Anthropic's 2026 product launches, and the one a local-agent builder should actually care about
If you searched for what Anthropic shipped this year, here is the short version, then the part the other write-ups skip. Anthropic spent 2026 moving Claude off the chat box and onto the computer. The models got the headlines. A quieter launch decided whether any of it could run on a machine you control.
In 2026 Anthropic launched, in order: Claude Opus 4.6 (Feb 5), Claude Sonnet 4.6 with a 1M context window (Feb 17), Claude Opus 4.7 generally available (Apr 16), and Claude for Small Business (May 13). Alongside the models it shipped Cowork, a macOS desktop agent, and disclosed the gated Mythos Preview. Opus 4.7 is the newest generally available model as of late May.
Checked against anthropic.com/news.
The 2026 timeline
Reverse the camera and a pattern shows up. The model cadence is steady, roughly one flagship and one mid-tier per quarter. But the non-model launches are the ones that change what Claude does: Cowork makes it act, Claude for Small Business points those actions at QuickBooks and HubSpot. The dates below are the checkable facts; the argument starts after them.
Feb 5, 2026 - Claude Opus 4.6 (GA)
Anthropic's flagship refresh: stronger coding, planning, and debugging, plus better financial analysis. Available on claude.ai, the API, and the major clouds.
Source: anthropic.com/news/claude-opus-4-6. This set the price floor that Opus 4.7 would later match.
Feb 17, 2026 - Claude Sonnet 4.6
The mid-tier model gained a 1M token context window and meaningfully faster throughput than Sonnet 4.5, arriving about two weeks after Opus 4.6.
Apr 16, 2026 - Claude Opus 4.7 (GA)
SWE-bench Verified rose from 80.8 percent to 87.6 percent, image resolution roughly tripled, and a new xhigh effort level landed between high and max. Pricing held at 5 dollars per million input and 25 dollars per million output tokens.
Available the same day across Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Anthropic conceded it trails the unreleased Mythos system.
May 13, 2026 - Claude for Small Business
A package of roughly 15 prebuilt workflows and connectors into QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365, running through Cowork. The launch came with a free AI fluency course and a 10-city training tour.
Across 2026 - Cowork and the gated Mythos Preview
Cowork brought Claude Code's agentic loop into the macOS desktop app for non-coding work. Mythos Preview, a more advanced cybersecurity model, stayed gated to a handpicked set of companies and was not released publicly.
The throughline: every launch moved Claude further from a chat box and closer to an agent that takes actions on a computer.
What every launch had in common
Opus 4.7's jump on SWE-bench Verified, from 80.8 percent to 87.6 percent, is a real coding gain. But notice what Anthropic launched around the model. Cowork is an agent in a desktop app. Claude for Small Business is a set of connectors that let that agent touch invoicing, CRM, and design tools. The story of 2026 is not a smarter chatbot. It is Claude reaching out of the text box and taking actions across software.
There is a catch hiding in all of it. Cowork runs in Anthropic's app. Claude for Small Business runs through Cowork against hosted connectors. The agent loop, the part that decides what to click and what to type, lives on Anthropic's side. For most people that is fine. For anyone who wants the loop running on their own machine, against their own logged-in browser, with source they can read, none of the headline 2026 launches deliver that directly.
The launch most roundups skip: an open protocol, not a model
The 2026 development that lets the agent loop run locally is not a Claude version at all. It is the Agent Client Protocol, ACP, settling into a stable shape over the year. ACP is an open, documented JSON interface between a host application and an agent backend. Because it is open, an app you run yourself can speak it to a Claude backend and get the same loop Anthropic ships in Cowork, without being Cowork.
This is the part you can verify rather than take on faith. Fazm is an open-source native macOS app that wraps Claude Code and Codex through their ACP adapters. Its bridge declares the exact protocol version it depends on:
// fazm/acp-bridge/package.json
"dependencies": {
"@agentclientprotocol/claude-agent-acp": "0.29.2",
"@zed-industries/codex-acp": "0.12.0"
}Two pinned lines. The first wires in the real Claude agent loop, the same one behind Anthropic's 2026 launches, reached through your own Pro or Max account so Opus 4.7 usage hits your existing plan. The second makes Codex a swappable backend per chat. There is no hidden indirection layer: clone the repo, grep for ANTHROPIC_BASE_URL, and you can see the loop is the genuine Claude SDK rather than a reimplementation.
Where the loop actually runs when it runs locally
The difference between a hosted agent launch and a local one is simply which box owns the loop. Here is the local path the same 2026 Claude models travel when the host is an app on your Mac.
A 2026 Claude launch, reached locally over ACP
Your Mac
A native Swift app holds the chat window, your files, and screen context via accessibility APIs.
Local ACP bridge
A Node subprocess fazm spawns, pinned to claude-agent-acp 0.29.2, speaks the open Agent Client Protocol.
Claude Opus 4.7 (your account)
The same 2026 agent loop, reached through your own Pro or Max plan. Codex and other backends are swappable per chat.
Back into your apps
Results act on the browser, native Mac apps, and Google Workspace, not just a transcript.
Because the boundary is an open protocol rather than a closed app, the same picture survives each launch. When Opus 4.7 went GA on April 16, nothing in this diagram had to change shape; the model resolved through the account and rode the existing bridge. The launches add capability at the third box; the first two stay yours.
So what should you take from the 2026 launches?
