Notice: This article was created with AI.
Whether a model finds security vulnerabilities, tops a leaderboard or triggers orders: this week, the model’s own performance hardly decided how useful it was. What mattered was what is organized around it, namely how quickly updates reach users, what a solved task costs, what rights an agent has and who checks the result. This pattern runs through attack and defense, stock market figures and price lists, enterprise software and everyday law firm work. The capabilities of the systems are considered proven. What remains open is who steers them, who pays for them and who stands behind their results.
Black October: How dangerous are AI models for cybersecurity?
The term was prompted by a vulnerability in PHP. The KI-Beratung blog states in PHP-Sicherheitslücke: So gelingt die KI-Verteidigung that it allows attackers to execute foreign code without a password. It was a zero-day vulnerability, meaning a flaw for which no fix existed when it became known. According to the blog, it was discovered by the model GPT-5.6 Sol. Institutions such as the Charité and the BKA subsequently took systems offline as a precaution. The blog also stresses an information gap: the US government knew five days before publication, while Europe, in its account, was warned inadequately.
The weighting of these events differs depending on the source. In its video “Schwarzer Oktober hat begonnen”, the channel Everlast AI also names the Charité and several federal ministries, but puts it into perspective: around 70 percent of the internet was not affected, and many organizations shut down their systems only out of caution. Both sources point to real attacks by a Chinese hacker group. According to KI-Beratung, it used AI agents to steal credentials from more than 2,250 company accounts. The incident shows that AI not only finds vulnerabilities but also carries out attacks.
Epoch AI provides a concrete example involving open models in Open-weight AI models can exploit the patch gap. Open models are those whose weights can be downloaded freely. The model GLM 5.3 reportedly exploited vulnerabilities in six current Firefox versions. It relied on the “patch gap”, the period between the release of a security update and its actual rollout to users. The attack read files from the computer and sent them to the attacker without the user having to do anything. Epoch puts the cost at about $1,000 per Firefox version. The researchers also found an unpublished vulnerability, which they combined with the known flaws. Mozilla has since provided an update. Epoch interprets the result as a growing threat from open models.
The newsletter Interconnects takes the opposite position. In Was I wrong about open models & cyber risks?, the author places the events under the term “Black October”, with PHP2Shell and GitLab as examples. His finding, in essence, is that the reaction of the AI safety community was exaggerated. Today’s open models did not have the feared catastrophic consequences. The greater risk comes from closed models and from the cyber industry as a whole. Cyber risks are a societal problem that the tech industry cannot solve alone. Epoch and Interconnects disagree less about the facts than about how to read them. The Firefox attack works only because updates arrive late. That is a problem of distribution and patching discipline, not of the model alone. At the same time, it shows that the entry barrier for open models is low.
That even experts disagree is also shown by the Interconnects podcast Kimi K3.1 Pro, the new cyber reality, and Muse musings. It covers the mixed reception of Kimi K3.1 Pro, which raises cybersecurity questions despite strong benchmarks. It discusses the differing perceptions of cyber and AI experts, the need to patch and the role of nations in cybercrime. The defenders are also arming up. According to Latent Space in Gemini 4 Carbon tops leaderboards, launches for cyber defenders, Google DeepMind reports 88.2 percent for vulnerability detection and 54.8 percent for exploitation with its new model. Access is initially limited to trusted defenders in the Fairwind program, with a broader opening announced. Here the question of control becomes concrete: the provider decides who gets models with offensive capabilities.
A practical consequence remains for companies. In Pentests für KI-Systeme, Computerwoche points out that generative AI already drafts code, searches knowledge bases and carries out actions via connected tools. Penetration tests, meaning controlled attacks on one’s own system, are therefore essential for AI applications as well. Beyond this week, what matters is that the debate is shifting from whether models can find vulnerabilities to how quickly organizations respond. The dispute over open and closed models depends on whether patches arrive in days or in weeks.
Anthropic on the stock market: What the numbers reveal about agentic AI
Anthropic has filed an application for an IPO in the US, the so-called S-1 document, in which a company must disclose its business before listing. According to The Batch, the stock is to be traded on the Nasdaq under the ticker ANT. Revenue in the first half of 2026 was reportedly $16.3 billion, compared with around $1.2 billion in the same period a year earlier. The KI-Update from heise summarizes this roughly as follows: 14-fold growth, which makes the IPO one of the largest in history. For the second quarter alone, Latent Space cites growth of 1,368 percent year over year and a valuation target of two trillion dollars.
More interesting than the growth is the cost side. Latent Space reports an operating profit of $974 million for the second quarter. At the same time, the books show a total loss of $80.7 billion for the half year, which according to this source stems mainly from non-cash expenses, meaning items where no money flows out. Cash holdings stand at $134 billion, and the amount actually burned so far is about $9.7 billion. The Batch adds the lever behind this: spending on compute and infrastructure was still about 6.5 times revenue in 2024 and fell to roughly half of revenue in the first half of 2026. The adjusted operating margin is therefore about 10 percent. The $80.7 billion loss thus says little about the ongoing business, while the compute cost per revenue dollar says a lot.
