# ChatGPT Just Became a Work Agent
*The AI Daily Brief — Friday, 2026-07-10 · https://aidailybrief.ai/e/2026-07-10*

**The coding playbook is now the knowledge-work playbook — and the race has moved to cost.**

Everything that worked in coding — agents, loops, goals, a harness that runs whole tasks — is being generalized to all knowledge work. OpenAI's ChatGPT Work is the clearest expression: define a goal, load context, let the agent do the loop. But the bigger structural shift is that every model shipped this week — GPT-5.6, Grok 4.5, Muse Spark 1.1, Cognition's Swe 1.7 — led its pitch with cost and efficiency, not frontier scores. The labs now openly compete on dollars-per-task, and that puts Meta and SpaceX AI back in the enterprise conversation they'd fallen out of.

---

## By the numbers
- **30%** — Of SWE-bench Pro tasks OpenAI found broken — and formally retracted support
- **$10B** — Meta's new Alberta data center, targeting 1GW of capacity
- **7GW** — Capacity Meta plans to deploy in 2026 — doubling the pace in 2027
- **6 weeks** — How fast one of Meta's in-house chip designs passed testing
- **$200K** — Tokens Theo burned building with GPT-5.6 Sol in a month
- **40%** — How much cheaper GPT-5.6 ran than Opus 4.8 on the analysis index
- **1/10** — Muse Spark 1.1's cost vs. Fable and GPT-5.5 in Vals testing
- **$0.73** — Cost for Muse Spark to build a Minecraft clone in ~5 minutes

## Headlines

### Cursor is building a general-purpose work agent called Sand `[00:00]`
Per The Information, Cursor began work in April, shortly after its SpaceX deal, on a Grok 4.5-powered agent aimed at anyone beyond coders. Codenamed Sand, it handles standard office tasks like email and spreadsheets, was rolled out internally in June, and could eventually unify with AI coding into one platform.
*For: Product, Ops*
Link: https://aidailybrief.ai/e/2026-07-10#cursor-sand-agent

### OpenAI declares the leading coding benchmark bunk `[01:20]`
OpenAI audited SWE-bench Pro, found 30% of tasks broken — public problems leaking into training data, hidden requirements, contradictory instructions, overly strict tests — and formally retracted support, saying it 'no longer reliably measures frontier coding capability.' Cursor, Cognition, and Databricks have all launched their own benchmarks.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-07-10#swe-bench-pro-bunk

### We're now at peak benchmark proliferation `[01:50]`
NLW predicts everyone will keep presenting all the benchmarks — including SWE-bench Pro — and users will just wait to see whether the vibes confirm or debunk each one.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-07-10#benchmark-vibes

### OpenAI publishes its own red lines `[02:30]`
OpenAI's new national security principles say it won't support mass domestic surveillance, high-stakes decisions including use of force without human judgment, or uses that evade legal oversight. NLW notes these are essentially Anthropic's red lines — which created chaos with the government — leaving the document's purpose unclear.
*For: Legal, Exec*
Link: https://aidailybrief.ai/e/2026-07-10#openai-national-security-principles

### Anthropic puts Ben Bernanke on its long-term benefit trust `[03:30]`
The former Fed chair joins the independent trust that sits above Anthropic's corporate board, can elect or remove board members, and will hold majority control by next year (with a shareholder supermajority override). No trust member may be a shareholder.
*For: Exec, Legal*
Link: https://aidailybrief.ai/e/2026-07-10#bernanke-anthropic-trust

### Meta's $10B Alberta data center comes with real community money `[05:00]`
Meta broke ground on a $10B Canadian site targeting 1GW, pledging C$60M for local roads and water, full infrastructure costs (with rates expected to drop), nonprofit funding, 3,000 peak construction jobs, and 300 ongoing roles.
*For: Ops*
Link: https://aidailybrief.ai/e/2026-07-10#meta-alberta-datacenter

### Community contributions are the easiest win hyperscalers keep missing `[05:30]`
NLW argues these local investments are a rounding error in data-center budgets but hugely valuable to communities — one of the easiest possible alignments between hyperscalers and the places they operate, if only they'd actually do it.
*For: Ops, Exec*
Link: https://aidailybrief.ai/e/2026-07-10#community-alignment-take

