Anthropic and OpenAI cut frontier prices on the same day. The next saving is in your harness, and the next fight is over which doors let agents in.

Tools

Build & ship

• ⭐ Strands harness — AWS's fully assembled general-purpose agent, not a coding agent: one line of Python or TypeScript, run it locally or deploy it, with any model from Bedrock, Anthropic, OpenAI or Google, or a local one through Ollama. AWS reports 28% lower cost than other harnesses on the same models across six benchmarks, and 77% under Claude Code with Fable 5 on Terminal-Bench 2.1. · Open source (Apache-2.0)
• Unreal Agent — An agent harness that runs every tool call in the background and logs it as in progress, so the model never waits or polls and you can steer it mid-call. Unreal Labs reports up to 40% less than Codex and 20% less than Pi with GPT-6 Astra, at matching pass rates. · Open source (MIT)
• Devin Cloud in your terminal — /handoff moves a task from the Devin CLI to its own cloud VM that you keep steering from the terminal; devin ssh gets you into that VM, and /handoff from the cloud pulls the PR branch back to a local session. · Free SWE-2 sessions until 8 October
• GeaStack examples — The app gallery for GeaStack's TypeScript-and-CSS app stack, with iOS and embedded-board targets. Each example is a small package with a gea manifest. · Open source (MIT; embedded board support GPL-3.0)

AI & agents

• ⭐ Claude Opus 5.5 — Anthropic's new Opus, at about Fable 5.1's level on most work by its own account. 66.4% on Terminal-Bench 4.0 against Opus 5's 52.3%, output more than 30% faster, and a fast mode at up to 2.5× speed. Thinking can no longer be switched off. claude-opus-5-5 on every major cloud. · $4 / $20 per million input / output tokens · $0.20 cache reads
• ⭐ GPT-6 Sol and Luna — OpenAI's mid and small GPT-6 models, trained like Astra. Sol scores 68.8% on DeepSWE v1.1 at max effort; Luna at higher effort matches GPT-5.6 Sol on OpenAI's factuality eval for about a hundredth of the cost. In ChatGPT Work and Codex for paid plans; Luna for Free and Go in the desktop app. · Sol $2 / $10 · Luna $0.10 / $0.50 per million tokens
• Aikido Altar — Aikido's open-weight security model: GLM-5.3 quantised and expert-pruned from 1.51 TB to 328 GB so it serves on one 4×H200 node. It runs Aikido Machine's pentesting inside your own network, air-gapped if needed. · Open weights

Design & create

• RadianceKit — 3D Gaussian splatting on a Mac: photos, video or 360° footage in, trained on the Apple Silicon GPU with no cloud upload, edited and exported in six formats. An iPhone capture app adds camera poses and LiDAR depth. · Free 3-day trial · US$7.99 once

Resources

Steal the template

• Eval skills — Hamel Husain and Shreya Shankar's plugin for coding agents. evals-start routes you to eval-audit if you already have an eval pipeline, or to error-discovery if you have traces nobody has read; that one builds a review app shaped to your data.
• Your company needs a pricing constitution — Elena Verna's five laws, written in Lovable's own wording: never manufacture pain to monetise it; monetise the outcome, not the steps; don't tax your own distribution; one primary usage metric and four plans at most; self-serve lands, sales expands. Plus a small monetisation council to settle the edge cases.
• No- or Lo-JS options — Aaron T. Grogg's growing collection of common JavaScript patterns you can replace with plain HTML and CSS, with a reading list going back to 2016.
• The Designer's Toolkit for the AI age — components, pre-built blocks, AI chat components, 26 brand themes and a theme generator that exports tokens, plus guides to design process with AI.

Learn the craft

• Advanced evals: find the hidden failures — Hamel Husain and Shreya Shankar on error discovery before metrics: annotate 10 traces yourself before the agent suggests anything, aim for 100, stop at the first upstream error. The preview is free; the full post is for paid subscribers.
• Designing with AI — Buzz Usborne on Help Scout: a design.md that writes down the unspoken parts of a 15-year-old product philosophy, and two weeks rebuilding the design system so more than 200 primitives and compositions are legible to an LLM.
• How AI will change Windows — Gergely Orosz and Ivan Klaric inside Microsoft's plan: agent identities in Task Manager, the On-Device Agent Registry for local MCP servers, MXC containment levels from process to full VM, and Windows ML. Paid; the preview covers the first four sections.
• Your presentations feel lifeless because they're missing a villain — Kai Wong: the villain in a design story is the status quo, and it is already winning when the story opens.

