In partnership with
In Today’s Issue:
☀️ GPT-5.6 Sol tops the evals and folds Codex and Work into a ChatGPT superapp
🧠 Meta's Muse Spark 1.1 blindsides everyone on the agent benchmarks
🚀 SpaceX and Cursor ship Grok 4.5 at a fraction of the cost
⚖️ Europe votes to legalize scanning half a billion people's chats
✨ And more AI goodness…
⚡ The Signal
This was the week the AI race stopped being a two-horse contest.
For two years it was OpenAI versus Anthropic, with everyone else a lap behind. Then, inside a few days, SpaceX-backed Grok 4.5 landed at GPT-5.5 and Opus 4.8 quality for a fraction of the cost, and Meta's Muse Spark 1.1 blindsided everyone by topping the agent benchmarks. Today OpenAI answered with GPT-5.6, folding Codex and ChatGPT Work into a single superapp. Four labs now trade blows at once, and the fight has moved from who is smartest to who is smartest per dollar. Anthropic still tops the independent rankings, but the pack chasing it has doubled overnight. And while the US and China sprint toward AGI, Europe spent its week legislating mass surveillance. A week of dizzying highs and one grim low.
All the best,
Kim Isenberg
(Meta)
🧠 Meta's Muse Spark 1.1 Blindsides the Agent Benchmarks
Meta Superintelligence Labs dropped Muse Spark 1.1 with almost no warning, and it promptly beat GPT-5.5 and Claude Opus 4.8 on the benchmarks that measure agents doing real work. The multimodal model tops MCP Atlas, JobBench, Humanity's Last Exam, and Meta's Finance Agent v2 for tool use, even as it trails those same rivals on pure coding tests like SWE-Bench Pro. Meta opened it to developers through the new Meta Model API, its first serious bid to sell frontier intelligence rather than give it away.
👉 tl;dr: Meta came out of nowhere with an agent model that outscores GPT-5.5 and Opus 4.8 on tool use. That is four serious frontier labs, up from two.
(SpaceX / Cursor)
🛠️ Cursor Builds "Sand" to Take On Claude Cowork
Cursor is quietly building a general-purpose AI agent, internally called Sand, to compete with Anthropic's Claude Cowork, The Information reports. It would be the coding startup's first product aimed at everyday users rather than developers, handling email, texts, and spreadsheets alongside engineering work. The timing is not subtle: Cursor is leasing compute from SpaceX's AI unit ahead of SpaceX's planned $60 billion acquisition, and the two just shipped Grok 4.5 together.
👉 tl;dr: Cursor wants to be more than a coding tool, and SpaceX's money is what makes the pivot possible.
(xAI / Cursor)
🚀 SpaceX and Cursor Ship Grok 4.5 at Half the Cost
SpaceXAI and Cursor released Grok 4.5, a 1.5-trillion-parameter model trained on real Cursor coding sessions that lands within striking distance of the frontier for a fraction of the price. It trades wins with GPT-5.5 and Claude Opus 4.8 across coding benchmarks like Terminal-Bench and SWE-Bench Pro while costing about 60% less to run, at roughly $2 per million input tokens. Elon Musk called it an "Opus-class model," and on price alone, that is hard to argue with.
👉 tl;dr: Near-frontier coding at about 60% off is the whole pitch, and it is working.
Stop trusting the benchmark charts and run your own.
With four labs now crowding the frontier at wildly different prices, the benchmark that matters is your own work, not a public leaderboard. Pick one real task you did this week and run it through GPT-5.6, Grok 4.5, and Muse Spark side by side. Why it helps: public evals measure average tasks; you only care about yours. A five-minute bake-off tells you which model to actually pay for better than any chart does. Try this: "Here is a task I did this week: [paste the task and the inputs]. Do it end to end. Then rate your own output 1 to 10 on accuracy, usefulness, and how much editing I would still need, and justify the score." Run it in each model and compare the self-scores against your own judgment.
🎬 Watch This
This is OpenAI's own launch film for ChatGPT Work, the superapp half of today's release. Powered by Codex and GPT-5.6, it shows ChatGPT stepping out of the chat box and into the job itself, reading your company's context, running multi-step tasks, and writing and executing code in the same place you draft an email. It is the clearest look yet at where OpenAI thinks the assistant is going: less a place you ask questions, more a colleague you hand projects to.
"I was clearly wrong about Anthropic. They are obviously currently the leader in AI. No company has released a model as good as Mythos/Fable."
– Elon Musk, CEO of SpaceX and xAI
A year ago, Musk said winning "was never in the set of possible outcomes" for Anthropic. This week he crowned them, and the timing is the tell: his own SpaceX-backed Grok 4.5 had just shipped as a budget play rather than a frontier-topper, while Anthropic's Mythos and Fable models sit on top of the independent rankings. The sharper point is leverage. Anthropic rents roughly $1.25 billion a month of SpaceX compute, enough for Musk to choke a rival at will, and he went out of his way to swear he never would. When the industry's most combative founder concedes the lead and promises not to pull the plug, the concession itself is the news.
The line in OpenAI's GPT-5.6 material that set X alight was not a benchmark. It was the claim that GPT-5.6 Sol autonomously post-trained its smaller sibling, GPT-5.6 Luna. Read literally, that is one model improving another with limited human hands on the wheel, the first whiff of recursive self-improvement in a shipping product. Researchers who dug in quickly cooled the hype: the likeliest reading is that Sol ran a bounded, well-scoped slice of the post-training pipeline, editing a config, adjusting a scheduler, and launching a run inside OpenAI's mature infrastructure, not inventing Luna from scratch. Still, a model that trains another model, even a sliver of the job, is a claim that would have read as science fiction a year ago, and OpenAI chose to put it in writing.