The sources shed light on how Anthropic earns its money from different angles. The Batch attributes the growth to agentic products, meaning software that completes tasks independently across several steps, namely Claude Code and Claude Cowork. As evidence it cites enterprise customer retention of about 900 percent. This apparently refers to a metric that measures revenue from existing business customers over time; the exact definition is not given in the material. Heise, by contrast, emphasizes that around 80 percent of revenue comes from fees for using the AI models. That is not a contradiction but a different perspective: the products drive consumption, but what is billed is essentially the model. Latent Space cuts the numbers yet another way and shows that 48 percent of revenue runs through cloud partners. Nearly half of the business thus reaches customers via third-party platforms.
The sources give several hints about where Anthropic wants to go next. According to Latent Space, the company states in the filing a goal of reaching more than a billion knowledge workers by 2028 with “virtual employees” that take on complex tasks. Claude Duet fits this, which The Batch describes as a new agent for task automation aimed at professionals and small businesses. Heise sketches a further step: Anthropic is reportedly consulting with law firms on how agents should negotiate prices, salaries and contracts for their users in the future. This is so far a report without confirmed details, but it shows the direction, namely agents that not only work but negotiate bindingly on behalf of people.
Beyond this week, the filing matters because it provides, for the first time, reliable figures on how much revenue agentic AI carries and how fast compute costs per revenue dollar can fall. Whether the margin of ten percent holds if usage keeps growing this steeply and the share of cloud partners stays high will be shown by the first quarterly reports as a listed company.
Who is ahead is only half the answer
The analysis firm Artificial Analysis reshuffled the top of the model ranking within two days. On October 27, it presented the fifth version of its Intelligence Index in Announcing Artificial Analysis Intelligence Index v5. Two benchmarks are new, meaning standardized collections of test tasks: FrontierCode 2.0 with real programming tasks and Terminal-Bench-Science with 70 research tasks in a terminal environment. In addition, the index now relies 60 percent on private datasets, whose tasks are not public and which the models are therefore not supposed to know in advance. There is also a Progress Index, which is meant to make models from different release periods comparable. In this first version, Claude Opus 5.5 led, followed by GPT-6 Astra and GPT-6.1 Sol.
One day later, Artificial Analysis reported a change in Gemini 4 Carbon is the new #1. Google’s model reaches 59 points, mainly thanks to strong results in complex reasoning and scientific tasks. In programming and knowledge it partly trails other models. In the AA-Omniscience v1.1 measurement, which also captures hallucinations, it leads with 54 points and shows a low hallucination rate. The caveat: Gemini 4 Carbon is so far open only to selected users, and an introductory price with a 50 percent discount applies. The note about lower cost per task compared with Opus 5.5 may therefore apply only to this introductory phase. Whether the advantage remains afterward is not stated in the material.
How thin the lead is becomes clear in The Batch. On October 29, the newsletter still names Opus 5.5 with 58 and GPT-6 Astra with 57 points at the top, although Gemini had already moved past with 59. A leaderboard is thus a snapshot, and one has to watch the index version and date. A one-point gap is also hardly a reliable verdict on everyday usefulness. The same applies to Kimi K3.1 Pro. Latent Space reports on October 24, before version 5 appeared, independent evaluations with 55 points and fourth place. The model has open weights, around 2.8 trillion parameters and 104 billion active parameters. The high-speed variant produces output two to three times faster and costs twice as much. According to Latent Space, reactions were mixed because some criticized the economics and the presentation of the benchmarks. The report does not say which index version is behind the 55 points.
The cost side is at least as revealing. In its index announcement, Artificial Analysis names GPT-6.1 Terra the cheapest model. The article GPT-6.1 Terra pushes the cost efficiency frontier quantifies this: prices per token (the smallest billed unit of text) are half or less compared with its predecessor GPT-5.6. The hallucination rate halves while accuracy is higher. In Terminal-Bench 4.0, Terra gains about 20 points and reaches values similar to GPT-6.1 Sol. In the Coding Agent Index it scores 48 points, above GPT-6 Luna and below Sol. A model not being in first place thus does not rule it out for many tasks.