Benchmarks you can re-run

• RecreationWorld — Qwen's environments for computer-use agents that rebuild a running app across five platforms, scored on observable behaviour; RecreationBench holds 250 tasks. MIT.
• JevBench — Benchmark Heaven's own benchmark for typed decision models: 16 systems on 534 decisions, scored on intelligence, calibration, speed and cost, with the harness and public tasks under MIT.
• Opus 5.5 on Artificial Analysis — the independent index score and, more usefully, how many tokens it took to earn it.
• Unreal Agent's benchmark runs — Terminal-Bench 4.0 and science runs with cost, tokens, turns and tool calls per trial, each linked to its Harbor run.

Reads

☕ Under 5 minutes

• AI comes for the if statement — Tom Tunguz swaps a quarter of an agent's model calls for Jev and SemIf and shows his numbers.
• GPT-6 Astra breaks the Enigma message MVUEH — Frode Weierud on a 1941 message unsolved since 2005. The model picked the target itself, found the crib, and wrote its own Enigma simulator and Bombe.
• Adobe's hover-to-prompt patent — a published application for generating the AI result from whatever the cursor is over, before the click. The cost is wasted work on every wrong guess.
• Local maximums — Sean Byrnes on staying at okay because every move off it hurts first. Staying is fine; staying by default is the trap.

🍵 5–10 minutes

• Will OpenAI eat Jev's lunch? — Arcturus Labs argues OpenAI has used its models as implicit classifiers for years and is placed to fast-follow, and that TypeSafe's moat is its training data.
• The design GM — what it takes for a designer to own a P&L, with Bruce Bell at Square as the rare example.
• AI can build your UI. Now Figma can tell it when it's wrong — Applitools uses a Figma frame as the visual baseline for what an agent built.
• SAML: a fractal of bad design — Trail of Bits on why SAML should be retired in favour of OpenID Connect.
• Data-only attacks are easier than you think — a tool called Einstein generates exploits that never touch control flow, and the authors argue vendors need to rethink mitigations built around control-flow hijacking.

📚 Longer, worth it

• Swarm scaling — Toby Ord on why a swarm of agents is a way to buy speed, not efficiency, worked through the 10,000-agent Navier–Stokes run.
• The business of building God — Rohit Krishnan on the labs as businesses: one or two models ahead of open source, and the traces you send them as the moat.
• The current balance of power in open models — Nathan Lambert's briefing to members of Congress on open-weight models and US–China competition.
• The design history of Wikipedia — It's Nice That on the people behind the encyclopedia that puts function above looks, from the puzzle globe to the mascots.

Numbers

• $4 / $20 · $0.20 · 40% — Claude Opus 5.5's price per million input and output tokens, per million cache reads (down 60%), and the cost cut Anthropic expects on typical workloads (Anthropic).
• $2 / $10 · $0.10 / $0.50 · 50% — GPT-6 Sol and Luna per million input and output tokens, and the cut against GPT-5.6 promotional pricing (OpenAI).
• 66.4% · 57.9% · 52.3% — Terminal-Bench 4.0 for Opus 5.5, GPT-6 Astra and Opus 5 (Anthropic; Astra as reported by OpenAI).
• 33.2% · $0.27 · 11.1× — GPT-6 Sol at xhigh on AutomationBench, its cost per task, and Opus 5's cost per task at max effort relative to it (OpenAI).
• 58 · 260M · 91M — Opus 5.5 on the Artificial Analysis Intelligence Index, the tokens it used to run the index, and the median model's (Artificial Analysis).
• $600 · $7,000 — daily token usage at API prices for OpenAI's median researcher and a researcher at the 90th percentile (OpenAI).
• 16 of 18 — Opus 5.5's financial reports that passed Anthropic's check for invented figures and quotes; Fable 5.1 and Opus 5 passed none (Anthropic).
• 28% · 77% — Strands harness's cost saving against other harnesses on the same models, and against Claude Code with Fable 5 on Terminal-Bench 2.1 (AWS).
• 47% → 80% · 98 — classification accuracy after moving from a generative model to a Jev decider, on hand-verified production email threads (Tom Tunguz).
• 3–5× — the longer single-agent chain of thought that ten times as many agents in a swarm is worth (Toby Ord).
• 730,000 · ~5 days — Muse downloads after its 8 September launch (Sensor Tower, CNBC).
• 38% · 28% — share of recent AI papers mentioning a Chinese open-weight model against an American one (Nathan Lambert).
• 2.2× · 68% · 41% · 35% → 83% — Shopify's workflow builder speed and cost against the frontier system it replaced, Cursor's cost cut from routing, and Ramp's receipt-matching accuracy, all credited to evals (Hamel Husain and Shreya Shankar).
• 9,937 — responses to a September developer survey that put Windows in third place, behind Linux (Gergely Orosz and Ivan Klaric).
• 1.51 TB → 328 GB — GLM-5.3 at full size and after Aikido's quantising and expert pruning for Altar (Aikido).
• ~$25 · 2 of 6 — what the average developer earns a month from Kickbacks' sponsored loading line in Claude Code, and the ad networks for that line that shut down within months (Dru Riley).
• $899 · 4 October — Googlebook's starting price and US shelf date (Google).
• 50%+ · 27 — Waymo riders in its most mature cities who also use public transit, and the Bay Area agencies in its new transit rewards (Waymo).

