Merge

yo fleet week was crazy fr

@youdotcom is now a connector in @merge_api Agent Handler.

Agent platforms on Merge can now add 9 You.com tools: search, page contents, cited answers, research, and finance research. Users bring their own API key.

Full Blog: you.com/resources/ydc-…

You.com (@youdotcom) is now available as a connector in Merge Agent Handler!

Agents can search the live web and pull current results into their work, from this morning's news to a competitor's updated pricing page.

Every search runs through the same auth, permissions and logging as the rest of your agent's tools.

Check it out: merge.dev/merge-agent-ha…

Which model uses MCP tools best?

We tested 6 models on the same 30 multi-step tasks, using real MCP servers in Claude Code.

Frontier models like Opus 5.5 led on completion and speed but open models like MiMo cost 87% less per attempt.

How much is Opus’s lead actually worth? 🧵

We just got ranked again on @Paraformtalent's Talent Density Index alongside, @AnthropicAI, @OpenAI, and @ThinkyMachines.

Everyone at Merge owns real problems, and the bar keeps going up because the people you work next to keep raising it!

Thank you to the team and if you want to grow with us, we're hiring.

paraform.com/talent-density…

Claude Haiku 5.5 from Anthropic (@AnthropicAI) is now available on Merge Gateway.

Anthropic calls it their fastest model yet, and early customers are seeing up to 2.5x faster inference per agent turn while costing 75% less than Haiku 4.5.

Check it out: gateway.merge.dev

Two frontier models a few points apart on the Intelligence Index can differ 7x in the tokens they burn per task.

GPT-6 Astra (@OpenAI) scores 52.7 and finishes an average Artificial Analysis Intelligence Index task in 27K output tokens. Claude Sonnet 5.5 scores 56.0 and uses 197K. GPT-6.1 Sol comes in at 38K, while Muse Spark 1.3, Gemini 4 Argon and MiMo-V2.6-Pro all land between 60K and 64K.

Output tokens are what you pay for. Sending simple requests to a leaner model is one of the easiest ways to cut inference spend, and Merge Gateway routes them for you.

Scores via artificialanalysis.ai, week of Oct 5.

.@FastCompany named @Merge_api one of the 7 next big things in foundational AI for 2026.

A year ago we launched Agent Handler on a bet that agents are only as useful as the systems they can safely reach.

Being recognized among the breakthroughs shaping how we work and live is an incredible milestone for our team and the customers building with us. Thank you, Fast Company!

Read the full article: fastcompany.com/91612044/found…

Merge has been named one of @FastCompany's 7 Next Big Things in Foundational AI for 2026!

Agents are only as good as the data they can reach. Merge is the layer that connects them to tools like @Salesforce, @Workday, and @GitHub, while keeping access secure.

Check it out: fastcompany.com/91612044/found…

Mistral Large 4 from @MistralAI is now available on Merge Gateway.

It's a big jump for Mistral. Large 4 scores 38 on the Artificial Analysis Intelligence Index, more than double their previous best.

Open weights ship this month.

gateway.merge.dev

Merge Agent Handler just got 11 new connectors!

- Salesforce Marketing Cloud (@salesforce)
- Salesforce Data 360 (@salesforce)
- Adobe Experience Manager (@Adobe)
- Adobe Workfront (@Adobe)
- Excel GCC High (@Microsoft), for US government tenants
- Tavily (@tavilyai)
- Kit (@kit)
- Checkly (@checklyhq)
- ConfigCat (@configcat)
- Alpha Vantage
- TinyPNG (@Tinypng)

Try them out: docs.merge.dev/merge-agent-ha…

Switching models is scary when the only real test is in production.

Gateway Evals grades a new model on your agent's real tasks, then shadows it on live traffic. You see cost, latency, and every answer before a customer does.

Try it now in Merge Gateway: docs.merge.dev/merge-gateway/…

Gemini 4 Argon (@GoogleDeepMind) makes up an answer about half as often as the next best frontier model.

On AA-Omniscience, when Argon doesn't know something it guesses wrong 15% of the time instead of saying so. Qwen3.8 Max is next at 29%, with Grok 4.7 at 29% and Muse Spark 1.3 at 32%. GPT-6 Astra has the highest accuracy of the six at 61%, but still guesses on 45% of the questions it misses.

Whether a wrong answer or no answer costs you more depends on the task. With Merge Gateway you can send each request to the model that fits it.

Scores via artificialanalysis.ai, week of Oct 5.

We're hosting a Vercel Ship Afterparty with @linear and @supabase!

Come hang out for food and drinks at Campus, just a short walk from the Palace of Fine Arts.

RSVP: luma.com/merge-y2z6

We co-hosted the official afterparty of The AI Conference (@AIconference) with Fireworks (@FireworksAI_HQ) yesterday.

Party in the Park took over The Great Lawn next to Oracle Park with live music, food, and drinks while the sun went down over the Bay.

Thanks to everyone who came out and to the Fireworks team for throwing it with us!

The biggest generational jump since June came from a model that costs $0.18 per million output tokens.

DeepSeek-V4-Flash-0731 (@Deepseek_AI) scored 26.8 points higher on Toolathlon than the release it replaced. GLM-5.3 gained 24.8 on the same board. Claude Opus 5.5 added 20.7 on SWE-bench Pro, and Claude Sonnet 5.5 added 18.1 there six days later.

Scores via llm-stats.com, week of Sep 28.

When an integration stops syncing, the error rarely says how to fix it.

AI-powered issue resolution is now live in Merge Unified. Open any issue and Merge explains what went wrong, drawing on years of supporting integrations in production.

If the fix is on your customer's side, you get an explanation ready to send them. If it's on yours, Merge walks you through the fix in the dashboard.

Learn more: merge.dev/blog/ai-powere…

Sending every agent task to the most capable model gets expensive fast.

@EragonAI now routes each task to the right model through one Merge Gateway endpoint. Customers get the same quality, while Eragon saves hundreds of thousands of dollars in inference.

Check out the full case study: merge.dev/case-studies/e…

Merge Agent Handler is now a partner recipe in NVIDIA (@nvidia) NemoClaw.

A NemoClaw agent reads Workday, org structure, and time off, with no access to payroll tools.

Deploy agents safely without exposing credentials.

The best open-weight model for code is third overall, at a 23rd of the price of the model above it.

MiMo-V2.6-Pro scores 60.9 on SciCode against Claude Opus 5.5's 66.9, and runs $0.87 per million output tokens against $20. Kimi K3 and GLM-5.3 follow within two points of it.

Six points for $19 per million tokens is a real tradeoff, and it's the kind a Build-Your-Own-Router policy on Merge Gateway lets you set per request rather than once.

Scores via artificialanalysis.ai, week of Sep 28.