Trick for optimizing your Grok usage:
Install @Zapier MCP and have Grok build automations instead of agents. In our modeling, using AI only for judgment steps cut workflow run costs by 71%
Wade Foster, CEO of @zapier, on The Product Podcast:
“Turns out the best no code product was always just an agent writing code for you.”
Zapier grew up as the no-code company. It’s now a code product. Users just never see it.
Zapier has changed a lot to match this. If you still picture it as clicking through steps manually, you’re in for a surprise.
Connect @Zapier MCP to any AI and ask it to build and host a workflow: zapier.com/mcp
The future of no-code is code.
But now AI writes the code, so you don't have to.
"No-code" made a promise: you could get working software without writing it yourself. That's still true! But instead of clicking through boxes on a screen you have an agent write the code for you.
Now our customers do the same with workflows. More and more of them build a workflow just by asking their AI tool.
Asking an AI to build something feels like a different category from no-code. But the person on the other end gets the same result: working software from someone who doesn't write code.
My rule: the value of AI is having it build your automations, not run them. On the running side, run it with code. Code is deterministic, reliable, and cheap. Whereas AI automations are probabilistic and far too often break in strange and unexpected ways.
OpenAI released GPT-6.2 Sol and GPT-6.2 Astra today.
They matched their best on AutomationBench for an eighth of the cost.
Sol, the cheaper of the two, scored 41.2% on AutomationBench at High effort, for 21 cents a task. The September record, GPT-6 Astra, scored 41.4% at Max effort for $1.73 a task (close enough that I'm calling it a tie)
GPT-6.2 Astra, @OpenAI's new top model, scored 52.4% at Max effort for 83 cents a task. That's 11 points past the September record, at less than half the price.
Insane progress. But cheap only matters if the work comes out right.
In one of our finance tests, the tax team posted 3 changes to this year's 1099 rules in a team chat, and the written guide never got updated. Sol found the changes, held one contractor under legal review, and exempted another whose tax status had changed. The previous version of Sol never opened the chat and sent the tax director an overstated total.
That's why we track cost per finished task at Zapier, and why we built our own benchmark. Labs publish their own scores. I want to know, on work that looks like ours, which steps in a workflow need a model at all, and the cheapest model that gets those steps right.
In most of my personal workflows, code runs the trigger, the data pull, and the send. An AI model only handles the judgment call in the middle. That step is where new models like these pay off: Sol can now handle it at September's record score for about an eighth of the price.
Any workflow shelved because the model step cost too much is worth another look.
Try GPT-6.2 Astra and Sol in @Zapier today.
Google's Gemini 4 Hydrogen is out today.
Argon-level scores on @Zapier's AutomationBench, at Flash-level pricing.
At high effort, Hydrogen scored 51.1%. Gemini 4 Argon scored 51.3%. Hydrogen costs about $0.82 a task to run, against $1.63 for Argon.
@GoogleDeepMind priced it the same as Gemini 3.8 Flash. Flash's best score is 29.7%, and Hydrogen beats that even on low effort (30.6%).
It's not the top scoring model. Gemini 4 Carbon (53.9%) and Claude Fable 5.2 (51.4%) still score higher. But they also cost more: about $1.94 a task for Carbon and $4.64 for Fable 5.2.
Best fit for Operations (64%) and Finance (59%). Weaker for Support (42%) and HR (34%).
In one sales workflow we tested, several urgent customer emails arrived at once, including a full production outage at one account. Hydrogen opened a high-priority case for the outage, alerted the account owner, and added a matching report to the existing incident instead of opening a duplicate. Gemini 3.8 Flash handled the first email and stopped. The outage never got a case, and nobody told the account owner.
If a workflow runs on Argon today, I'd test it on Hydrogen next. For high-volume work, Medium effort scored 46.7% at only $0.41 a task.
Zaps now fix themselves, with AI.
Build your workflow in Claude, ChatGPT, etc. Zapier keeps it running.
When it hits an error? An agent reads it, tests a fix, and ships a new version.
Small fixes ship on their own. Anything bigger, or anything touching a login, waits for your OK.
All part of the new @Zapier. Describe the workflow, let your agent build it & host it in Zapier. We'll fix it when it breaks.
Just connect your AI to Zapier MCP: zapier.com/mcp
Agents now attempt my tasks before I even see them.
Email me, tag me, etc? My agent takes first crack at solving and responding.
I review (and often edit) every message, so it's still mine. Nothing goes out without my green light, and I'm saving a ton of time. Here's the setup:
- Anything where I get tagged or need to take action (Slack, email, Granola, etc) routes through a Zap into Todoist
- Each task kicks off a @Zapier cloud agent that's authenticated to all my tools: Slack, Gmail, Glean, Databricks, HubSpot, you name it
- It assigns an agent to attempt the task, propose what it thinks I'd do, and drafts the reply using a 'writing-as-Wade' skill I created.
- At my desk, I run a /sweep command in Cursor and blitz through all the low effort stuff: send, done, or edit.
- On my phone, I use a private Slack channel I created. "🚀" means send, "✅" means done, and a "🧵" reply means I have edits. The rest is usually voice-to-text, so I can handle stuff while I'm out.
It's my own agent factory, and it's working pretty dang good. But it only works because I worked really hard to make it sound like me. When AI dresses up a thought with the "veneer of smartness," it bugs me to no end.
