has offered and since removed, with the day it left PolyCog and the day (if any) its provider retired it. Prices for all of them stay on /pricing/.
| Model | Left PolyCog | Retired by provider | Why |
|---|---|---|---|
Gemini 3.7 Flashgemini-3.7-flash |
Oct 10, 2026 | still served | Deprecated by Google, which now routes its requests to Gemini 3.8 Flash at the same price. |
Gemini 3.5 Flashgemini-3.5-flash |
Oct 10, 2026 | still served | Deprecated by Google, which now routes its requests to Gemini 3.6 Flash. |
DeepSeek V4 Flashdeepseek-v4-flash |
Sep 11, 2026 | Sep 10, 2026 | Retired upstream; the id is temporarily routed to V4.1 Flash, which took its place. |
o3o3 |
Aug 10, 2026 | retires Dec 11, 2026 | Dropped ahead of OpenAI’s announced API removal (2026-12-11). |
Claude Opus 4.1claude-opus-4-1-20250805 |
Jul 25, 2026 | Aug 5, 2026 | Retirement notice issued; excluded at the 14.8 catalog refresh. |
GPT-4ogpt-4o |
≈ Jul 2026 | retires Jan 14, 2027 | Deprecated upstream at the time; OpenAI still serves the id with no shutdown date. |
o3-minio3-mini |
≈ Jul 2026 | Oct 23, 2026 | Deprecated upstream; never supported web search. |
Claude Sonnet 4.5claude-sonnet-4-5-20250929 |
≈ Jul 2026 | still served | Superseded by Sonnet 4.6 / 5; still served. |
Claude Opus 4.5claude-opus-4-5-20251101 |
≈ Jul 2026 | still served | Superseded by Opus 4.6+; still served. |
Gemini 2.5 Progemini-2.5-pro |
≈ Jul 2026 | retires Dec 1, 2026 | Closed to new Google accounts; Gemini 3.1 Pro took the Pro slot. |
Gemini 2.5 Flashgemini-2.5-flash |
≈ Jul 2026 | retires Dec 1, 2026 | Closed to new Google accounts (‘no longer available to new users’); grandfathered keys still work. |
DeepSeek Chat (V2 → V2.5 → V3 → V3.1 → V3.2)deepseek-chat |
≈ Jul 2026 | Jul 24, 2026 | Replaced by the V4 line (V4 Pro / V4 Flash). |
DeepSeek Reasoner (R1 → R1-0528 → V3.1/V3.2 thinking mode)deepseek-reasoner |
≈ Jul 2026 | Jul 24, 2026 | Folded into the V4 line’s thinking mode. |
Before PolyCog: 58 models retired by their providers outside PolyCog’s catalog
| Provider | Models | Retired |
|---|---|---|
| ChatGPT | GPT-5.4 nano | Apr 1, 2027 |
| ChatGPT | GPT-5 | Dec 11, 2026 |
| ChatGPT | GPT-3.5 Turbo · GPT-4 (8K) · GPT-4 Turbo · o1 · o4-mini | Oct 23, 2026 |
| ChatGPT | o1-mini | Oct 27, 2025 |
| ChatGPT | o1-preview | Jul 28, 2025 |
| ChatGPT | GPT-4.5 (research preview) | Jul 14, 2025 |
| ChatGPT | GPT-4 (32K) | Jun 6, 2025 |
| ChatGPT | GPT-3 Davinci · GPT-3 Curie · GPT-3 Babbage · GPT-3 Ada · text-davinci-003 | Jan 4, 2024 |
| ChatGPT | GPT-4.1 nano · GPT-5 mini · GPT-5 nano · GPT-5.1 nano · GPT-5.2 · GPT-5.3-Codex | still served |
| Claude | Claude Opus 4 · Claude Sonnet 4 | Jun 15, 2026 |
| Claude | Claude 3 Haiku | Apr 20, 2026 |
| Claude | Claude 3.5 Haiku · Claude 3.7 Sonnet | Feb 19, 2026 |
| Claude | Claude 3 Opus | Jan 5, 2026 |
| Claude | Claude 3.5 Sonnet · Claude 3.5 Sonnet (upgraded, Oct 2024) | Oct 28, 2025 |
| Claude | Claude 2 · Claude 2.1 · Claude 3 Sonnet | Jul 21, 2025 |
| Claude | Claude (v1: claude-1.0 / 1.1 / 1.2 / 1.3) · Claude Instant (1.0 / 1.1 / 1.2) | Nov 6, 2024 |
| Gemini | Gemini 2.0 Flash · Gemini 2.0 Flash-Lite | Jun 1, 2026 |
| Gemini | Gemini 3 Pro Preview | Mar 9, 2026 |
| Gemini | Gemini 1.5 Pro · Gemini 1.5 Flash · Gemini 1.5 Flash-8B | Sep 29, 2025 |
| Gemini | Gemini 1.0 Pro (Gemini Pro) | Feb 18, 2025 |
| Grok | Grok 3 · Grok 4 · Grok Code Fast 1 · Grok 4 Fast (Reasoning) · Grok 4 Fast (Non-Reasoning) · Grok 4.1 · Grok 4.1 Fast (Reasoning) · Grok 4.1 Fast (Non-Reasoning) | May 15, 2026 |
| Grok | Grok Beta · Grok 2 (1212) · Grok 3 Fast · Grok 3 Mini · Grok 3 Mini Fast · Grok 4.20 (Reasoning) · Grok 4.20 (Non-Reasoning) · Grok Build 0.1 | still served |
How this page is made
- Roster
- The same catalog the app runs. When a model is added to or removed from PolyCog, it appears here or moves to the archive on the same deploy.