### Meta's in-house chip program is back from the dead `[06:30]`
Per a Reuters-cited memo, Meta will begin producing its first chips in September, with one design clearing testing in just six weeks. Built with Broadcom, made at TSMC with Samsung memory, Meta plans a new chip every six months from next year — and reaffirmed deploying 7GW in 2026, doubling in 2027.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-07-10#meta-chips-back

## Main episode

### GPT-5.6 splits into a three-model family `[08:00]`
OpenAI's first tiered lineup: flagship Sol, mid-size Terra, and small cost-efficient Luna. All are now fully released and publicly available, with official benchmarks landing after a week of sanctioned early-tester impressions.
*For: Eng, Exec*
Link: https://aidailybrief.ai/e/2026-07-10#gpt56-family

### OpenAI reframes benchmarks around cost, not just score `[09:30]`
Instead of a simple table of numbers, OpenAI now leads with performance-per-cost charts — score against API cost, latency, and output tokens. The emphasis: 5.6 not only performs better but does so far more cheaply.
*For: Finance, Eng*
Link: https://aidailybrief.ai/e/2026-07-10#benchmarks-cost-charts

### 5.6 Sol is a huge step forward for dollars per task, as are Terra and Luna. `[10:15]`
*— Sam Altman, OpenAI*
Altman explicitly tied the launch to enterprise cost concerns, framing the whole family around efficiency rather than raw capability.
*For: Finance, Exec*
Link: https://aidailybrief.ai/e/2026-07-10#altman-dollars-per-task

### 5.6 nearly matches Fable 5 at a third of the cost `[10:45]`
On the artificial analysis index, GPT-5.6 finished a single point behind Fable 5 but at one-third the cost and 40% cheaper than Opus 4.8. On the coding agent index it's the new state-of-the-art, and mid-size Terra matched Fable-level coding at far lower cost.
*For: Eng, Finance*
Link: https://aidailybrief.ai/e/2026-07-10#gpt56-cost-performance

### Luna matches an open-weight model and undercuts it `[11:30]`
Simon Smith noted 5.6 Luna matches GLM 5.2 on the intelligence index at 43% cheaper — evidence, he argues, that frontier labs optimizing for both intelligence and efficiency negate the need to switch to open-weight models purely to save money.
*For: Eng, Finance*
Link: https://aidailybrief.ai/e/2026-07-10#luna-vs-glm

### Two frontier models that behave very differently `[12:00]`
Early consensus: Fable 5 is the big, slow, autonomous model for massive long-running tasks, while GPT-5.6 Sol is a fast, cheaper daily driver for tasks where you want to stay involved in intermediate decisions.
*For: Eng, Product*
Link: https://aidailybrief.ai/e/2026-07-10#two-different-frontier-models

### Sol is the first model I've trusted to run whole loops of knowledge work, not just help with individual tasks. `[13:15]`
*— Dan Shipper, Every*
Dan Shipper said Sol 'shifted my job from doing the work to tending the system that does it,' called it half Fable's price and his default for almost everything, and even preferred it for writing as clearer and more concise than Anthropic models.
*For: Exec, Ops*
Link: https://aidailybrief.ai/e/2026-07-10#shipper-knowledge-work-loops

### Theo burned $200K in tokens building with 5.6 Sol `[13:40]`
In a portfolio-style review, Theo described interacting with 5.6 as a collaborator you work alongside — reinforcing the split from Fable, which you let run off on its own.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-07-10#theo-200k-tokens

### Anthropic has not changed their data retention policy on Fable... we're going hard on GPT 5.6 Sol as a result. `[15:00]`
*— A developer, via Gergely Orosz*
Gergely Orosz relayed a dev at a large, AI-bullish company saying data-retention concerns on Fable — that Anthropic would store their data — pushed them to standardize on GPT-5.6 Sol.
*For: Legal, Eng*
Link: https://aidailybrief.ai/e/2026-07-10#data-retention-switch