That the model price is not all that counts is shown by GitHub Copilot Fusion. The system distributes tasks across models from Anthropic and OpenAI to balance performance, cost and response time. According to Artificial Analysis, it reaches 71 points in Terminal-Bench 4.0 at $5.59 per task. Claude Code with Opus 5.5 costs $11.42. Fusion needs 7.2 million tokens per task, about half as many as Opus 5.5. In Coding Agent Index v2.0, it ranks third in the FrontierCode rating with 51 points, behind Claude Code and Devin Fusion. One thus gives up part of the top performance but saves about half the cost. Language also plays a role: The Batch points to the so-called tokenizer tax, under which some languages consume significantly more tokens for the same content and therefore cost more. In 5 Hebel, um KI-Kosten zu senken, Computerwoche describes the cost block as even more opaque than the cloud bills for which companies have built FinOps departments over the past ten years.
Beyond this week, what matters is that the question is shifting from “Which model is best?” to “What does a solved task cost?”. A top spot with a one-point lead and a time-limited introductory price is less useful for this than the per-task calculation that Artificial Analysis now tracks systematically.
The end of ERP and the office job? What AI replaces in companies
The sharpest forecast of the week comes from Emanuel Böminghaus, CEO of AvenDATA. According to the summary on the KI-Beratung blog, he considers it possible to replace SAP as a central ERP system with AI-supported software. ERP stands for Enterprise Resource Planning, the software in which a company manages finance, purchasing, production and personnel. His argument: many companies use only a fraction of the functionality, so a self-built system would not have to do everything. He sees a serious threat to SAP if retail chains such as Lidl or Aldi switch to in-house solutions. He frames this with a finding about the German economy: it is stagnating because of high taxes and expensive energy, and labor costs are well above those in China. His prediction fits this: call centers will be replaced by AI assistants, and in ten years only about a tenth of today’s office jobs will remain.
The vendors themselves tell the story differently. Computerwoche reports from SAP TechEd, where technology board member Phillip Herzig made the “Autonomous Enterprise” more concrete. This is the vision SAP formulated at the start of the year: rethinking every business process end to end with AI. Herzig drew a comparison to the change thirty years ago, when the entire technology stack shifted. At the center is “Joule Work”, which he described as a fundamentally new way of using software, along with five autonomous SAP domains. SAP thus describes the upheaval as something happening within its own house. Böminghaus, by contrast, speaks of the house possibly becoming superfluous. Oracle is also on the offensive: Computerwoche headlines its report from Oracle AI World with the forecast of the end of ERP. CEO Mike Sicilia told the audience of the conference, which draws around 25,000 people every year, that AI must give people more time to be human. How the end of ERP is supposed to look in concrete terms, the report leaves open in the part available here.
A practical counterpart to the rhetoric comes from The Decoder. With Nuclos 5, Novabit brings an AI agent called Nuclosian into its open-source low-code platform, an environment in which applications are mostly configured rather than programmed. The agent takes ERP requirements via chat and configures the application directly, five to ten times faster than before according to the manufacturer. The limitation is notable: all generated configurations remain manually editable and can also be used without AI. This is closer to Böminghaus’ build-it-yourself thesis than to SAP’s platform promise, but it does not replace the ERP; it speeds up its setup and keeps humans in the process. The speed figure comes from the manufacturer and has not been independently verified so far.
On office jobs, the picture is more sober when set against the podcast KI verstehen from Deutschlandfunk. It cites 57 percent of large companies in Germany already using AI, but only 3 percent making strong use of its potential. Anyone wanting to gauge the pace of the transformation must therefore first look at this gap. The consequences for employment are disputed: Nvidia CEO Jensen Huang stresses that AI creates jobs, but journalists reacted skeptically, also with a view to Microsoft’s announced cut of 6,000 positions. The podcast points to new work, for example in data centers, and to the widespread fear of job losses, and it notes that deployment must be done carefully.
How realistic, then, is the thesis of the end of ERP and the office job? The sources support a shift rather than a replacement. That systems can be configured via chat and that vendors are rebuilding their products around agents is documented. The figure of ten percent remaining office jobs, by contrast, is an estimate by a single company head, for which the material gives no derivation. And the vendors announcing the end have an interest in selling it as their own product. Beyond this week, the key question is who will own a company’s process logic in the future: the software house that supplies it, or the business that builds it itself with AI.
Agents in regular operation: leading, securing, building properly
The pilot with a handful of users usually runs smoothly. The real test begins in regular operation. Computerwoche describes this in its article KI-Agenten im Regelbetrieb: Governance darf nicht warten. When thousands of employees work with different devices, access to production systems exists and agents no longer just deliver answers but trigger orders or send emails themselves, it becomes clear whether responsibilities, permissions and data flows have been settled. Without governance, meaning a binding framework of responsibilities and rules, isolated AI islands emerge, according to the article. They are hard to control and drive up costs and compliance risks. The question of who leads, steers and is liable thus comes at the beginning of regular operation, not at its end.