I'd rather just get the sloppy human version of the output than the dressed up Claude version.
Fable 5.2 is here.
On Zapier workflows, it performs 2.3x better than Fable 5.1. It's the best
@AnthropicAI model we've tested to date.
At max effort, Fable 5.2 scores 51.4% on Zapier's AutomationBench.
@Claudeai's Sonnet 5.5, their previous best, scored 44.6%. Opus 5.5 scored 40.0%, and Fable 5.1 scored 22.4%. Huge jump over the previous Fable model, and a clear step past Sonnet and Opus too.
You don't need max effort to see the jump. Even at low effort, Fable 5.2 scores 39.7%, roughly the same as Opus 5.5 at max effort, for about the same cost per task ($1.30 vs $1.28).
Fable 5.2's biggest gains over 5.1:
1. Finance: from 16% to 53%
2. Sales: from 21% to 54%
3. HR: from 9% to 37%
Finance is also where it has the largest improvement over Opus 5.5 (36% to 53%)
Biggest difference with Fable 5.2 is how it works on each task. It averages about 97 tool calls per task, against 54 for Fable 5.1. It works harder and checks its work before it touches anything.
You do pay for that extra thoroughness. At max effort, Fable 5.2 averages about $4.64 per task, against $1.28 for Opus 5.5. So figure out which workflows need Fable 5.2, and run the rest on Opus 5.5 ($1.28 per task) or Sonnet 5.5 ($1.00 per task)
Fable 5.2 is live on @Zapier now.
We run every new model through Zapier’s AutomationBench: 657 workflows across 6 departments, to tell you what AI to use for your business.
Read the full leaderboard here: zapier.com/benchmarks
An agent shouldn't run your whole workflow.
Mine get the messy middle.
In fact most of my workflows follow the same 4-step shape: trigger, collect data, agent work, send it out.
The trigger, the data collection, and the 'send' run the same way every time. Code handles it all faster and cheaper than a model would.
Then the AI agent reads a pile of context and decides what matters. That's the only part that needs judgment, and the only part worth paying for AI for.
Keeping the model on the judgment step also keeps the bill down. Workflows that use AI only for judgment calls 𝗰𝗼𝘀𝘁 𝟳𝟭% 𝗹𝗲𝘀𝘀 𝘁𝗼 𝗿𝘂𝗻 than ones that send every step through a model.
One of my favorite workflows preps my to-do queue and drafts potential Slack replies in the background. The drafts show up right in Slack, so I just review, edit, and send.
The agent in that middle step is a hosted cloud agent. It carries context from run to run, which makes it a much stronger AI step than a one-off prompt.
I save the AI for the part that needs judgment, and let cheap code and automation handle everything else.
Any agent that connects to Zapier MCP can do that middle step: Claude Code, ChatGPT, Muse, Grok Bot, etc.
Which step in your workflows actually needs the agent?
With Next Gen Zaps, @Zapier figures that out for you. Repeat steps run as code, so you only pay AI for the steps that need it.
Next Gen Zaps fix themselves when they break.
In just one week, that saved 100,000 tasks that would have failed.
How it works: Say an app renames a field and your workflow fails. An agent reads the error, works out a fix, tests it, and publishes a new version of the Zap. Then it emails you what changed.
(Don't worry, the agent still checks in when it should. Small fixes go through on their own, but the agent asks you whenever a fix would make any big changes or touch a login)
For most of our history, a broken Zap meant a different email: your Zap broke, you gotta go fix it. We've sent a lot of those 😅
But more and more of our customers build workflows just by asking Claude, ChatGPT, Muse, Grok Bot, etc. That’s what we call Next Gen Zaps.
If you never touched the steps, you shouldn't be the one fixing them at 2am.
Connect @Zapier MCP anywhere you build with AI.
Something I'm hearing in more meetings: "Saying this for the agent."
Since AI call recorders showed up, people use that line to say the uncomfortable thing. Naming the elephant in the room. Asking someone to commit to a deadline and deliverables.
It's like the 2026 version of "with all due respect"
Part of it works like an accountability buddy. It reminds you what's actually needed from a call. Wanting a quick, painless AI summary pushes us to ask for the uncomfortable truths. Some of this might have felt a little harsh to say to another human but with an agent in the room, you get away with it.
So I've started inviting it. "If you have any concerns, say it now so the agent scoops it up." Now it's part of the process.
Basically "Speak now or forever hold your peace" 😂
The team, and our agents, seem better off for it.
Agents are "probabilistic." That means they 𝘱𝘳𝘰𝘣𝘢𝘣𝘭𝘺 work.
Your business processes need to be 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝘁𝗶𝗰.
Ask an agent to do the same task twice and you'll get two different runs. Even if it gets the same result, it likely got there two different ways. That's useful for research and brainstorming. But updating the CRM, sending the invoice, routing the ticket? Those need to happen the same way every time, or you end up with bad data.
With Zapier, your agent builds the automation, and Zapier runs it. The agent does the part that needs judgment: it reads your request, figures out the steps, and sets it up. Then the automation runs on a trigger, like a new email, with the same steps every time. If a step breaks, Zapier switches back to an agent to figure out the fix.
It's cheaper too. Using AI only where it's needed cuts run costs by 71%.
Install @Zapier wherever you use AI (Claude, ChatGPT, Grok Bot, Muse, etc) and build an automation from a single prompt.