The thing reading your pricing page isn't a customer any more

The 60-second catch-up

Something else is reading your pricing page, and it mostly can't. Kyle Poyar and Nikolas Laskaris ran 7,600 AI responses to pricing questions about the Cloud 100 across six engines. A company's own pricing page appeared somewhere in 46% of answers — but was cited first in only 12%. No company in the set had its own page cited first in a majority of runs. The mechanical reason is duller than the strategic one: of 77 public pricing pages, 57 were fully readable to bots, and ten hid at least 40% of their body content behind client-side rendering, accordions, iframes or a robots.txt nobody has looked at since 2023. What fills the gap is Vendr (18.7% of runs), Reddit (18.6%) and G2 (15.9%). Separately, Trellner traced three linked sites that between them published 215,128 generated "best software" pages, and found Perplexity citing them; Wikipedia was cited three times in 7,534 responses. Your pricing is being described whether or not you participate. → do this: curl your own /pricing and read what actually comes back. If the numbers aren't in the initial HTML, no crawler has ever seen them.

The price that fell this week was the price of repetition, not the price of intelligence. Anthropic cut Fable 5.1's cache reads by 75%, to $0.25 per million tokens, and puts the resulting saving at roughly 25% on typical workloads and up to 45% on highly agentic ones. Reducto's new parser charges a cent a page where its previous pipelines charged three to six. Google says Gemini 3.8 Flash Cyber patches at 47.2% pass@1 against a leading frontier model's 47.8%, at significantly lower cost. Meta will sell you the same Muse Spark model at $0.10 or $1.25 per million input tokens depending on whether it may learn from your traffic. And one researcher hit 44% on ARC-AGI-1 for 67 cents of total lifetime compute — training from scratch included. None of these is a capability story. All of them change what you can afford to run in a loop. → do this: find the block of context your agent rereads on every single turn, and price that one thing again at today's numbers.

Capability went up, and so did the volume of work nobody asked for. Fable 5.1 more than doubled Anthropic's own agentic-science score, 52.6% against 24.7%. Then read Anthropic's prompting guide for the same model, which tells you to keep scope tight, to say explicitly that you want the whole assignment finished, and to tell it not to helpfully fix nearby bugs or add tests you didn't request. That is a release note and a bug report in the same document. Two of the better essays this week come at it from the other side: agents finish the task without teaching you anything, and speed applied to a bad decision is still a bad decision arriving sooner. → do this: put the boundary in the prompt, not in the review. "Do not touch anything outside these files" costs you one sentence and saves you a diff.

Tools

Build & ship

• ⭐ Wigolo — Local-first web search, fetch, crawl and research for coding agents, over MCP. The ranking, embeddings and browser engine run on your machine, so there is no per-query meter — the trade is about 1.5 GB of disk. Works with Claude Code, Cursor, Codex, Gemini CLI and any MCP client. Public beta. · Open source (AGPL-3.0)
• @huggingface/kernels — 207 WebGPU kernels published as individual Hub repositories, plus a JavaScript loader that downloads and runs them in the browser. Every kernel ships a manifest, correctness tests and benchmark cases rather than a bare shader. · Open source (Apache-2.0)
• WebLLM — In-browser LLM inference on WebGPU, with worker and service-worker APIs so the model runs off the main thread. · Open source (Apache-2.0)

AI & agents

• ⭐ Reducto r-1 — Document parsing aimed at the cases that break pipelines: strikethroughs that change meaning, watermarked pages, nested tables, documents that follow no template. One pass, one rate, no credit multipliers to reach high accuracy. Reducto reports a 20% error-rate reduction against its own legacy agentic pipelines in early preview. · $0.01 per page
• Muse Spark 1.3 — Meta's agentic coding model on an OpenAI-SDK-compatible endpoint: 1M context, native video/image/document perception, plus computer-use and search-grounding tools. Note the two price tiers for the identical model — the cheap one is cheap because Meta may use your traffic to improve its products. · $1.25 / $4.25 per Mtok in/out, or $0.10 / $0.20 on the contributor tier

Resources

Steal the template

• AI pricing visibility audit — Kyle Poyar and Nikolas Laskaris turned the pricing-page findings above into a Claude skill you point at your own site. Free download.
• The Fable 5.1 prompting guide — Anthropic's own list of what to change: re-test effort from scratch, keep conversation history append-only, batch independent tool calls, ask for surgical edits rather than whole-file rewrites, and say when you want it to search rather than answer from memory.

