There is one thread running through the week: what remains of a product when the component that powers it changes without the end user noticing. Replit moves the default from Anthropic to OpenAI, OpenAI closes training to the live web, Anthropic switches on an agent that works while you sleep. Three different moves, the same underlying question about where value lies: in the model or in the architecture around it. The convenient answer is "it depends." The useful one is narrower: it depends on how many parts of the system remain yours when the vendor changes its mind.
Image generated entirely with AI.Does the default beat the benchmark?
At TechCrunch Disrupt, Amjad Masad said that Replit has stopped relying on Anthropic as its main provider. Since August there has been a partnership with OpenAI, but only on Thursday did the explicit statement arrive. The stated reasons: costs, strength of the partnership, performance measured on an internal benchmark and a public one. The models cited are Sol and Astra, more solid at building "vibe-coded" applications from scratch.
The technical detail that matters is another: Replit remains model agnostic. Anyone who wants to select Anthropic models can do so, they are simply no longer the default choice. It seems small, and it is everything, because most users do not change models by hand. Moving the default moves traffic and spending. Masad confirmed that Replit was, until recently, among Anthropic's top five clients.
The official version does not convince everyone. Sources close to Anthropic maintain that the partnership is still active, and on Replit's late-September blog, Claude Fable 5 and Fable 5.1 appear in production. Meanwhile the numbers tell another story: enterprise went from 2% to 15% of revenue, B2B exceeds 50%, and by 2027 Masad expects more than half of revenue to come from companies and government.
The operational lesson is called default. In every production system, the default choice weighs more than the benchmark: the user stays on the already selected model and traffic shifts on its own. The battlefield is not cost per token, it is who controls routing. An agent that hardcodes a single provider into the code rebuilds half the product the day that vendor changes price or policy. With a multi-provider layer and a cost log per individual call, traffic shifts in half an hour and the delta is measured on the bill, not in the press release. The useful question is just one: how many in-house workflows survive a change of default? If the answer is few, the problem is the architecture, not the vendor. It is the same reasoning needed when the token price collapses and the weights become open.
How do you read old code?
REA (Reverse Engineer Anything) is an open-source MCP server that connects the coding agent to a stack of analysis tools: native binaries, Electron apps, Android APKs, .NET assembly, websites and firmware. One-command installation, Node 22 or higher, automatic configuration of Claude Code, Codex, Cursor, Gemini CLI and Grok Build, with backups of the files it replaces.
The correct framing is that REA acts as a coordinator on top of Ghidra, IDA and the other decompilers. Without understanding this, you install the server, point it at a binary and nothing happens.
The field test with Obsidian 1.14.4 is the most useful part. Before touching the agent, fifteen minutes building a throwaway vault and recording the quick-switcher rankings with timestamps. The tier model died immediately: a weighted numeric score is needed, with penalties for fragmentation, mid-word characters and increasing position. The falsifiable prediction that emerges is sharp: a word boundary is worth between 3 and 10 positions. Then the agent avoided the classic trap, distinguishing the 472KB bootstrap shim from the real 27MB application and targeting exposed functions instead of brute-forcing the minified code.
A plausible output on an obfuscated binary remains a hypothesis until proven otherwise.
On legacy code the discussion is the same, shifted by thirty years. OpenAI has published ChatGPT evaluations on COBOL, Fortran, Visual Basic and SQLite migrations, and the new GPT-6 models show gains precisely where previous generations stopped: code without tests, without documentation, with conventions that no one in the company remembers anymore. Banks and insurers have millions of lines of COBOL in settlement systems, and every modernization burns years of consulting before producing a single new line. The fact that the evaluations come out alongside the release is the signal that matters to whoever signs a budget.
The method that holds up is a cost hierarchy: cheap reconnaissance first, then brute force only on the confirmed target. Migrating means reconstructing behavior, and that means tests. Equivalence tests generated from the original code, the new version run through the tests, human review on every diff that touches money. The model acts as translator, the tests remain the judge. Implicit business rules are the point where a model that seems to have understood can make things up, and at that edge method matters more than the model, as in any project that touches processes and teams.
Image generated entirely with AI.
How much does an always-on agent cost?
