- Cognition's SWE-2 gets within a point of Anthropic's Fable 5.1 on Cognition's own coding benchmark. The base model underneath it belongs to Moonshot.
- NASA and IBM put a lunar science model on Hugging Face under Apache-2.0, trained on roughly 2 million image tiles of the Moon.
- OpenAI opened the Codex harness to everyone through a new Agents API in public beta, at no fee beyond tokens.
- A fresh paper finds agent "skills" get brittle on long jobs while subagents hold up better, and explains exactly when that flips.
- The scarce resource of the day turned out to be electricity. If you'd rather build than watch, the unglamorous first move is still writing the spec.
Line up the day's three biggest moves and they look unrelated. A startup landed within a point of a frontier coding model without training one. A space agency gave its lunar model away. The UN spent the afternoon warning that the binding constraint on AI isn't intelligence, it's megawatts.
They're the same story. The model is turning into the cheap part. What surrounds it, the base weights somebody else paid for, the context you feed it, the power it burns, is where the cost quietly moved.
The Front Page: Cognition Built a Near-Frontier Coding Agent on Somebody Else's Model
Cognition shipped SWE-2 on September 10, its most capable coding agent so far. The company laid out the numbers in its own announcement: 50.0% on FrontierCode 1.1 Main, against 50.9% for Anthropic's Fable 5.1. Cognition also says SWE-2 medium matches its previous model's score while taking 58% fewer turns and costing 81% less, and that it makes its first real code edit after a median of 18 steps instead of 48. Coverage of the launch ran here.
Now the part the headline number buries. SWE-2 is post-trained from Kimi K3, Moonshot's open-weight 2.8-trillion-parameter model. Post-training means taking a base model somebody else already built and training it further for one job, in this case agentic coding. Cognition didn't pay for the base. It paid for the specialization.
Two caveats before anyone reorganizes a roadmap around a 0.9-point gap. FrontierCode is Cognition's own benchmark, measuring whether an AI-written pull request would survive a human maintainer's review. Useful thing to measure, and also a leaderboard where the host writes the questions. The cost comparisons are against Cognition's older model, not Anthropic's pricing.
What it means: The moat is sliding. Training a 2.8-trillion-parameter base is a capital expense roughly a dozen organizations on earth can carry. Post-training one into a specialist costs a different order of magnitude, and the best base for the job was open-weight and Chinese. If that arbitrage holds, the competition stops being "who has the smartest model" and becomes "who wraps the best harness around weights anyone can download." Much larger field of players.
Releases & Features
NASA and IBM open-sourced a Moon model. The NASA-IBM Lunar Foundation Model went up on Hugging Face under an Apache-2.0 license, trained on roughly 2 million image tiles from the Lunar Reconnaissance Orbiter. NASA says it beats widely used methods by as much as 23% at spotting ice deposits, craters, and the volcanic features called irregular mare patches. It shipped with SomBench, a matched evaluation dataset. NASA's write-up, and a second read.
OpenAI put the Codex harness behind one API call. The new Agents API entered public beta the same day, with no fee beyond the tokens and tools an agent consumes. A harness is the plumbing around a model: it keeps the session alive, compacts context when the window fills, retries after failures, and splits work across subagents. OpenAI now hosts that layer for you. The announcement.
What it means: Two opposite giveaways pointing the same way. NASA released the weights because the value lives in what researchers do with lunar data. OpenAI is renting out the harness because it thinks the value moved into the harness. Both bet the model is becoming a commodity input, and a small team can pick up either gift tomorrow morning.
In the Lab
A paper posted this week takes on something anyone running long agent jobs has felt in their teeth: Subagents vs Agent Skills. An agent skill is a reusable package of instructions and scripts loaded into the agent's context window so it can repeat a familiar job. A subagent is the alternative: spawn a separate agent with a clean window for that subtask, and hand back only the result.
The researchers find skills grow brittle as task horizons stretch, because reasoning quality drops as the context window fills with accumulated information. Subagents degrade more gracefully, since no single window has to hold everything. The limitation is in the paper too: subagents only win when the skill encodes procedural knowledge with a clear input and output, runnable without the main agent's history.