Learn the craft

• AI Systems Atlas — 78 diagrams across five collections: agent harnesses, safety and security, evals, context and memory, coding agents. One visual grammar throughout, with attack/defence toggles on the security plates.
• The AI engineering skills map: software engineering fundamentals — which engineering areas still have to be in your head for a coding agent to ship production software rather than a brittle demo.
• Training frontier knowledge-work agents: a 397B RL guide — Mercor and SkyRL document the full recipe and task format, including the parts that went wrong, so you can point it at your own data.

Benchmarks you can re-run

• 44% on ARC-AGI-1 in 67 cents — a transformer trained from scratch on a rented 5090. The write-up includes the objections and answers them, notably that 67 cents is lifetime compute, not inference-only.
• Cat Doom — one of those recurring one-file browser-game prompts people use to compare frontier models. Playable, which is more than most benchmarks manage.

Reads

☕ Under 5 minutes

• I used Fable to rewrite 65kLoC to Rust. It cost $400 — the structure that made a large language translation tractable, and where the cheap subagents went.
• Inside Meta's Infrastructure Lab — a one-minute look at the Menlo Park hardware effort.

🍵 5–10 minutes

• Fluid Compute — Vercel on the compute layer underneath its builds, sandboxes and functions, and why bursty agent workloads broke the old shape.
• What comes after HBM — magnonics, vertical FeRAM, and an honest assessment of how far away any of it is. Also a good read on what "hundreds of millions to find out" does to a founding team.
• AI can make you suck faster too — the counter-case to velocity: output went up, quality did not, and the expertise that would have caught it never got built.
• Exit the Cave — a wrestling essay that is really about shipping. Controllable inputs are comfortable precisely because they are controllable, which is not the same as them determining the outcome.
• Reasons robotics is hard — a list worth reading before you assume the atoms part is a solved integration problem.

📚 Longer, worth it

• What AI agents really think about your pricing — Kyle Poyar and Nikolas Laskaris, on 7,600 responses across six engines. The section on Plaid, whose billing docs out-cite its pricing page, is the part to steal.
• Three sites made 215,128 "best software" pages for AI. Perplexity cites them — Trellner's forensics: same registrar window, same nameserver pair, same template, three different rankings for the same category.
• Building autonomous goal loops that deliver — the harness matters more than the model: exposing failures, locating missing capabilities, and keeping the accountability line visible.
• 9 harsh truths nobody told you about success — Sahil Bloom. The local-maximum framing and the "margin of freedom" idea are the two that survive a second reading.
• Mastery still comes from doing the reps — Addy Osmani on skill decay: if the agent finishes the task, nothing taught you anything, so the learning has to be made deliberate.
• A deep dive into StyleX — typed style objects compiled to deduplicated atomic CSS. The argument that verbosity matters less when an agent is doing the typing is the interesting half.
• Optimizing on-device inference for Apple silicon — unified memory, sparse MoE routing, and what you can actually get out of one Mac.
• Improving our alignment and security practices — Anthropic's own account of the July and August incidents, and the containment it put in place afterwards.
• Claude Fable 5.1 can do the work. The hard part is managing it — a long live stress test: games, computer use, Blender, and the moment the model explains a style decision nobody asked it to make.
• Hugging Face attack postmortem — 98 minutes, and the longest argument you will read this week that the engineering-failure framing is the wrong one.

