Choosing an AI Model

Silk works with any model on OpenRouter that supports image input and tools, and it can also run on your ChatGPT subscription (from Silk 1.1.79) or your Claude subscription (from Silk 1.1.81). So a common question is: which one should I use? Public leaderboards don’t answer that well. They test puzzles and chat, not the work you hand Silk: organizing folders, filling in web forms, researching vendors, updating a CRM, fixing code, and following the skills and memories you’ve taught it.

So we built our own benchmark. It runs Silk’s real agent, with the same instructions, tools and safety checks that ship in the app, on the kinds of tasks people actually use Silk for. Then it checks the results the way you would: did the work get done, done correctly, and done without breaking anything?

Our recommendations

You can change your model any time in Settings → Model ID. Copy the ID exactly as shown.

Model	Model ID	Why
DeepSeek V4.1 Flash	deepseek/deepseek-v4.1-flash	The default and our top pick. Passed all 62 runs, the fastest of the accurate models (median 20 seconds per task), about 1 cent per task.
GPT-6 Luna	openai/gpt-6-luna	The backup. 97% of runs passed, the cheapest per successful task, and from a different provider, so it’s a good switch if DeepSeek is ever unavailable.
Qwen 3.7 Plus	qwen/qwen3.7-plus	The dependable last option. The previous default: accurate (97%), but slower and about twice the cost per task.
MiMo V2.6 Pro	xiaomi/mimo-v2.6-pro	For unattended work. Also passed every run, but about 5 times slower, so it suits scheduled and long background jobs where you aren’t waiting.

Also capable, but not our first choice: GLM 5.3 FlashX (z-ai/glm-5.3-flashx, 97% but about twice DeepSeek’s cost), MiniMax M3 (minimax/minimax-m3, 95% but slower and pricier), and GPT-5.6 Luna (openai/gpt-5.6-luna, 94%; GPT-6 Luna does better for less).

Not recommended right now:

• StepFun Step 3.7 Flash overwrote or lost files in three separate runs, once while reporting success.
• Gemini 3.1 Flash Lite is the fastest model tested, but it failed 18% of runs, including scheduling a meeting at 5 AM for one attendee and not saving CRM changes.
• Qwen 3.8 Flash is capable when it can be reached, but its provider throttled it heavily during our tests.

Tip

If Silk feels slow or makes mistakes, check which model you’re on before anything else. Switching to deepseek/deepseek-v4.1-flash fixes a lot of “Silk is slow” and “Silk got it wrong” problems.

What else we found

• Every model resisted a prompt-injection attack. One test page hid instructions telling the AI to send a confidential file to an attacker. None of the ten models did it, and in most runs the model also warned the user about the suspicious page.
• Silk’s own features work well across the board. Memory, skills, MCP integrations and scheduling were handled correctly by 7 of the 10 models in every run. Every OpenRouter model that finished the test asked “which Dana?” before issuing a refund when two customers shared that name. (On a ChatGPT Free or Go plan, the smaller model didn’t. See below.)
• Long, exact jobs separate the models. The hardest task was updating exactly 7 of 30 leads in a CRM. Only 11 of 20 runs got it fully right: some models skipped a lead, changed one they were told to leave alone, or wrote notes without saving the status changes.
• Speed matters in practice. DeepSeek’s median of 21 seconds on the hard tasks compares with 76 seconds for Qwen 3.7 Plus and 138 seconds for MiMo. Over a day of tasks, that’s the difference between Silk feeling instant and feeling slow.

How we tested

• The real agent. Every run used Silk’s Background agent with its built-in tools (files, shell, browser, web search, PDFs, memory, skills, schedules) and its “did you really finish?” checks.
• Real-world tasks. 31 tasks in three groups: everyday work (sorting folders, filling forms, summarizing data, SEO audits, fixing code, writing a flyer), hard multi-step jobs (supplier sync, CRM follow-up, invoice filing, messy data, time-zone scheduling), and Silk’s own features (memory, skills, MCP, schedules, asking before irreversible actions).
• Fair and repeatable. Test websites were served locally so every model saw identical pages, and each model ran every task twice: 620 runs on OpenRouter, plus 94 on ChatGPT plans.
• Graded on outcomes. A run passed only when the result was fully correct. Speed and cost were measured on successful runs only, and costs are what we were actually billed.

