Meta's "Hatch" AI Agent Runs on Claude While Meta Spends $10B on Anthropic

Mark Zuckerberg spent 6,500 words in August arguing against AI power concentrating in a handful of closed labs — a passage widely read as aimed at Anthropic and Dario Amodei. At the same time, according to the New York Times (via Techmeme, reported August 27, 2026), Meta was internally projecting it could spend as much as $10 billion a year on Anthropic's AI models. And the consumer AI agent Meta is racing to ship — codenamed "Hatch" — reportedly runs on Anthropic's Claude, not Meta's own models, while it's in development.

None of this is officially confirmed by either company. But the reporting is consistent enough, across enough outlets, to treat as a real and useful signal about how even the biggest AI labs actually behave versus how they talk in public.

TL;DR

What Hatch reportedly is

Hatch is Meta's internal codename for an "always-on" consumer AI agent, positioned as Meta's answer to OpenClaw — an agent that does things rather than just answers questions. According to The Information and the Financial Times, Meta is targeting integration directly into Instagram and WhatsApp, letting users hand off tasks like comparing products, booking a table, following up on a delivery, or triaging calendar and email before summarizing what needs attention.

To get there, Meta reportedly built mock training environments resembling DoorDash, Etsy, Reddit, Yelp, and Outlook, using them to teach the agent tool use and web navigation. Internal testing was targeted for the end of June 2026, and a public rollout — possibly gated behind a waitlist — has been reported as realistic for late August or September 2026. Meta is also said to be considering a premium subscription tier priced as high as $199.99 a month for higher usage limits, alongside a parallel Instagram shopping-agent effort aimed at Q4 2026.

Here's the detail that makes this story worth a second look: while Hatch is in development, it's reportedly running on Anthropic's Claude Opus 4.6 and Claude Sonnet 4.6 — not Meta's own Muse Spark reasoning models, and not Llama. Claude is described as a "transitional layer" that Meta plans to swap out for Muse Spark before or around public launch — the difference, as one analysis put it, between renting the capability and owning it.

Meta has not confirmed any of this on the record. It has confirmed, separately, that it's building its own coding agent — Muse Code, which Zuckerberg announced in beta on August 5-6, 2026, and which Meta's own published benchmarks show trailing Claude Code on some tasks. That gap is relevant context for why Hatch, a harder and newer agentic problem, would reportedly lean on Claude rather than wait for an in-house model to catch up.

The $10 billion Anthropic spend — and the deal it's easy to confuse it with

According to the New York Times (as summarized by Techmeme and Quartz on August 27, 2026), Meta has internally projected it could spend up to $10 billion annually on Anthropic's AI tools and models — spanning Meta's own developers adopting Claude Code at scale internally, plus Anthropic models powering products like Hatch. The Times reporting frames this as a striking contrast: Zuckerberg was publicly criticizing AI companies he didn't name (widely read as targeting Anthropic) in his August 10 essay, "The Future Is for Everyone," while Meta was quietly one of Anthropic's largest customers.

This is genuinely a different story from the one you may have seen in July. In mid-July 2026, CNBC and others (including Quartz's earlier piece) reported that Anthropic was in early talks to lease up to $10 billion of compute from Meta over two years — chips, power, and data-center capacity to keep Claude running, with Anthropic paying Meta in monthly instalments. That's money flowing the opposite direction, for a different purpose (compute infrastructure, not model access), and it's still described as an early-stage, non-final negotiation with either side able to exit.

Put together, the two reports describe an unusually tangled relationship: Anthropic potentially renting raw compute capacity from Meta's data centers, while Meta potentially pays Anthropic for the finished product — Claude's model access and Claude Code. Meta and Anthropic build directly competing frontier models, Llama/Muse Spark and Claude, and would be each other's customer and vendor at the same time on two separate, overlapping deals. Both companies have declined to comment on the terms of the July compute-lease talks; neither has confirmed the $10B annual figure from the NYT report.

For scale: Anthropic projected its own annualized revenue would exceed $65 billion this year, with Claude Code alone reportedly accounting for roughly a third of that. A $10 billion Meta relationship, if it materializes at that scale, would be a meaningful chunk of Anthropic's business from a single customer that also competes with it.

What this means for enterprise AI strategy

Strip away the personalities and the headline-grabbing "frenemies" framing, and the pattern here is one that shows up constantly for teams choosing between AI vendors, not just at hyperscaler scale:

Even a company with its own frontier lab still buys the model that performs better for a given job. Meta has invested tens of billions of dollars building Meta Superintelligence Labs, Muse Spark, Muse Code, and Muse Glimmer. It still reportedly built its flagship new consumer agent on a competitor's model, because — per the reporting — Claude had a capability edge on this specific agentic task, and closing that gap with an in-house model would have pushed the launch date out by months. Build-vs-buy isn't a one-time decision; it's revisited per product, per task, and it can flip mid-project.

Single-vendor loyalty is rare even inside AI-native companies. Meta's own engineers reportedly adopted Claude Code broadly, in parallel with Meta shipping its own competing Muse Code. If the company building foundation models still lets its own developers reach for a different model when it routes work better, that's a strong signal against assuming any one vendor covers every task well.

"Transitional" dependencies are a real strategy, not just a stopgap. Using a stronger third-party model to ship now, with a defined plan to swap in an owned model later, lets a team hit a deadline without permanently ceding the category. The risk is real too — Meta reportedly restricts employee access to Claude and OpenAI's Codex over concerns about model distillation, per earlier Information reporting, showing the tension between "use the best tool" and "don't teach your rival's tool your playbook."

Compute and model-access relationships can run in both directions simultaneously. Anthropic reportedly leases compute from Meta while Meta reportedly buys model access from Anthropic — the same two companies, two different resource flows, two different negotiations. For any team mapping its own AI vendor relationships, it's worth being explicit about which direction money and dependency actually flow for each deal, rather than assuming one relationship implies the other.

None of this requires taking Zuckerberg's essay or Meta's PR at face value — the practical lesson is narrower and more useful: pick models per task based on what actually performs, keep a credible in-house or alternate-vendor path, and expect that path to take longer than the roadmap says.

Related reading

• Update — September 7, 2026: Pre-launch tests reportedly found Hatch changing account passwords without permission — see Meta's Hatch agent changed passwords without permission.
• Update — September 7, 2026: Anthropic has reportedly lined up roughly $517 billion in total compute commitments and reportedly surpassed OpenAI in revenue — see Anthropic's reported $517B compute commitments, explained.
• Zuckerberg's "The Future Is for Everyone": Meta's Open-Source Pledge
• Meta Muse Code: Terminal Coding Agent Powered by Muse Spark 1.2
• Muse Spark 1.1: Meta Model API, 1M Context, and Agentic Coding Upgrade
• Is OpenClaw Safe? Anthropic's Ban and Peter Steinberger's Response
• Cursor Router: Automatic Model Selection for Teams and Enterprise
• Meta's 73 Trillion Tokens: Spotify, Shopify, and AI Engineering at Meta
• Dario Amodei and Gavin Baker Debate AI Regulation

Official sourcing on the underlying reports: New York Times reporting summarized via Techmeme, Quartz on the $10B Anthropic tools spend, and Quartz on the earlier Anthropic-leases-Meta-compute talks.

This post reflects reporting available as of August 28, 2026. Hatch's launch date, pricing, model backend, and the exact terms of any Meta-Anthropic spending arrangement are based on unconfirmed, anonymous-sourced reporting and may change — neither Meta nor Anthropic has made an on-the-record statement confirming these details as of publication.