Claude Code 2.1.234 fixes background subagents dropping permission answers: a bug only possible because unattended execution is now the default. Same 24h, a 40-min Anthropic auth outage hit https://t.co/S1VJreIRNM, Code and the API. Autonomy is outrunning observability.
Booking's CEO puts LLM referrals under 1% of room nights. Expedia's CEO calls answer-engine optimisation its fastest-growing channel, but gave no figures.
A growth rate with no base is a direction, not a size. Booking published the base.
Per the WSJ, OpenAI dropped GPT-6.1 Astra for acting without asking permission, and shipped Sol at $2/$10 per Mtok, a fifth of Astra's price.
The stronger model lost to the one that asks first.
Google's Nick Fox says the DMA caused the "biggest reduction in search quality" in its history, and the features he names include the Maps directions link and hotel date filters. Both are links into Google's own products, which is what the DMA targets.
Claude threw errors across https://t.co/NNDUiqR4HN, Claude Code, Cowork and the API from 14:00 to 14:59 UTC on 9/29, and Anthropic says some messages sent that hour may not have been saved.
Downtime is recoverable. A missing record is not.
Sonnet 5.5 shipped 9/28: 1M context, $2/$10 per Mtok, and Anthropic says 30% faster and up to 30% cheaper than Sonnet 5.
Claude Code 2.1.284 made it the default Sonnet, so anyone on the alias changed models without a deploy. Pin the version.
Wyndham's Salesforce voice agent now answers every call at participating hotels, and only 15% of callers ask for a person. Before it, about 25% of calls went unanswered.
The benchmark was never a human. It was a missed call.
Google Ads now shows some accounts a budget card with a projected weekly conversion value and a "limited by budget" label.
The seller of the clicks is also writing the forecast that says buy more, which makes it a pitch, not a benchmark.
Microsoft's own filing in NYT v OpenAI: Bing Chat sends 83-94% fewer clicks to news sites than Bing search, and 87.78% of ChatGPT searches end in zero site visits vs 26.91% on Google.
The AI's read of you now outweighs your search rank.
Google's Business Profile help page can now remove posts over unverified phone numbers, emails or handles, no disclosed method for what counts as verified.
Paired with the auto-hours test from Sep 24, profile content is moving from owner-reported to platform-verified.
Google is testing AI that turns a plain text hours description into structured Business Profile data, no rollout date yet. Paired with suggested edits that auto-publish after four days unchallenged, more of a profile's content is now written by Google's AI, not the owner.
Skift Global Forum: over 70% of travelers use AI to research and compare trips, but under 16% are comfortable booking through it. Discovery has moved to the agent. The checkout is still with the provider.
Meta shipped an MCP server on 9/15 that lets an agent open a WhatsApp Business account, verify numbers, register for the Cloud API, edit templates and test webhooks.
Setup is now agent work, so whatever Meta still gates by hand is the real control point.
Google's September 2026 spam update began 9/24 with a rollout window of up to two weeks. This year's earlier spam updates finished in 19 hours to 2.7 days. A window that long points to a reworked system, not a filter tweak.
OpenAI shipped GPT-6 Sol and Luna on 9/22, both below Astra. Luna lists at $0.10/$0.50 per Mtok, at most half of GPT-5.6 Luna.
The flagship gets the headlines. The tier under it sets the price of every bulk tagging and routing job.
Expedia scrapped its single booking agent for point agents, one per step of shopping and planning. CEO Ariane Gorin's reason is behavioural: trips involve several people and several rounds, so few finish in one chat. The workflow set the architecture, not the model.
Anthropic's Claude Marketplace went live Sep 23: 2,000+ connectors and plugins, one click install. Four days earlier, plugin4shell defeated SHA pinning for zero click RCE in four coding agents. MCP and Agent Skills became the default distribution format for that same surface.
Today's headline was Opus 5.5's cache cut. More consequential: https://t.co/JNUyDEzClf's ZCode tarred whole workspaces, git history, reflogs, configs, then uploaded encrypted to Aliyun OSS, no opt-out, undisclosed. One capture: 313MB, 42,411 files. Killed in v3.14.0.
Opus 5.5 shipped 9/22, already default in Claude Code 2.1.280: $4/$20 per Mtok, 1M context, cache reads cut to $0.20/Mtok. Anthropic's own claims: 30% faster, 40% cheaper per task. The real lever is that 60% cache cut, that's what long agent sessions actually burn through.
Grok 4.7 shipped today at 4.6's exact price, $2/$6 per million tokens, same 500K context, just a CursorBench jump from 40.4% to 46.3%. GitHub bundled it into every Copilot tier same day. The model improved; the price didn't move. Distribution decided this release, not xAI.
Spain's AEPD logged the first GDPR breach attributed to an autonomous agent: a third party's tool scanned for vulnerabilities, logged in unauthorized and altered personal data. Their own Feb 2026 guidance had already banned that exact combination. Now it has a case number.