GPT-5.6 Arrives, and the Chatbot Becomes a Coworker
The Takeaway
👉 OpenAI released GPT-5.6 in three tiers, Sol (top intelligence), Terra (balanced), and Luna (cheapest), at $5, $2.50, and $1 per million input tokens.
👉 GPT-5.6 leads the frontier on Terminal-Bench 2.1 (91.9%), BrowseComp (90.4%), and Agents' Last Exam (53.6%), ahead of Claude Mythos 5 and Gemini 3.1 Pro.
👉 Codex and ChatGPT Work are folded into one ChatGPT superapp that can plan a task, then write and run the code to finish it.
👉 A native ChatGPT desktop app and a million-token context window on the API push GPT-5.6 from answer engine toward autonomous coworker.
For two years, using ChatGPT meant typing a question and waiting for an answer. Today OpenAI tried to end that arrangement. GPT-5.6, released across ChatGPT, the API, and a new desktop app, is less a smarter chatbot than a bid to make the assistant do the work itself. It ships in three tiers OpenAI now treats as durable products rather than sizes: Sol for maximum intelligence, Terra for balance, and Luna for speed and low cost. The number marks the generation; the name marks the tier.
(Artificial Analysis)
On OpenAI's own benchmarks, GPT-5.6 takes the top spot. Its best configuration tops Terminal-Bench 2.1, a test of an agent working in a real command line, at 91.9%, ahead of Anthropic's Claude Mythos 5 (88.0%) and Gemini 3.1 Pro (70.7%). It leads again on BrowseComp (90.4%), which measures live web research, and stretches the gap on Agents' Last Exam (53.6% versus Mythos 5's 40.5%), a brutal test of long, multi-step autonomy. These benchmarks reward a model for finishing real jobs, and that is the point.
(OpenAI)
The launch OpenAI half-buried under the benchmarks is the superapp. Codex, its coding agent, and ChatGPT Work, its enterprise workspace, are merging into a single ChatGPT that reads your context, plans a task, then writes and runs the code to do it, all in one window. Paired with a native desktop app and, for developers, a million-token context window on the API, it reframes the product: you are no longer prompting a chatbot, you are handing a brief to something that behaves like a junior colleague.
The caution is the one every agent launch earns. A score in the low 50s on Agents' Last Exam means the hardest long tasks still fail more often than they land, and a coworker you cannot fully trust still needs a manager. But the direction is unmistakable, and OpenAI is no longer shipping it alone.
Why it matters: The frontier is no longer only about raw intelligence; it is about who turns intelligence into finished work at a price businesses will pay. GPT-5.6 plus the superapp is OpenAI's answer, and it arrived the same week two rivals proved the answer is now contested.
Name A Better GTM Resource
GTM Atlas by Attio is back with more frameworks, systems, and insights for early teams navigating growth from scratch, straight from the people shaping modern GTM.
Read six new entries from operators at Anthropic, Notion, Stripe, Linear, Granola, and Wispr Flow.
Mapped by operators. Curated by Attio.
The chart: Artificial Analysis measured the cost to finish one task on its Intelligence Index. GPT-5.6 Sol runs about $1.04, against $1.80 for Claude Opus 4.8 and $2.75 for Claude Fable 5, the one model that narrowly outscores it. The cheaper GPT-5.6 tiers go lower still: Terra at $0.55, Luna at $0.21.
The lesson: The frontier has bunched up on quality, so the real fight is now price. Sol delivers near-top intelligence at roughly a third of Fable 5's cost, and Grok 4.5 and the Luna tier drag the floor down again. This is what "smartest per dollar" looks like on a chart.
The caveat: Cost per task climbs with how much a model reasons, and heavy reasoners burn tokens fast. These are Artificial Analysis's blended estimates, not your invoice; a cheaper model that needs more retries can still cost more in practice.
⚖️ Europe Just Legalized Reading Everyone's Messages
⚡ Bottom line: The EU Parliament let Chat Control 1.0 pass, allowing suspicionless scanning of private messages on major US platforms until 2028.
💡 Why it matters: A majority of the MEPs who voted actually opposed it; it advanced anyway on a procedural technicality rushed through before the summer recess.
🔎 What it means: While the US and China race for AGI, Europe's defining AI-era move is to legalize mass surveillance of half a billion people.
On July 9, the European Parliament did something quietly remarkable: it advanced a law that most of its own voting members opposed. The motion to reject Chat Control 1.0 needed an absolute majority of 361 votes. It drew 314, against 276 in favor with 17 abstentions, so more MEPs voted to kill the measure than to keep it, and it survived anyway. The rule lets platforms like Instagram, Discord, Snapchat, Gmail, and iCloud keep scanning private messages for suspected illegal content, with no suspicion and no warrant, until 2028.
What makes it worse is the how. The same text had been rejected twice in March. This time it was forced through an expedited procedure, with no real urgency, timed for the days when most of Parliament had already left for summer break. Patrick Breyer, the former MEP who has fought the law for years, did not soften it: "The fact that Chat Control is moving forward against the will of the majority of voting MEPs is a farce and damages democracy." He compared suspicionless scanning to "frantically mopping the floor while the faucet is still running."
(Patrick Breyer)
There is an AI subtext Brussels prefers not to say out loud. Scanning messages at continental scale is itself an AI system, a wall of classifiers making automated guesses about hundreds of millions of private conversations, with all the false positives that implies. So while OpenAI, Anthropic, SpaceX, and Meta spent the week racing toward more capable models, Europe's most consequential AI deployment of the week is aimed at its own citizens. Permanent rules get negotiated in September. This was supposed to be the version everyone could live with.
See the whole platform. No guided tour.
Skip the sales call. Walk through Gladly's interface yourself — the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.