### ChatGPT Work extends the Codex harness to all knowledge work `[16:00]`
OpenAI's answer to Claude CoWork: an agent that acts across your apps and files, stays on a project for hours, and turns a goal into finished work. It has connectors for Notion, Google Drive, and Microsoft 365, scheduled tasks, cloud execution, and enterprise security controls.
*For: Ops, Product, Exec*
Link: https://aidailybrief.ai/e/2026-07-10#chatgpt-work-harness

### ...generated a weekly executive dashboard that revealed seven figures in potential sales. `[16:45]`
*— Angela Ferrante, head of enterprise at Zapier*
Zapier's head of enterprise said ChatGPT Work built a repeatable system to review thousands of leads monthly, tracing touchpoints across CRM and email to find where follow-ups broke down.
*For: Sales, Ops*
Link: https://aidailybrief.ai/e/2026-07-10#zapier-seven-figures

### OpenAI is running its own teams on ChatGPT Work `[17:30]`
Sales used it to turn a discovery call into a tailored proof of concept in 24 hours (normally weeks); finance cut month-end close and forecasting from days to hours by finding source data, moving it into Excel or Sheets, reconciling, and building slides.
*For: Sales, Finance*
Link: https://aidailybrief.ai/e/2026-07-10#openai-internal-usage

### I think the ChatGPT Work versus Codex thing is confusing. `[18:30]`
*— Peter Yang*
Peter Yang argued it should all just be called Codex with no tabs or toggles. Ethan Mollick echoed the confusion, and Dan Shipper's takeaway was that the merged app is 'fine' — not exactly the reaction OpenAI wants from a big launch.
*For: Product*
Link: https://aidailybrief.ai/e/2026-07-10#work-vs-codex-confusion

### Updated Sites turns knowledge work into shareable web apps `[20:00]`
The Sites feature lets you turn any output into a website or web app shareable across your company, even with non-ChatGPT users. NLW argues building websites instead of traditional artifacts, with better hosting, will meaningfully change how people output work.
*For: Ops, Marketing*
Link: https://aidailybrief.ai/e/2026-07-10#sites-feature

### Meta shocks with Muse Spark 1.1 `[21:00]`
Zuckerberg tweeted for the first time in three years to announce Muse Spark 1.1, competitive with Opus 4.8 and GPT-5.5. It beats both on Humanity's Last Exam and leads on personal agentic tasks — state-of-the-art on MCP Atlas and ahead on Job Bench — Meta's biggest LLM leap since Llama 3.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-07-10#muse-spark-meta

### The model is so cheap I almost don't believe it... one-tenth the cost of both Fable and GPT-5.5. `[23:30]`
*— Rayyan, Vals AI*
Vals' Rayyan flagged Muse Spark 1.1 at one-quarter Opus's latency and dramatically cheaper — 92 cents to test on Vibe Code Bench vs. $5.09 for Opus and $12.51 for Fable. It built a Minecraft clone inside Julius in five minutes for 73 cents.
*For: Finance, Eng*
Link: https://aidailybrief.ai/e/2026-07-10#muse-spark-cost

### The labs now openly compete on efficiency, not just frontier scores `[24:20]`
NLW argues every model this week — Grok 4.5, Cognition's Swe 1.7, Muse Spark 1.1, even GPT-5.6 — led on cost and efficiency. SpaceX AI and Meta, out of the enterprise conversation days ago, are ending the week firmly back in it.
*For: Exec, Finance*
Link: https://aidailybrief.ai/e/2026-07-10#cost-is-the-new-vector

### Meta is the only hyperscaler on track to be world-class at data, talent, and compute. `[25:00]`
*— SemiAnalysis*
SemiAnalysis argued Meta has the best chance of catching Anthropic and OpenAI, holding all three ingredients of a true frontier model.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-07-10#meta-three-pillars

---
Transcript: https://aidailybrief.ai/e/2026-07-10/transcript.md
Listen: https://pod.link/1680633614 · Ad-free: https://patreon.com/aidailybrief
© 2026 The AI Daily Brief — Until next time, peace ✌