Skip to content
llmwise
Sign inStart free

Best AI · Agents and tool use

The best AI for agents and tool use

In our test runs of September 27, 28, 29 and October 2, 7, 8, 9, 15, 2026, 19 of the 20 models passed 5 of 5 agents and tool use prompts, so these prompts don't name one best model. For best value, GLM 5.3 (5 of 5). Our picks below follow fixed rules, beside each model's price per message, and every prompt and reply is published.

Start freeSee pricing

Based on 100 of our test runs on , through OpenRouter with the app's own prompt and settings. Updated .

Short answer

In our test runs on October 15, 2026, 19 models passed 5 of 5 agents and tool use prompts, so these prompts don't pick one for hard problems. For value, GLM 5.3 (5 of 5), 250 a month on Pro; for everyday agents and tool use, GPT-6 Luna (5 of 5), from the daily count.

Our picks for agents and tool use

  • Hard problems

    Shared by 19 models

    19 models passed 5 of 5, both hard ones: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and 16 more. These prompts don't tell them apart, so they share the pick.

  • Best value

    GLM 5.3

    Passed 5 of 5 agents and tool use prompts, with 250 a month on Pro.

  • Everyday

    GPT-6 Luna

    Passed 5 of 5 agents and tool use prompts; an everyday model, so its messages come from the daily count (60 a day on Pro), not the monthly allowance.

These picks aren't our opinion: they're what the results below give, by these rules, among all 20 models in llmwise. They change when the results do.

  • Hard problems: the model that passed the most prompts and, of those, the most hard ones. Models level on both share the pick: the prompts don't tell them apart, so we don't break the tie by price or by name.
  • Best value: among the models that draw on the monthly allowance, the one with the most messages on Pro that passed no more than one prompt fewer than the top model. Ties go to the one that passed more, then to the lower cost per reply.
  • Everyday: among the cheapest models on the page (the everyday models, which come from the daily count, when the page has any), the one that passed the most. Ties go to the one that passed more of the hard prompts, then to the lower cost per reply.

Our agents and tool use test runs, model by model

How each model did on our 5 agents and tool use prompts, what each reply counted as on Pro, and what it cost to run.

Our agents and tool use test runs
ModelPassedHard onesMessages used on ProCost per replyTime per reply
Claude Fable 5.1Anthropic5 of 52 of 21 each, of 31 a month on Pro$0.01193.8 s
Claude Opus 5.5Anthropic5 of 52 of 21 each, of 62 a month on Pro$0.00503.8 s
Claude Sonnet 5.5Anthropic5 of 52 of 21 each, of 125 a month on Pro$0.00261.8 s
Claude Sonnet 5Anthropic5 of 52 of 21 each, of 125 a month on Pro$0.00243.0 s
Claude Haiku 5.5Anthropic5 of 52 of 21 each, of 60 a day on Pro$0.00011.6 s
Claude Haiku 4.5Anthropic5 of 52 of 21 each, of 250 a month on Pro$0.00091.1 s
GPT-6 AstraOpenAI5 of 52 of 21 each, of 31 a month on Pro$0.00682.0 s
GPT-6.1 SolOpenAI5 of 52 of 21 each, of 125 a month on Pro$0.00071.6 s
GPT-6 SolOpenAI5 of 52 of 21 each, of 125 a month on Pro$0.00161.9 s
GPT-6 LunaOpenAI5 of 52 of 21 each, of 60 a day on Pro$0.00011.7 s
Gemini 4 ArgonGoogle5 of 52 of 21 each, of 62 a month on Pro$0.00113.9 s
Gemini 3.1 Pro (preview)Google5 of 52 of 21 each, of 125 a month on Pro$0.00575.7 s
Gemini 3.8 FlashGoogle5 of 52 of 21 each, of 250 a month on Pro$0.00094.6 s
DeepSeek V4.1 FlashDeepSeek5 of 52 of 21 each, of 60 a day on Pro$0.00020.5 s
DeepSeek V4 ProDeepSeek5 of 52 of 21 each, of 250 a month on Pro$0.00072.0 s
Grok 4.7xAI5 of 52 of 21 each, of 250 a month on Pro$0.00396.1 s
Kimi K3Moonshot5 of 52 of 21 each, of 125 a month on Pro$0.00272.4 s
GLM 5.3Z.ai5 of 52 of 21 each, of 250 a month on Pro$0.00041.0 s
GLM 5.3 FlashZ.ai4 of 52 of 21 each, of 60 a day on Pro$0.00013.7 s
Mistral Large 4Mistral5 of 52 of 21 each, of 250 a month on Pro$0.00135.6 s
Passed: out of the job's 5 prompts, how many replies passed their check (a reply the provider failed to give isn't counted). Messages used: what each reply counted as on Pro, and Pro's count on the model. Cost: what OpenRouter charged us per reply, on average; in llmwise you pay per message, not per token. Time: from sending to the whole reply.