What it means: If your agent works beautifully for ten minutes and falls apart at ninety, the fix may be architectural rather than a bigger model. Split the job into pieces with clean handoffs. That's the discipline behind every decent agent workflow: define each step's input and output before you automate it, and long jobs stop collapsing under their own context.
The Oversight Desk
The UN Economic Commission for Europe warned that data-intensive technology, AI data centers above all, is being built faster than the electricity systems feeding it. The commission points to IEA projections of data center consumption climbing from 485 terawatt-hours in 2025 to roughly 950 by 2030, about 3% of global demand. The timing mismatch should worry you more: a data center gets built and connected in two to five years, while the transmission lines it depends on routinely take over ten. The UNECE release, and UN News on it.
What it means: No benchmark measures this, which is why it keeps getting waved through. The company that builds a data center is rarely the one paying to reinforce the lines around it. Somebody pays, usually everyone else on the same grid. Ask where the power for a shiny new AI region is coming from, and notice how often the answer is a press release rather than a signed interconnection agreement.
Today's research says long jobs break because one context window has to hold too much. The fix starts with cutting the job into pieces. Tell BYOBot what you're trying to automate and get that breakdown back as a spec.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- A perfect-10 flaw in Google's agent toolkit. CVE-2026-79696 hits Google Cloud's Agent Development Kit for Python 2.0.0 through 2.6.0 wherever pytest is installed, scoring CVSS 10.0 and letting an unauthenticated remote attacker run code via a crafted test session replay. If you ship on ADK, patch before you finish this list. Advisory.
- Microsoft wants to triple its footprint. Bloomberg reports plans to grow data center capacity from about 12 gigawatts today to more than 38 by 2032, AI-specific capacity rising from 2 gigawatts to roughly a third of that. The finished build would exceed New York State's peak electricity draw. Bloomberg.
- Every possible human DNA typo, precomputed. Google DeepMind published AlphaGenome Atlas, predicted molecular effects for all 9 billion single-letter variants in the human genome, roughly a petabyte, open for non-commercial research. The authors are clear it's evidence, not diagnosis. DeepMind.
- A European deadline lands Tuesday. Providers of general-purpose models trained above the 10^25 FLOPs threshold owe the EU AI Office their first systemic risk evaluations by September 15, covering red-teaming method, energy disclosures, and the copyright training summary. Commission.
- Adobe's AI line item, in proportion. Q3 revenue rose 13% to $6.76 billion, and AI-first annual recurring revenue grew over 150% to pass $650 million. Real growth on a number still under 3% of Adobe's $27.5 billion total ARR, and the stock slipped on soft Q4 guidance regardless. Results.
The Bottom Line
Today's lesson repeats in three keys: the thing everyone competes over is becoming the thing nobody can charge much for. Base weights are going open, harnesses are going rented, and the genuinely scarce assets turned out to be transmission lines and a decade of planning permission. Watch whether anyone independent reproduces Cognition's FrontierCode gap, and watch the interconnection queues more than the leaderboards. Meanwhile a lunar science model is sitting on Hugging Face for free, the Codex harness costs tokens, and the best thing you can do with a commoditized model is point it at your own boring Tuesday.
Frequently Asked Questions
-
Post-training is taking a base model somebody has already trained and training it further for a specific job. Cognition's SWE-2 is post-trained from Moonshot's Kimi K3, an open-weight model with 2.8 trillion parameters. The expensive part, building the base, was already paid for by someone else, which is why a startup can land near frontier performance on coding tasks without a frontier training budget. The specialization on top is where the craft lives, and it's the same craft behind designing a multi-tasking agent.
-
It depends on how long the task runs. Agent skills load reusable instructions into one shared context window, which gets brittle as that window fills up. Subagents spawn a fresh context per subtask and degrade more gracefully. The catch is that subagents only work when the subtask has a clear input and a clear output and can run without the main agent's history. If you're choosing where that work executes, the tradeoffs between runtimes are laid out in our guide to hosted AI agents.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