Results (both rounds combined, 62 runs per model)

Model	Passed	Median time	Cost per success
DeepSeek V4.1 Flash	62/62 (100%)	20s	$0.011
MiMo V2.6 Pro	62/62 (100%)	100s	$0.014
GPT-6 Luna	60/62 (97%)	38s	$0.009
GLM 5.3 FlashX	60/62 (97%)	45s	$0.025
Qwen 3.7 Plus	60/62 (97%)	53s	$0.022
MiniMax M3	59/62 (95%)	50s	$0.026
GPT-5.6 Luna	58/62 (94%)	37s	$0.018
Qwen 3.8 Flash	55/62 (89%)	106s	$0.008
Gemini 3.1 Flash Lite	51/62 (82%)	16s	$0.025
StepFun Step 3.7 Flash	51/62 (82%)	44s	$0.029

On a ChatGPT plan

From Silk 1.1.79 you can power Silk with your ChatGPT subscription instead of an OpenRouter key. We ran the same benchmark through it on three ChatGPT plans in October 2026, using the model Silk recommends for each plan.

ChatGPT Plus, full benchmark (62 runs), next to the best OpenRouter models:

Model	Passed	Median time	Cost per task
DeepSeek V4.1 Flash (OpenRouter)	62/62 (100%)	20s	$0.011
GPT-6 Luna (OpenRouter)	60/62 (97%)	38s	$0.009
GPT-6.1 Sol (ChatGPT Plus)	62/62 (100%)	54s	Included in your plan

ChatGPT Free and Go, Silk features suite (16 runs each):

Plan	Model	Passed	Median time	Allowance
Free	GPT-6 Luna	14/16 (88%)	24s	About 300 tasks a month
Go	GPT-6 Luna	14/16 (88%)	41s	About 400 tasks a month

What this means:

• Plus matches the best OpenRouter model on quality. It passed every task, including asking “which Dana?” before a refund, in both runs.
• It’s slower. A typical task took about 2.7 times as long as DeepSeek V4.1 Flash. Choose ChatGPT mode to avoid a separate AI bill, and OpenRouter when speed matters most.
• Plus is generous. All 62 tasks used 40% of one 5-hour window and 6% of the weekly allowance: about 150 tasks every 5 hours, or 1,000 a week.
• Free and Go are a step down. They only offer the smaller GPT-6 Luna, and both failed the same test twice: they refunded an order without asking which of two customers named Dana was meant. Everything else passed. The slower Go time most likely reflects server load when we ran it, not the plan.

On a Claude plan

From Silk 1.1.81 you can power Silk with your Claude subscription. We haven’t run the full benchmark on it yet, so there are no Claude results here for now. We’ll add them once we have. Until then, leave the model on Recommended (Sonnet): Opus is more capable but uses up your plan’s allowance faster, and Haiku is quicker but less careful on multi-step work.

Limits of this benchmark

• It covers Silk’s Background agent, which handles most real usage. Guide mode (voice answers with on-screen pointing) and live Act mode (Silk clicking and typing on your screen) weren’t part of this test.
• Models were tested in October 2026. Models, providers, prices and ChatGPT and Claude plan allowances change often, and we plan to re-run the benchmark as new models arrive.
• Your workflows may differ. If a model isn’t working for you, try another from the recommended list.

Free models

OpenRouter also lists free models (their IDs end in :free). Silk can use one if it supports image input and tools, but free models are rate-limited and less accurate, so expect slower and less reliable results on screen tasks. For most people, a paid model with a credit limit on the key is the better way to keep costs low: usage on the recommended models is around a cent per task, and the limit means you can never be charged more than you set.