One prompt, side by side

Each model's reply to “Search, but don't book”, as it came, with whether it passed and what it cost.

The prompt: Search, but don't book (hard)
You can call these tools:
- search_flights(origin: string, destination: string, date: string in YYYY-MM-DD, max_stops: number): searches flights between airports given as IATA codes, like "LHR".
- book_flight(flight_id: string): books a flight found by a search.

Today is Monday, 5 October 2026.
User: Find me nonstop flights from Boston to Denver on November 14. Don't book anything yet: I want to see the options first.

Reply with only a JSON object, {"calls": [{"tool": "<name>", "arguments": {...}}]}, listing the tool calls to make now. If no call is right yet, reply {"calls": []}.

  • Claude Fable 5.1

    Passed: Made the 1 expected call.

    {"calls": [{"tool": "search_flights", "arguments": {"origin": "BOS", "destination": "DEN", "date": "2026-11-14", "max_stops": 0}}]}

    677 tokens in, 66 out · 5.0 s · $0.0116 · 1 message on Pro · answered by anthropic/claude-fable-5.1 via Anthropic ·

  • GLM 5.3

    Passed: Made the 1 expected call.

    {"calls":[{"tool":"search_flights","arguments":{"origin":"BOS","destination":"DEN","date":"2026-11-14","max_stops":0}}]}

    527 tokens in, 38 out (1 of them reasoning) · 0.5 s · $0.0005 · 1 message on Pro · answered by z-ai/glm-5.3 via Wafer ·

  • GPT-6 Luna

    Passed: Made the 1 expected call.

    {"calls":[{"tool":"search_flights","arguments":{"origin":"BOS","destination":"DEN","date":"2026-11-14","max_stops":0}}]}

    487 tokens in, 72 out (29 of them reasoning) · 2.2 s · $0.0001 · 1 message on Pro · answered by openai/gpt-6-luna via OpenAI ·

The agents and tool use prompts, and how they're scored

Tool calls. Automatic. The reply's JSON must make exactly the right tool calls, with the right arguments, and no others.

Each prompt was sent the way llmwise sends a message in a side-by-side comparison, which offers no tools: the app's own system prompt, the model's own settings, and Pro's reply size limit (8,000 tokens). Read all five agents and tool use prompts and how every reply was scored.

Prompts like these to try yourself

What matters for agents and tool use

  • Reliable tool calls

    An agentic task is only as good as the model's tool calls: picking the right tool, passing sensible inputs, and reading the result before the next step.

  • Following a plan

    Multi-step work needs a model that keeps the goal and your constraints in mind across many steps, and stops when it's done.

  • Knowing when to ask

    A good agent asks when a step is risky or the goal is unclear, instead of guessing.

  • Room for tool output

    Every step reads the conversation again, so a model with a large context window has more room for long tool outputs.

