Guides › Choose model & effort
Advanced AI model guide: every model, effort levels, best picks
Official pages and independent leaderboards only. New models appear here automatically.
For experienced users: what each tracked model is meant for, where it ranks, how much it can read, which effort levels it has and which plans include it. New to this? Start with the 60-second version.
Best model by task
The top three current models on the one leaderboard we use for each task, with the plans whose official pages name them. Rankings are copied exactly as each leaderboard publishes them; we never blend scores.
| Task | Top current models | Leaderboard |
|---|---|---|
| Writing and chat Plans for this → |
| Arena Text (overall) 7 |
| Coding Plans for this → |
| LiveBench 2.0 Coding 8 |
| Research and reasoning Plans for this → |
| LiveBench 2.0 Reasoning 8 |
| Data, maths and all-round work Plans for this → |
| LiveBench 2.0 Overall 8 |
| Making images Plans for this → |
| Arena Text-to-Image 9 |
| Making videos Plans for this → |
| Arena Text-to-Video 10 |
| Agents and automation | Ranked by what the agent can do, not one leaderboard: see AI agents compared and best AI for agents. | |
How effort levels work, app by app
Higher effort (also called thinking or reasoning) means the model works longer before answering: better on multi-step problems, slower, and it uses more of your plan's allowance. For simple questions it adds little.
| App | What the levels do | Source |
|---|---|---|
| xAI (Grok) | Fast for quick replies, Expert thinks hard, Heavy runs a team of agents, Auto picks Fast or Expert. Expert, Heavy and Auto need a paid tier. | 1 |
| Anthropic (Claude) | Low and Medium stretch your usage; High is the best balance; Extra high is for long coding and agent tasks; Max is the most thorough. Higher effort is slower and uses limits faster. Opus 5.5 and Haiku 5.5 default to Medium; Fable 5.1, Opus 5 and Sonnet 5.5 default to High. | 2 |
| Google (Gemini) | Standard thinking is the default and faster; Extended thinking reasons longer for complex problems; Deep Think (AI Ultra, Pro model) runs parallel reasoning and can take minutes. Bigger models and higher levels use more of your limit. | 3 |
| OpenAI (ChatGPT) | Instant, Medium, High, Extra High, Pro in ChatGPT; low to max on the API. Plus gets Medium and High; Pro adds Extra High and Pro. If you hit a thinking limit, ChatGPT may continue with another model. | 4 |
| Microsoft (Copilot) | Smart adapts to the task; Think Deeper is slower and deliberate; Quick response is fastest. | 5 |
| Perplexity | Best picks a model per query; on paid plans choose a model and turn Thinking on for multi-step questions; Deep Research picks its own models. | 6 |
Leaderboards often test several effort levels of the same model; our rankings show the level next to the name (High, XHigh, Max).
Grok (xAI) models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| Grok 4.7 | Our flagship model for code and everything else: agentic tool calling, minimal hallucinations, configurable reasoning | Highest placing: all-round tests (#10 on LiveBench 2.0 Overall); writing and chat (people's votes) (#96 on Arena Text (overall)) Weaker: writing and chat (people's votes) (#96 on Arena Text (overall)) | 500K tokens Configurable reasoning (API); in the Grok app: Fast · Expert · Heavy | Plan pages don't name it Powers Expert and Heavy in the Grok app on paid X Subscription tiers. API: US$2.00 in / US$6.00 out per 1M tokens 11,1 |
| Grok 4.3 | Older Grok model still offered on xAI's API Watch out: Older generation: Grok 4.7 is xAI's flagship. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | 1M tokens Not published on the page we checked | Plan pages don't name it API: US$1.25 in / US$2.50 out per 1M tokens 11 |
Anthropic models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| Claude Fable 5.1 | Next generation intelligence for long-running agents Watch out: Anthropic's most expensive model; uses limits fastest. | Highest placing: coding (#1 on LiveBench 2.0 Coding); all-round tests (#1 on LiveBench 2.0 Overall) Weaker: reasoning (#5 on LiveBench 2.0 Reasoning) | Up to 1M tokens (varies by model) Low · Medium · High (default) · Extra high · Max; thinking always on | Claude Max, Claude Ultra Max plans (up to 50% of weekly limits); Pro only through paid usage credits. API: US$10 in / US$50 out per 1M tokens claude.com/pricing, 17,2 |
| Claude Opus 5.5 | Daily driver for agentic coding and enterprise work | Highest placing: overall capability (Epoch index) (#1 on Epoch Capabilities Index (ECI)); writing and chat (people's votes) (#2 on Arena Text (overall)) Weaker: reasoning (#4 on LiveBench 2.0 Reasoning) | Up to 1M tokens (varies by model) Low · Medium (default) · High · Extra high · Max; thinking always on | Claude Pro, Claude Max, Claude Ultra Pro and Max (not Free). API: US$4 in / US$20 out per 1M tokens claude.com/pricing, 17,2 |