What happened

Models & math

• Anthropic released Claude Opus 5.5, the first of the Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks, higher five-hour limits on paid plans and a saved rate-limit reset for subscribers (Anthropic).
• OpenAI released GPT-6 Sol and GPT-6 Luna the same day and improved prompt caching for GPT-6, with 90% off cached input reads (OpenAI).
• GPT-6 Astra broke the German Army Enigma message MVUEH of 10 July 1941, unsolved since 2005, after choosing it from a list of unbroken messages and writing its own simulator and Bombe in Python and C++ (Frode Weierud).
• Aikido released Altar, its first open-weight security model, to run its pentesting appliance inside customer infrastructure (Aikido).
• Qwen released RecreationWorld, which trains computer-use agents by having them rebuild a running app (Qwen).

Safety & policy

• OpenAI and Anthropic were negotiating a legally binding agreement to stress-test each other's commercial models; whether it was finalised is unclear (The Information).
• Pentagon investigators found flawed intelligence, outdated imagery and over-reliance on AI contributed to the February missile strike on a school in Minab, Iran, that killed at least 123 children (Bloomberg).
• Opus 5.5 ships with cyber and biology safeguards similar to Fable 5.1's, re-routes most cybersecurity tasks to Opus 4.8, and applies preserved thinking against distillation to API accounts created from 31 August (Anthropic).
• ShinyHunters claims it holds data on all FBI employees and applicants; 404 Media reviewed a sample of 5,000 records with names, addresses and spouses' details (404 Media).
• WordPress disclosed an unauthenticated path traversal in page-template resolution that can lead to code execution on themes with a top-level page- folder, including Twenty Twelve and Twenty Fourteen (WordPress).
• The US embassy in Canberra called Australia's proposal to let users opt out of social media algorithms "censorship of protected speech" (BBC).

Business moved

• Amazon blocked Muse from shopping on Amazon.com after asking Meta to remove Amazon from the experience (GeekWire).
• Shopify put Shop Pay agentic checkout into Muse across its stores, announced by Tobi Lütke and Mark Zuckerberg; Shopify shares rose about 7% (24/7 Wall St.).
• Muse overtook ChatGPT as the top free iOS app in the US (Sensor Tower, CNBC).
• Kickbacks and five other networks began selling sponsored messages in Claude Code's loading line from June; two shut down by September without Anthropic blocking anything (Dru Riley).
• Waymo launched transit rewards, paying riders Waymo Cash when a ride connects to public transit, starting with employees in the Bay Area (Waymo).

Platforms shifted

• Google opened Googlebook pre-orders: five laptops from $899 to $1,299 with Gemini built in through Magic Pointer, shipping in the US on 4 October (Google; TipRanks).
• AWS released Strands harness, extending its Strands Agents SDK from May 2025 (AWS).
• Cognition put Devin Cloud sessions in the terminal with SSH into the agent's VM (Cognition).
• Windows is adding agent identities through Entra ID, an On-Device Agent Registry for local MCP servers, and MXC containment, already in use by OpenClaw for Windows (Gergely Orosz and Ivan Klaric).
• Applitools released Figma Design Baselines and expanded its Eyes MCP server so coding agents can review visual tests (WebDesignerDepot).
• Adobe's patent application for hover-generated AI prompts was published on 17 September (Patentlyze).
• Unreal Labs open-sourced Unreal Agent (Unreal Labs).

We skipped five items we covered yesterday (Grok 4.7, MiMo-V2.6, Kev, AX and the OpenAI maths advisory group), a Claude Projects redesign we covered last week, two iPhone rumours, a trend roundup for TikTok creators, a founder's travel diary, a paywalled Reddit playbook and a pile of sponsor placements.