Claude Duet is the week's announcement on the agent front. Anthropic describes it as an agent that keeps working when the user is offline: it runs on Claude Opus 5.5, connects to Gmail, Slack, Stripe and thousands of other services through over 6,000 connectors, prepares replies, manages tasks and financial analysis, and asks for explicit approval before sensitive actions such as payments. Waitlist access for Pro and Max, starting in the United States. The competition is moving along the same line: Grok Bot, Meta Muse, OpenAI's Dots. The game is played on connector quality and permission design.
The approval gate is the right choice. An agent running at night with access to email, chat and payments and no checkpoint is an accident waiting to happen.
The real issue is cost. An always-on agent reasoning over email and tasks is a meter running by itself: spending must be tracked per feature and per model, with a daily cap. In the first months the loudest discussion will be the bill, not the capabilities. Two metrics tell whether the design holds: how many actions end up in the approval queue and how many go smoothly after two weeks of real use. In the same week, the news that Anthropic nearly halted junior hiring while agents work six times as much went almost unnoticed: it is the organizational face of the same shift in default.
On the demand side, Italy has become a market. 18.9 million unique visitors in August 2026, more than double the roughly 9 million in January 2025, with four out of ten online Italians using AI assistants. Growth is slowing, however: nine percentage points in the first half of 2025, just two in the first six months of 2026. Intensity is rising, 80 minutes per month per user versus 44 the year before, still below the 94 in the United States and 118 in the United Kingdom. ChatGPT remains first with 13.9 million visitors but falls from 90% to 74% of the assistant audience, Gemini exceeds 9 million, and those using at least two tools in the same month go from one in six to one in 3.5. AI Overviews summaries reach over 20 million Italians and in June appeared in 22% of desktop searches on Google.
Multi-tool use has become normal behavior, and this rewards architectures that do not tie themselves to a vendor. The next number to watch is not the audience, it is the cost of agent autonomy.
Image generated entirely with AI.
SundAI comes out once a week, but AI news doesn't wait for Sunday. Every morning I pick the best one and comment on it.
TodayOctober 25, 2026
Coding agents cost triple and pass the same tests
Who verifies what the agent produces?
OpenAI will stop training frontier models on the live web in real time, citing significant risks. On the same day came a technical note on strengthening containment in training environments: offline environments and compute-intensive red-teaming. In the background, the audit by METR and Redwood Research, which over three weeks on-site found up to 25% of agent traffic unmonitored for GPT-5.5-class models and above.
The technical point is the perimeter. A model that reads the web during training absorbs poisoned content, hidden instructions and data that changes under its feet. A sealed environment makes training reproducible and verifiable. Those building on top of the APIs will see fewer contractual changes and more downstream defense work. The practical test to see if they are right is simple: in six months, ask a frontier model what happened the week before. If it answers well, containment works. If it hallucinates, the problem has only moved to retrieval.
On security, the week gave the opposite and more urgent case. Two critical vulnerabilities in PHP, dubbed PHP2Shell, opened a race against time across more than 70% of the web. Attackers began exploiting them before patches were available, and AWS recorded exploit attempts four days before the fixes were released. The discovery comes from AI research: OpenAI and Google's Wiz identified old holes. AI is getting good at finding old flaws at scale, and the same tools help attackers map dependencies and write variants faster. For those managing CMS, plugins or shared hosting, the differential is the time between disclosure and fix, not the elegance of the CVE writeup.
The pattern linking the two news items is the same: separate those who generate from those who verify. A judge model that checks another model's output, a pipeline that treats the retrieval result as untrusted input, a firewall between the agent and the world. Without that step, monitoring is theater, and the 25% of unseen traffic proves it: it is the same lesson as agents that leave the perimeter.
To put on the list this week
- DeepSeek Platform API: access to V4.1 Flash with Search and Fetch, for agents that read the web at low cost.
- REA: the MCP server for reverse engineering, with cheap reconnaissance before deep analysis.
- Voxtral 3: Mistral's voice stack with speech-to-text, text-to-speech and an SDK for real-time voice agents.
- Bedrock AgentCore MCP Gateway: a single authentication for MCP servers and A2A agents distributed across multiple clouds.
- 1Password for AI agents: agents that access your accounts on your behalf without ever seeing the password.
