ChatGPT Just Became A Work Agent

July 10, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
July 10, 2026

HEADLINES

Cursor Is Building Its Own Cowork Competitor

Cursor is reportedly developing a general-purpose agent called Sand to go after the knowledge-work market that Claude Cowork has staked out. Work began in April, shortly after Cursor's deal with SpaceX, and the agent — which runs on Grok 4.5 — is designed for everyday office tasks like email and spreadsheets rather than just coding. It's already live internally as of June, though it's still unclear if or when it gets a public release.

OpenAI Says the Top Coding Benchmark Is Broken

OpenAI audited SWE-bench Pro, the leading coding benchmark, and found that 30% of its tasks were broken — contaminated by public problems leaking into training data, hidden requirements, contradictory instructions, and incomplete grading criteria. They're formally retracting their support for it, joining a broader move away from the benchmark that Cursor, Cognition, and Databricks were already ahead of with their own proprietary versions. Expect a lot more benchmark fragmentation from here, with SWE-bench Pro still getting cited anyway while everyone quietly goes with their vibes.

OpenAI Draws Its National Security Red Lines

OpenAI published a new statement on government and military partnerships, pledging not to support mass domestic surveillance, high-stakes use-of-force decisions without human accountability, or uses that evade legal oversight. The principles land awfully close to Anthropic's own red lines — and given how much friction those created with the government, it's not obvious what problem this document solves. Still, it's a clear marker for OpenAI to build from in future conversations with Washington.

Bernanke Joins Anthropic's Long-Term Benefit Trust

Anthropic has appointed former Fed Chair Ben Bernanke to its Long-Term Benefit Trust, the independent oversight board that sits above the corporate board and is on track for majority board-appointment power next year. Bernanke chaired the Fed through the 2008 crisis and the QE era that followed, and remains a polarizing pick depending on whether you credit him with averting a depression or blame him for widening the wealth divide. Asked for a dispassionate read, Claude described him as generally well-regarded but viewed by critics on both the left and right as "emblematic of an unaccountable technocracy protecting elite interests." FinTwit had its fun with the pick too, with Tom Bruni asking the obvious question: "Will he bring token prices to zero like he did with rates?"

Meta Breaks Ground on $10B Canada Data Center

Meta is heading north with a $10 billion, one-gigawatt data center in Alberta, breaking ground this week. The company is pledging $60 million CAD for local infrastructure like roads and water, covering the full cost of electricity infrastructure (with officials expecting local rates to actually drop as a result), and funding local nonprofits directly, on top of 3,000 construction jobs and 300 permanent ones. It's a good reminder that even as Meta talks about selling off excess compute, the build-out shows zero signs of slowing down.

Meta's In-House Chips Head to Production in September

Meta's long-troubled custom chip program is back from the dead, with production set to begin in September after at least one design cleared testing in just six weeks. Meta is working with Broadcom on design, TSMC for processors, and Samsung for memory, and plans to ship a new chip every six months starting next year — well ahead of the standard industry cadence. The company also reaffirmed plans to deploy 7 gigawatts of capacity this year and double that pace in 2027.

MAIN STORY

A New AI Harness for Work

This week was always going to be about new models — GPT-5.6's full family finally shipped, Grok 4.5 rattled the leaderboard, and Meta's Muse Spark 1.1 came out of nowhere. But the real story of 2026 isn't just which model is smartest, it's the harness wrapped around it, and OpenAI closed the week by putting that idea front and center with ChatGPT Work.

NEW MODELS AND HARNESSES KEEP COMING

GPT-5.6: Sol, Terra, and Luna
Two frontier models, two very different jobs
OpenAI's first three-tier model family is officially out, benchmarks and all — presented less as leaderboard tables and more as cost-per-performance charts. Sol on max settings lands a single point behind Fable 5 on the Artificial Analysis Index at a third of the cost, and actually beats Fable outright on the Coding Agent Index. Terra matches Fable's coding performance for a lot less, and Luna reportedly matches GLM-5.2 at 43% cheaper, per Simon Smith.

Fable vs. Sol
Fable does it for you, Sol does it with you
Early testers are converging on the same split: Fable 5 is the model you hand a massive, autonomous task and walk away from, while Sol is fast enough to stay in the loop with you. Every CEO Dan Shipper called Sol his default for almost everything and said it's the first model he's trusted to run whole loops of knowledge work, while Theo said he burned over $200K in tokens building with it in a month. Not everyone's sold on Fable regardless of capability — one enterprise dev told Gergely Orosz they're avoiding it entirely because Anthropic hasn't changed its data retention policy.

ChatGPT Work
OpenAI's answer to Claude Cowork, and then some
Alongside the model family, OpenAI shipped ChatGPT Work, a new harness that takes Codex's agentic approach and extends it to knowledge work generally — connectors for Notion, Google Drive, and Microsoft 365, cloud execution that keeps running after you close your laptop, and enterprise-grade access controls. OpenAI says its own sales and finance teams are already using it to compress week-long processes into hours, and Zapier's head of enterprise credited it with surfacing seven figures in missed pipeline.

The Naming Backlash
So developers aren't working?
Consumer reaction to the ChatGPT Work / ChatGPT Codex split was more muted than OpenAI probably hoped. Peter Yang argued the whole thing should just be called ChatGPT or Codex with no tabs, Ethan Mollick said he still doesn't understand what he's gaining or giving up versus Codex, and even Dan Shipper admitted he was "extremely worried" about the merge before landing on "the merged app is fine."

Muse Spark 1.1
Zuck's back on Twitter, and Meta's back in the race
Meta capped the week by having Mark Zuckerberg tweet for the first time in three years to announce Muse Spark 1.1, a model that's competitive with Opus 4.8 and GPT-5.5 on benchmarks and state-of-the-art on agentic tasks like MCP Atlas and JobBench. The real headline is cost and speed — Vals AI's Rayyan called it a quarter the latency of Opus 4.8 and a tenth the cost of Fable or GPT-5.5, and the Julius team used it to build a working Minecraft clone in about five minutes for 73 cents in tokens.

The Race Reshuffled
Every lab is now competing on cost, not just IQ
Every model released this week — Grok 4.5, Swee 1.7 from Cognition, Muse Spark 1.1, even GPT-5.6 itself — leaned hard on cost and efficiency in its pitch, not just raw capability. A week ago, SpaceX AI and Meta weren't really part of the enterprise model-selection conversation; they're ending it very much back in the room, and that's a good thing for everyone shopping for models.