| Claude Sonnet 5.5 | High-performance model for coding and agents | Highest placing: coding (#5 on LiveBench 2.0 Coding); all-round tests (#5 on LiveBench 2.0 Overall) Weaker: writing and chat (people's votes) (#29 on Arena Text (overall)) | Up to 1M tokens (varies by model) Low · Medium · High (default) · Extra high · Max; thinking can't be turned off | Claude Free, Claude Pro, Claude Max, Claude Ultra All plans, including Free. API: US$2 in / US$10 out per 1M tokens claude.com/pricing, 17,2 |
| Claude Haiku 5.5 | Fastest, most cost-efficient model Watch out: Smallest Claude model: less capable on hard problems. | Highest placing: coding (#7 on LiveBench 2.0 Coding); all-round tests (#11 on LiveBench 2.0 Overall) Weaker: all-round tests (#11 on LiveBench 2.0 Overall) | Up to 1M tokens (varies by model) Low · Medium (default) · High · Extra high · Max | Claude Free, Claude Pro, Claude Max, Claude Ultra All plans, including Free. API: US$0.10 in / US$0.50 out per 1M tokens claude.com/pricing, 17,2 |
Google models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| Gemini 4 Argon | Our flagship reasoning model with a unified 1M context window and long decodes up to 1M tokens | Highest placing: writing and chat (people's votes) (#1 on Arena Text (overall)); reasoning (#1 on LiveBench 2.0 Reasoning) Weaker: coding (#9 on LiveBench 2.0 Coding) | 1M tokens In the Gemini app: Standard · Extended thinking; Deep Think on AI Ultra | Google AI Pro, Google AI Ultra 5x, Google AI Ultra 20x Google AI Pro and Ultra (not Free or AI Plus). API: US$2.00 in / US$10.00 out per 1M tokens 13,12 |
| Gemini 3.8 Flash | Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents and complex enterprise workflows | Highest placing: writing and chat (people's votes) (#9 on Arena Text (overall)) Weaker: Only ranked on one leaderboard we carry | Not published on the page we checked Not published on the page we checked | Plan pages don't name it API: US$0.75 in / US$3.75 out per 1M tokens 13 |
| Gemini 3.6 Flash | Our previous-generation Flash model, balancing speed and multimodal capabilities across general agentic and everyday tasks Watch out: Previous generation: Gemini 3.8 Flash replaces it. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | Not published on the page we checked Not published on the page we checked | Plan pages don't name it API: US$0.75 in / US$3.75 out per 1M tokens 13 |
| Gemini 3.1 Pro Preview | Older Pro model, still listed as a preview on Google's API Watch out: Preview model from an older generation. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | Not published on the page we checked Not published on the page we checked | Plan pages don't name it API: US$2.00 in / US$12.00 out per 1M tokens 13 |
| Gemini 3.5 Flash Lite | Our fastest, most cost-effective 3.5 model for high-throughput execution Watch out: Built for speed and cost rather than hard problems. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | Not published on the page we checked Not published on the page we checked | Plan pages don't name it API: US$0.30 in / US$2.50 out per 1M tokens 13 |
OpenAI models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| GPT-6 Astra | Our most capable model for the most demanding work; OpenAI's pick for complex reasoning and coding Watch out: The most expensive OpenAI API model here; usage-limited even on Pro. | Highest placing: overall capability (Epoch index) (#2 on Epoch Capabilities Index (ECI)); reasoning (#2 on LiveBench 2.0 Reasoning) Weaker: writing and chat (people's votes) (#32 on Arena Text (overall)) | 1.05M tokens (128K max output) low · medium · high · xhigh · max | ChatGPT Plus, ChatGPT Pro 100, ChatGPT Pro 200, ChatGPT Pro 500 In ChatGPT it powers the Pro thinking level (GPT-6 Pro) on Pro plans; Plus gets it in ChatGPT Work and Codex only. API: US$10.00 in / US$50.00 out per 1M tokens 18,4 |
| GPT-6.1 Sol | Near-Astra performance for complex work at a lower cost | Highest placing: overall capability (Epoch index) (#3 on Epoch Capabilities Index (ECI)); reasoning (#3 on LiveBench 2.0 Reasoning) Weaker: writing and chat (people's votes) (#27 on Arena Text (overall)) | 1.05M tokens (128K max output) low · medium · high · xhigh · max | ChatGPT Plus, ChatGPT Pro 100, ChatGPT Pro 200, ChatGPT Pro 500 In ChatGPT it's a Work and Codex option, not a Chat option. API: US$2.00 in / US$10.00 out per 1M tokens 18,4 |
| GPT-6 Luna | Our most efficient model for focused, high-volume tasks Watch out: Built for cost and volume rather than the hardest problems. | Highest placing: writing and chat (people's votes) (#91 on Arena Text (overall)) Weaker: Only ranked on one leaderboard we carry | 1.05M tokens (128K max output) none · low · medium · high · xhigh · max | ChatGPT Free, ChatGPT Go, ChatGPT Plus, ChatGPT Pro 100, ChatGPT Pro 200, ChatGPT Pro 500 API: US$0.10 in / US$0.50 out per 1M tokens 18,4 |
