Claude Sonnet 5 lands everywhere at once

Anthropic shipped Claude Sonnet 5 today, and within hours it was live in AWS Bedrock, Cursor, Devin, Augment’s Cosmos, and GitLab Duo. On Cursor’s own CursorBench it posts 57% against Sonnet 4.6’s 49%. Cognition’s FrontierCode Extended, the benchmark that grades whether an agent’s changes actually merge, puts Sonnet 5 at 53.8% with a 57.6% pass rate, higher than Opus 4.8 on that harness. The pitch is the one Sonnet has always made, now sharper: near-Opus quality at Sonnet cost. Augment is running an introductory $2 per million input tokens and $10 per million output through August 31, and Cognition says Sonnet 5 burns roughly 30% less quota than 4.6 in Devin over the same window. Anthropic’s launch post leads with autonomy: planning, browser and terminal use, long unattended runs. When one model release lands in five vendor surfaces the same afternoon, the number to watch isn’t the benchmark delta, it’s how fast the tool ecosystem now swaps its default.

The other Anthropic story today is about access, not capability. Claude Fable 5 returns globally after what the company calls productive conversations with the US government, redeployed with new classifiers that block a wider set of cybersecurity tasks. Routine coding and debugging will fall back to Opus 4.8 while those classifiers get tuned down for false positives. Anthropic also says it has begun drafting a shared framework with Amazon, Microsoft, and Google for rating jailbreak severity, and is giving the government pre-release access to models and safeguards for evaluation. This is the same thread that ran through last week’s Mythos 5 redeployment and the government-gated GPT-5.6 preview: frontier models increasingly ship through negotiated, partner-limited access, and that arrangement is hardening from exception into norm.

Google put out ADK Go 2.0, a rewrite of its Agent Development Kit around a graph-based workflow engine, with built-in human-in-the-loop checkpoints and dynamic orchestration. The framing matters: instead of chaining prompts and hoping, you describe the agent as a graph of steps with explicit control flow, which is the direction most serious multi-agent code has been converging toward anyway. Google paired it with a separate developer post on driving an “agent quality flywheel” from inside your coding agent, aimed at the problem where a prompt tweak that fixes one error regresses ten others in production.

The AI Daily Brief put a figure on the macro picture: AI is running at roughly a $175 billion annualized revenue rate, reading new Exponential View research as evidence the boom is more revenue-validated than the bubble discourse assumes. Token demand, compute, and power draw are the three curves reshaping the economics under that number. It’s worth holding against the licensing story above: the same models generating that revenue are the ones now gated behind government review, and those two facts will have to be reconciled.

On the research side, a widely circulated post argues that matrix orthogonalization improves memory in recurrent models. Applying Newton-Schulz orthogonalization to the recurrent state of an mLSTM keeps long-range information from collapsing into itself as the sequence grows, a concrete lever for the linear-attention revival that keeps trying to reclaim long context from quadratic transformers. It’s the kind of small, testable result that matters more than it looks if the recurrent line keeps closing the gap.

swyx amplified a sharper argument from Geoffrey Litt and Charlie Holtz about the words we use for AI-assisted work. Their target is “factories” as the reigning metaphor for AI-augmented software teams, a framing they read as a genuinely bleak vision of what the work becomes. The point underneath the aesthetics is real: the metaphors a field adopts shape the tools it builds and the outcomes it treats as success, and “factory” optimizes for throughput over craft by construction.

The monetization turn is arriving on hardware too. Meta put its AI smart glasses behind a paywall and added rate limits, per the Verge, an early signal that always-on, on-device assistants carry a per-query cost someone has to absorb, and Meta has decided it won’t be them.

Worth watching next: arXiv announced its next chapter, a rare first-party post on the future of the preprint server that most of the research in every one of these issues passes through first. When the substrate everyone builds on says it’s changing, it’s worth reading what it plans to change into.