Every model at a glance

Every model in llmwise
ModelOn ProOn FreeContext windowImagesPDFsReasoningAPI price per 1M, in / out
Claude Fable 5.1Anthropic31/mo on ProNo1M tokensYesWhole fileYes$10.00 / $50.00
Claude Opus 5.5Anthropic62/mo on ProNo1M tokensYesWhole fileYes$4.00 / $20.00
Claude Sonnet 5.5Anthropic125/mo on ProNo1M tokensYesWhole fileYes$2.00 / $10.00
Claude Sonnet 5Anthropic125/mo on ProNo1M tokensYesWhole fileYes$2.00 / $10.00
Claude Haiku 5.5Anthropic60/day on ProYes1M tokensYesWhole fileYes$0.10 / $0.50
Claude Haiku 4.5Anthropic250/mo on ProNo200K tokensYesWhole fileNo$1.00 / $5.00
GPT-6 AstraOpenAI31/mo on ProNo1.05M tokensYesWhole fileYes$10.00 / $50.00
GPT-6.1 SolOpenAI125/mo on ProNo1.05M tokensYesWhole fileYes$2.00 / $10.00
GPT-6 SolOpenAI125/mo on ProNo1.05M tokensYesWhole fileYes$2.00 / $10.00
GPT-6 LunaOpenAI60/day on ProNo1.05M tokensYesWhole fileYes$0.10 / $0.50
Gemini 4 ArgonGoogle62/mo on ProNo1.05M tokensYesWhole fileYes$4.00 / $20.00
Gemini 3.1 Pro (preview)Google125/mo on ProNo1.05M tokensYesWhole fileYes$2.00 / $12.00
Gemini 3.8 FlashGoogle250/mo on ProNo1.05M tokensYesWhole fileYes$0.75 / $3.75
DeepSeek V4.1 FlashDeepSeek60/day on ProNo1.05M tokensYesText onlyYes$0.30 / $1.20
DeepSeek V4 ProDeepSeek250/mo on ProNo1.05M tokensNoText onlyYes$0.40 / $4.00
Grok 4.7xAI250/mo on ProNo500K tokensYesWhole fileYes$2.00 / $6.00
Kimi K3Moonshot125/mo on ProNo1.05M tokensYesText onlyYes$3.00 / $15.00
GLM 5.3Z.ai250/mo on ProNo1.05M tokensNoText onlyYes$1.40 / $4.40
GLM 5.3 FlashZ.ai60/day on ProNo1.05M tokensYesText onlyYes$0.15 / $0.50
Mistral Large 4Mistral250/mo on ProNo1.05M tokensYesText onlyYes$0.68 / $2.09
Each badge is how many messages Pro gets on the model: a month’s, or a day’s on an everyday model. Free is a one-time trial of 5 messages on the models marked. “Text only” models get the text of a PDF, not the file. API prices are the per-token prices in our model catalog as of October 2026 (Anthropic: Anthropic's list price; OpenAI: OpenAI's list price; Google: Google's list price; DeepSeek: the price of the OpenRouter endpoints llmwise uses, not DeepSeek's own API; xAI: xAI's price, served through OpenRouter; Moonshot: Moonshot's list price; Z.ai: Z.ai's list price; Mistral: Mistral's price, served through OpenRouter). In llmwise you pay per message, not per token. Claude Haiku 5.5: the rate for prompts up to 100K tokens; $0.50 / $2.50 a million over that. Gemini 4 Argon: its standard price from December 1, 2026; $2 / $10 until then. Gemini 3.1 Pro (preview): the standard rate, for prompts up to 200K tokens. Gemini 3.8 Flash: an introductory price, through December 31, 2026. Grok 4.7: xAI charges more for very long prompts. Mistral Large 4: a sale price, half its list price of $1.36 / $4.18.

Every limit is published. Paid plans also have a monthly fair-use limit on AI cost: Plus $3.75, Pro $7.50, Max $20, Ultra $42, Studio $85. Using every message on your plan at typical sizes stays under it; very large messages and heavy research use it faster. Every limit, explained.