- Strengths
- Twelve fixed areas, scored 1–5 by PolyCog from provider model cards, published evaluations and what we see in debates. Editorial, dated, revised at every catalog refresh, and written in the same register for every lab.
- Guardrails
- Provider usage policies read on the verified date, plus observed behaviour. Regional censorship is its own axis; the subject-by-subject table is the source the matcher uses.
- Matcher
- A dictionary of project types, computed in your browser. Each category carries weights over the twelve areas and a content class, and each leaf holds the words people actually type. No text you type leaves the page.
- Prices
- From /pricing/, the provider’s own list price; PolyCog adds nothing.
- Data
- Machine-readable at /api/models.json, CC BY 4.0.
Spotted something wrong or out of date? Send a correction and it gets fixed, with the change noted on this page.
Questions people ask
Which AI model is best for coding?
It depends on the job. This is PolyCog’s own read, dated at the top of the page. For hard work on large codebases, the three frontier flagships are peers with different strengths: GPT-6.2 Astra (OpenAI builds it for the hardest projects from start to finish), Claude Fable 5.2 (Anthropic stresses demanding reasoning and long agentic runs) and Gemini 4 Carbon (Google stresses hard reasoning, complex coding, science and security work). For everyday development at a mid-tier price, try GPT-6.2 Sol, Claude Opus 5.5 or Sonnet 5.5, or Gemini 4 Neon, 4 Hydrogen or 3.8 Flash. xAI pitches Grok 4.7, and Z.ai pitches GLM 5.3, at coding and long-running agents. DeepSeek V4 Pro is a low-cost reasoning seat. In PolyCog you can put three of them on the same bug and let them argue.
Which AI is best for current events and news?
The ones with live web search. xAI’s models search the web and X natively. The GPT-6 and 5.x, Claude and Gemini models search too. DeepSeek searches with its own search tool on your key, and the Wildcard labs search through OpenRouter’s web plugin.
Which AI model is the least censored?
On general subjects (borderline but legal questions, contested politics, dark fiction, medical and legal detail), Venice and Nous in the Wildcard seat and Grok filter the least. Most of the other Wildcard labs and DeepSeek sit in the light-to-moderate band, and OpenAI, Google and Anthropic in the moderate-to-firm band. The guardrails table sorts them. The picture flips on China-related topics: DeepSeek, Qwen, Kimi, GLM and MiniMax decline or deflect, while the US and European models answer freely. Where a written policy is stricter than a model’s reputation, this page goes by the written policy.
Which AI will write sexual content?
Going by each provider’s own terms, as read on the verified date: xAI’s policy doesn’t restrict adult content between adults, and Mistral’s doesn’t mention it. OpenAI, Anthropic and Google refuse it. DeepSeek and the Chinese Wildcard labs forbid it in their terms. Llama, Venice and Nous are served under host terms that bar obscene or pornographic content, and Meta’s Muse terms bar adult sexual imagery, so they’re marked “may soften” and drop out when a request is explicit. The box at the top applies exactly this table, so a request like that only returns the families whose terms allow it.
Is DeepSeek censored?
DeepSeek’s written terms forbid sexually explicit, offensive and vulgar content, so the table marks those subjects as refused. On topics the Chinese government treats as sensitive, such as Tiananmen, Taiwan, Xinjiang and the Party’s leadership, it declines or changes the subject. Qwen, Kimi, GLM and MiniMax in the Wildcard seat do the same. This page reports that behaviour. It doesn’t offer ways around it.
Which AI is best for writing?
This is our read, and it’s dated. The three frontier flagships (GPT-6.2 Astra, Claude Fable 5.2 and Gemini 4 Carbon) are peers for long-form and structured writing. Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.2 Sol are the everyday picks, and Kimi K3.1 Pro and Mistral Medium 3.5 are the Wildcard seat’s writers. For anything that has to read well, run two of them and compare.
Which AI can read images and long documents?
Every current Gemini model takes images, video, audio and PDFs with a million-token window; GPT-6.2 Astra, Claude Fable 5.2 and Claude Opus 5.5 read images natively with a million tokens each. DeepSeek V4.1 Flash and sixteen of the nineteen Wildcard models (Qwen, Kimi, GLM 5.3 Flash, Mistral, Llama 4 Maverick, Muse and MiniMax) read images natively too; DeepSeek V4 Pro, GLM 5.3, Venice Uncensored and Hermes 4 405B get them through PolyCog’s vision bridge, which describes each image in words first.
How does the "what are you working on?" box decide?
It matches what you type against a dictionary of project types stored on this page. Each type has a subject class and weights across twelve capability areas. The box checks every provider’s policy for those subjects, drops the ones that would refuse, and ranks the rest by fit and price. It all runs in your browser: nothing you type is sent anywhere, and it keeps working if you lose your connection after the page loads.
How are the strengths and guardrails scored?
By PolyCog, from each provider’s model card and usage policy, published evaluations, and what we see in debates. They are editorial, dated with the verified date at the top, written in the same words for every lab, and revised at every catalog refresh. Corrections are welcome and noted on the page.
What happens to a model that leaves PolyCog?
It moves to the archive on this page with the day it left and, if its provider has retired it, that date too. Its price stays in the price history so old sessions still show what they cost. Saved chats that used it keep their responses.
Does PolyCog favour a provider?
No. The synthesis model is always your explicit pick; there is no house default. The scores here use the same axes and the same words for every seat and are published so you can disagree with them.
Run the whole lineup instead of picking one
PolyCog puts up to six of these models on the same prompt, lets them debate, and has one write the answer. Your own keys, provider list prices, no markup.
Free tier forever · $10/mo · $250 lifetime · your API keys, your bill