If you just want the best hosted model, the answer is short: Opus 4.7, generally available since April 16, is Anthropic's most capable shipped model, and Claude for Small Business is the fastest way to point it at common back-office tools. If you run a team that lives in QuickBooks and HubSpot, that May 13 launch is probably the one that matters to you.
But if you are a developer who wants the agent loop on your own hardware, reading your own screen, driving your own logged-in browser, then the model version is not the interesting axis. The interesting axis is whether the loop is reachable over an open protocol you can run and inspect. In 2026 it is, and a pinned two-line dependency proves it. That is the launch the roundups skip, and it is the one that decides whether Anthropic's agentic turn is something you watch or something you own.
Want the 2026 Claude agent loop running on your own Mac?
Walk through how fazm pins the open protocol so each Anthropic launch lands locally, against your account, with source you can read.
Questions about Anthropic's 2026 launches
What has Anthropic launched in 2026, in order?
Claude Opus 4.6 went generally available on February 5, 2026. Claude Sonnet 4.6 followed on February 17 with a 1M token context window and faster throughput than Sonnet 4.5. Claude Opus 4.7 reached general availability on April 16, with SWE-bench Verified climbing from 80.8 percent to 87.6 percent, roughly 3x image resolution over Opus 4.6, and a new xhigh effort level. Claude for Small Business launched on May 13 as a package of prebuilt workflows and connectors running through Cowork. Alongside the models, Anthropic shipped Cowork, a macOS desktop agent that brings Claude Code style agentic work to the desktop app, and disclosed Mythos Preview, a gated cybersecurity model not released to the public.
Is Claude Opus 4.7 the newest model, and what does it cost?
As of late May 2026, Opus 4.7 is Anthropic's most capable generally available model. It launched April 16 at the same price as Opus 4.6: 5 dollars per million input tokens and 25 dollars per million output tokens. The gains over 4.6 are concentrated in hard software engineering tasks and vision. Anthropic publicly conceded that 4.7 does not match the unreleased Mythos system, which it has held back over safety concerns and offered only to a handpicked set of cybersecurity companies as Mythos Preview.
What is Claude for Small Business and does it run on my own Mac?
Claude for Small Business launched May 13, 2026. It bundles roughly 15 ready-to-run workflows and a set of connectors into the tools small businesses already pay for: Intuit QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365. It runs through Claude Cowork, Anthropic's workspace, so the work happens against Anthropic's hosted agent and the partner APIs, not as a process you run and inspect on your own machine. If you want the agent loop itself executing locally on a Mac, with the same Claude models reachable, that is a different architecture, which is the point of the second half of this page.
What is Claude Cowork?
Cowork is Anthropic's desktop agent. It brings Claude Code's agentic capabilities into the Claude desktop app for knowledge work beyond coding, and it is macOS only on the Pro tier. Cowork is the clearest signal of the 2026 direction: Anthropic is no longer shipping only a model you chat with, it is shipping an agent that takes multi-step actions. Claude for Small Business sits on top of Cowork. The open question Cowork raises for builders is whether the same loop can run outside Anthropic's app, which the Agent Client Protocol answers.
Which 2026 Anthropic launch matters most if I want a local agent?
Not a model number. It is the Agent Client Protocol (ACP) reaching a stable point in 2026. ACP is the open protocol that lets a host application drive an agent backend over a documented JSON interface. Because the protocol is open, an app you run yourself can speak it to a Claude backend, a Codex backend, or another provider, and get the same agent loop Anthropic ships in Cowork without being Cowork. Fazm pins @agentclientprotocol/claude-agent-acp at 0.29.2 in its acp-bridge/package.json, which is the concrete, checkable version that wires the 2026 Claude agent loop into a native Mac app you can read the source of.
Do I need a new Anthropic subscription to use Opus 4.7 in a local agent?
No. Fazm lets you bring your own Claude Pro or Max account, so Opus 4.7 usage hits your existing plan rather than a separate API bill. The 2026 model launches resolve through your account, so when Opus 4.7 went GA on April 16 it became selectable without paying twice. There is a separate question of whether an app surfaces a new model the day it ships, which depends on whether the app reads the model list dynamically or hardcodes it. That mechanism is covered in the May model page linked at the end of this article.
Where can I verify the protocol version claim?
Clone github.com/m13v/fazm and open acp-bridge/package.json. The dependency block pins @agentclientprotocol/claude-agent-acp at 0.29.2 and @zed-industries/codex-acp at 0.12.0. There is no hidden indirection. The Node subprocess that fazm spawns to talk to Claude installs exactly those packages, and the Swift app drives them over ACP. You can grep the repo for ANTHROPIC_BASE_URL to confirm the agent loop is the real Claude SDK, not a reimplementation.
Is fazm affiliated with Anthropic?
No. Fazm is an independent open-source project that wraps Claude Code and Codex through their published ACP adapters. It is not built or endorsed by Anthropic. It tracks Anthropic's 2026 launches the way any client of an open protocol does: by depending on the published adapter package and resolving models through your own account.
More on running Anthropic's 2026 releases on your own terms
Keep reading
Anthropic's new model in May 2026, and the one thing that decides whether your tool can use it
Why apps that hardcode a model id need a release before you can pick a new model, and how dynamic resolution avoids it.
Anthropic Claude announcement, April 2026: where the news lands inside a Mac app, one env var deep
The single environment variable that decides whether the April launches arrive through Anthropic, a proxy, or a local bridge.
Local Claude desktop agent
Running the Claude agent loop as a process on your own machine instead of a hosted workspace.
Comments (••)
Leave a comment to see what others are saying.Public and anonymous. No signup.