| GPT-5.6 Sol | Complex work across coding, knowledge work and research, cybersecurity, science, computer use and design Watch out: Previous generation: GPT-6 is now the default in Chat. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | Not published on the page we checked In ChatGPT: Medium · High · Extra High · Pro (GPT-5.6 Sol Pro), by plan | ChatGPT Plus, ChatGPT Pro 100, ChatGPT Pro 200, ChatGPT Pro 500 Still selectable in ChatGPT on paid plans; not on Free or Go. API: US$4.00 in / US$20.00 out per 1M tokens 4 |
Mistral models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| Mistral Large 4 | A state-of-the-art, open-weight, general-purpose multimodal model | Highest placing: writing and chat (people's votes) (#111 on Arena Text (overall)) Weaker: Only ranked on one leaderboard we carry | Not published on the page we checked Not published on the page we checked | Plan pages don't name it API: US$0.68 in / US$2.09 out per 1M tokens 14 |
| Mistral Medium 3.5 | Our frontier-class multimodal model optimized for agentic and coding use cases | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | 256K tokens Not published on the page we checked | Plan pages don't name it API: US$1.5 in / US$7.5 out per 1M tokens 14 |
| Mistral Small 4 | Hybrid model unifying instruct, reasoning and coding in a single efficient model Watch out: Small model: less capable on the hardest tasks. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | Not published on the page we checked Not published on the page we checked | Plan pages don't name it API: US$0.15 in / US$0.6 out per 1M tokens 14 |
DeepSeek models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek's larger V4 model (DeepSeek-V4-Pro-0813) Watch out: Off-peak API prices are half the peak price. | Highest placing: Not on the leaderboards we carry yet Weaker: Not ranked yet | 1M tokens Not published on the page we checked | Plan pages don't name it Free in the DeepSeek chat app (no paid consumer plan). API: US$1.32 in / US$3.96 out per 1M tokens 15 |
| DeepSeek V4.1 Flash | DeepSeek's Flash model (now DeepSeek-V4.1-Flash-1017) | Highest placing: writing and chat (people's votes) (#34 on Arena Text (overall)) Weaker: Only ranked on one leaderboard we carry | 1M tokens Not published on the page we checked | Plan pages don't name it Free in the DeepSeek chat app (no paid consumer plan). API: US$0.30 in / US$1.20 out per 1M tokens 15 |
Meta AI models
| Model | Meant for | Best at / weaker at (rankings) | Context · effort | Plans · API price |
|---|---|---|---|---|
| Muse Wonder 1.0 | Meta's larger Muse model Watch out: Standard mode only on the API (your content isn't used for training). | Highest placing: all-round tests (#7 on LiveBench 2.0 Overall); reasoning (#10 on LiveBench 2.0 Reasoning) Weaker: coding (#12 on LiveBench 2.0 Coding) | 1M tokens Not published on the page we checked | Plan pages don't name it API: US$4.00 in / US$20.00 out per 1M tokens 16 |
| Muse Spark 1.3 | Meta's Muse Spark model Watch out: Cheaper Contributor mode lets Meta train on your data; Standard mode doesn't. | Highest placing: coding (#8 on LiveBench 2.0 Coding); writing and chat (people's votes) (#10 on Arena Text (overall)) Weaker: reasoning (#12 on LiveBench 2.0 Reasoning) | 1M tokens Not published on the page we checked | Plan pages don't name it API: US$1.25 in / US$4.25 out per 1M tokens 16 |
When an older model is still better
When is an older model better?
Newer usually wins, but not always. On these leaderboards an older version currently ranks above the newest version of the same model family:
- Claude Fable 5 ranks #4 on Arena Text (overall), above the newer Claude Fable 5.1 (#5).
- Grok 4.5 ranks #52 on Arena Text (overall), above the newer Grok 4.7 (#96).
Older models can also be cheaper, faster or use less of your limit. OpenAI says the GPT-6 rollout doesn't retire older selectable models in ChatGPT (OpenAI4). If a new model does worse on something you do often, try the same prompt on the older one.
How this page is built
How is this page built?
- "Meant for", context, inputs/outputs and effort levels: the company's own model and help pages (linked in each row).
- "Best at" and "weaker at": the model's best and worst placings on the independent leaderboards we carry. We don't run our own tests or invent benchmarks.
- Plans: only plans whose official page names the model.
- When a new model is added to our price data it appears here automatically; its description is filled in at the next weekly check.
Sources (18)
- grok.com
- Claude Help Center: Model, effort and thinking settings
- Gemini Apps Help: Usage limits
- OpenAI Help: GPT-6 and other models in ChatGPT
- Microsoft Support: Copilot conversation modes
- perplexity.ai/help-center/en/articles/10354919-what-advan…
- arena.ai/leaderboard/text
- livebench.ai
- arena.ai/leaderboard/text-to-image
- arena.ai/leaderboard/text-to-video
- docs.x.ai/developers/models
- gemini.google: subscriptions
- ai.google.dev/gemini-api/docs/models
- docs.mistral.ai/getting-started/models/models_overview
- api-docs.deepseek.com/quick_start/pricing
- dev.meta.ai/docs/pricing-rate-limits
- platform.claude.com/docs/en/about-claude/models/choosing-…
- platform.openai.com/docs/models
Last checked Oct 26, 2026.