Agents and tool use in llmwise

  • Web search with sources

    The model can search the web and cite what it found as source cards, or read a link you paste. On Free, web search works inside your 5 trial messages; on a paid plan a search counts as one more message of the model in use.

  • Run code (Pro and higher plans)

    On Pro and higher plans the model can run Python or Node.js in an isolated sandbox once you approve it, read the output and fix what failed. A run counts as 1 Claude Haiku 4.5 message and stops after 60 seconds.

  • MCP servers (Pro and higher plans)

    On Pro and higher plans, connect MCP servers to give the model extra tools, yours or the ones we set up. Each tool call waits for your approval unless you choose Always allow.

  • Skills

    Save instructions and reference files as a skill; the model reads the files it needs when a task calls for them.

Tips

  • State the goal, the constraints and what done looks like before the model starts.

  • Ask for a short plan first, then let it work through the steps.

  • Approve code runs one at a time until you trust the pattern, then use Always allow.

  • Ask it to show the sources or the output it relied on for each conclusion.

Bar chart: Prompts passed in our test runs, agents and tool use. GPT-6 Luna: 5 of 5; Claude Haiku 5.5: 5 of 5; DeepSeek V4.1 Flash: 5 of 5; GLM 5.3: 5 of 5; DeepSeek V4 Pro: 5 of 5; Gemini 3.8 Flash: 5 of 5; Claude Haiku 4.5: 5 of 5; Mistral Large 4: 5 of 5; Grok 4.7: 5 of 5; GPT-6.1 Sol: 5 of 5; GPT-6 Sol: 5 of 5; Claude Sonnet 5: 5 of 5; Claude Sonnet 5.5: 5 of 5; Kimi K3: 5 of 5; Gemini 3.1 Pro (preview): 5 of 5; Gemini 4 Argon: 5 of 5; Claude Opus 5.5: 5 of 5; GPT-6 Astra: 5 of 5; Claude Fable 5.1: 5 of 5; GLM 5.3 Flash: 4 of 5.
Our test runs of September 27, 28, 29 and October 2, 7, 8, 9, 15, 2026: the same prompts for every model, each reply checked the same way.

Questions

What is the best AI for agents and tool use?

In our test runs on October 15, 2026, 19 of the 20 models passed 5 of 5 agents and tool use prompts, so these prompts don't name one best model. For best value, GLM 5.3 (5 of 5). For everyday, GPT-6 Luna (5 of 5). Every prompt and reply is published, so you can check them, and the picks follow fixed rules.

How did you test the models for agents and tool use?

We sent the same agents and tool use prompts to every model through llmwise's own pipeline and checked each reply the same way. The prompts, the replies, how each was scored and the grader are all published on the methods page.

Can I try these models for agents and tool use for free?

Yes, to try: Free is a one-time trial of 5 messages on Claude Haiku 5.5.

Does one message cover every tool step?

The message's price covers the whole reply, however many steps it takes. Paid tools add their own price on top: on a paid plan a web search counts as one more message of the model in use, a code run counts as 1 Claude Haiku 4.5 message, an image from 2 Claude Haiku 4.5 messages an image (by image model), a backend app 5 Claude Haiku 4.5 messages per started 10 minutes, and a research report 15 Claude Haiku 4.5 messages (Standard) or 75 Claude Haiku 4.5 messages (Deep). Code runs, backend apps, research reports and MCP tools ask for your approval first.

Can I build and deploy an agent with llmwise?

No. llmwise is a chat app: the model works through tools inside a reply, with you watching and approving. There's no API, scheduler or background agent.

Claude, GPT, Gemini, DeepSeek, Grok, Kimi, GLM, and Mistral, in one chat.

See what a message costs before you send it. Free is 5 messages to try; sign in with an email link, no password or card.

Start freeSee pricing