Numbers

• $0.25 per million tokens — the new price of a Claude Fable 5.1 cache read, down 75%. Anthropic puts the resulting saving at about 25% for typical workloads and up to roughly 45% for highly agentic ones (Anthropic).
• 52.6% against 24.7% — Fable 5.1 versus Fable 5 on Terminal-Bench-Science 0.1, Anthropic's agentic scientific-research benchmark. Opus 5 scored 29.0% (Anthropic).
• 12% — share of AI pricing answers in which a company's own pricing page was cited first. It appeared somewhere in 46% of the 7,600 responses tested (Kyle Poyar and Nikolas Laskaris).
• 57 of 77 — public pricing pages among the Cloud 100 that were fully readable to bots. Ten hid at least 40% of their body content (Kyle Poyar and Nikolas Laskaris).
• 18.7% · 18.6% · 15.9% — citation share of Vendr, Reddit and G2 respectively in those pricing answers. Two of the three are places you cannot edit (Kyle Poyar and Nikolas Laskaris).
• 215,128 — "best software" pages published by three linked sites that Perplexity cites. Wikipedia was cited three times in 7,534 responses (Trellner Research).
• 13.9% to 19.2% — rise over the past year in the share of ChatGPT queries carrying commercial intent (Profound).
• $0.01 per page — Reducto's r-1 parsing price, against three to six cents on its legacy agentic pipelines (Reducto).
• $1bn at about $47bn — the round Cognition is closing, having fielded close to $10bn of investor interest (Bloomberg).
• 47.2% against 47.8% — Gemini 3.8 Flash Cyber versus a leading frontier model on CWE-Bench patching, at what Google describes as significantly lower cost. The same model scored 54.9% on HLE-Verified (Google).
• 16.11% to 27.29% — APEX-Agents pass@1 for Qwen3.5-397B-A17B after RL post-training, a 70% relative gain. APEX-Agents is 480 realistic knowledge-work tasks (Mercor).
• 67 cents — total lifetime compute, training from scratch plus inference across every task, for 44% on ARC-AGI-1 (Mithil Vakde).
• 12.5× — the gap between Meta's two published prices for the same Muse Spark 1.3 model: $1.25 per million input tokens, or $0.10 if Meta may use your traffic to improve its products (Meta).

What happened

Models & math

• Anthropic released Claude Fable 5.1 and Mythos 5.1, cutting cache-read pricing by 75% and loosening safeguards it says had been interrupting legitimate work — biology interventions on benign requests down 85%, cyber-safety interventions in Claude Code down roughly 60% (Anthropic).
• Google shipped Gemini 3.8 Flash and 3.8 Flash Cyber. The Cyber variant carries more permissive cybersecurity mitigations and is restricted to vetted defenders (Google).
• OpenAI said its forthcoming Astra model is the first to cross its own 'Critical' cybersecurity capability threshold, able to find and exploit unknown flaws without step-by-step human guidance. Access to those capabilities will be limited at launch (CNBC).
• Meta introduced Muse Voice Transcribe, its first real-time audio perception model: streaming recognition, diarization past 20 speakers, and mid-sentence code-switching. Meta reports first place on Artificial Analysis for streaming speech-to-text as of 1 September (Meta).
• Google added agentic video understanding to several Gemini models, letting the model choose which moments, frames, speeds or transcript segments to inspect (Google).
• Hugging Face published @huggingface/kernels — 207 Apache-2.0 WebGPU kernels with a browser loader and a crowdsourced benchmarking tool (Hugging Face).

Safety & policy

• Anthropic published a postmortem on unauthorised access to Claude models in July and August, tracing it to misconfigurations in third-party evaluation environments. It paused those evaluations and published best practices for external partners (Anthropic).
• A US court declined to break up Google's ad tech business. Judge Brinkema had found Google held a monopoly over publisher tools and the technology connecting publishers to advertisers, but rejected the government's request that Google sell its ad exchange. Google has said it plans to appeal the underlying finding (The New York Times).
• Mistral's documented process for opting out of having your inputs and outputs used for training did the rounds this week. Worth reading for the default it implies rather than the option it offers (Mistral).

Business moved

• Cognition is set to close around $1bn at a roughly $47bn valuation, with talks ongoing (Bloomberg).
• Manus resumed independent operations. Some users hit temporary data-access interruptions requiring backup and restoration (Manus).
• Meta opened up its Infrastructure Lab in Menlo Park, focused on hardware for next-generation AI (Meta).

Platforms shifted

• Vercel described Fluid, a unified compute layer that reconfigures for different workloads and absorbs burst capacity. It says the system already serves over a trillion requests a month across builds, sandboxes and functions (Vercel).