On leadership, Christoph Magnussen starts from his video KI braucht Führung: Warum du nicht jeden Schritt vorgeben solltest. His observation: many treat AI like a conventional computer and give instructions step by step. Treating it like a human employee would make more sense. You define the goal clearly and leave the solution path to the AI. Managers would have to learn to lead agents instead of overloading them with detailed specifications. This stands in tension with Computerwoche’s governance view. Magnussen argues for freedom in the how, while the governance perspective argues for hard limits on what an agent may access. The two can be read together: the goal is set openly, while the space for action remains technically limited. This is, however, an interpretation that arises from the two sources. Neither of them formulates it this way.
There is also a clear recommendation on how to build. According to The Decoder, Anthropic advises companies to abandon multi-agent architectures. This refers to setups in which many specialized subagents take over subtasks. Instead, a single main agent should work, loading context only when needed. In one customer case, speed reportedly doubled; for the conversion, Anthropic provides a migration kit on GitHub. This is a single case and not a general series of measurements. A second idea comes from the Latent Space podcast How Anthropic ships Claude with Claude. It deals with separating an agent’s brain from its hands. The logic and the execution run in different environments, which is meant to serve security and efficiency. The episode also stresses the value of evaluations, meaning systematic tests with which performance is continuously checked. Few agents and a clean dividing line between thinking and acting thus follow the same principle: fewer moving parts that need to be secured.
The vendors are responding with management tools. At the ServiceNow AI Summit for the DACH region in Munich, the focus, according to IT-Business, was the transition from the experimentation phase to productive workflows. ServiceNow wants to help companies capture, monitor and control their agents and applications. IBM presented a whole bundle of tools at TechXchange, as Computerwoche reports in So will IBM den KI-Agenten-Wildwuchs bändigen. The article refers to a Gartner analysis according to which an average Fortune 500 company will have more than 150,000 AI agents in use by 2028. As a weak point it names, among others, persistent credentials, meaning static, long-lived or over-privileged access. Where so many agents run, the first need is an overview of which agents exist at all and what rights they have.
How early the market still is, Gartner analyst Annette Zimmermann says in a conversation with Computerwoche published under the title KI-Agenten – was Gartner CIOs empfiehlt. In her assessment, companies are still experimenting a lot and in part cautiously probing what is possible. The conversation also covers how to keep costs and control in check. Beyond this week, what matters is that leadership, architecture and permissions must be decided together rather than one after another. Anyone giving agents a goal also needs an inventory, limited rights and tests before the pilot becomes an everyday tool.
Whose answers law firms and medical practices can believe
Anyone working in a profession with high liability asks less about speed than about reliability when it comes to AI. This tension runs through the week’s material. On one side are products meant to shorten routine work in law firms and practices. According to Anthropic erweitert Claude for Legal um vorgefertigte Vorlagen für Kanzleien at The Decoder, Anthropic has added interactive overviews and tables to its legal offering. They are meant to make setup easier for legal teams and speed up routine tasks such as contract reviews. In addition, more than 60 interfaces allow firms to connect their own databases directly. The emphasis is thus less on a smarter model than on how quickly a team becomes operational.
The podcast Claude for Legal und Gemini 4 Carbon sind da! by “Prompt mich mal” puts this in a larger context. AI is evolving from a pure text generator into a problem solver that analyzes legal documents and generates answers independently. As an example, the podcast names the Model Context Protocol (MCP), a standard through which language models access external data sources and tools. This would allow the AI to work locally in law firm databases without endangering sensitive data. That is the show’s assessment. The material lacks evidence of how this protection is implemented technically, and that is exactly what a law firm would need to clarify before use. For professionals bound by confidentiality, the usefulness of a tool is often decided by where the data flows, even before the quality of the answer comes into play.
At the same time, the choice is becoming less clear. In its overview Die 25 beliebtesten KI-Tools für Anwälte und Kanzleien, the FAZ puts the number of legal AI offerings at 1,400, twice as many as a year ago. Microsoft, OpenAI, Google and Anthropic have now joined in as well. For law firms this means: alongside specialized providers, the large platforms are pushing into a market in which many small providers have so far competed for trust. Buyers must therefore check whether a tool fits their own way of working and cannot rely on name recognition alone.
On trust in the results, the week helps only partly. On Hacker News, the article Dayjob: Tax was discussed, which measures AI agents on 50 professional tax tasks. It received 198 points and 124 comments. The material at hand, however, contains only the title and the response, not the results. Which tasks the agents handled and where they failed cannot be said here. The design is interesting: testing is done on realistic professional tasks rather than exam questions, and this kind of test is what a specialist department needs when it wants to assess vendor promises.
The question of when one may believe the machine is addressed rather generally by Manager Magazin in Wann Teams KI-Antworten wirklich trauen können. AI delivers answers in seconds and often sounds remarkably certain while doing so. Teams must therefore find out when they can trust the technology and when their own judgment counts. The point hits the core: the danger lies less in the obvious error than in the confident tone that makes a check seem unnecessary. Unlike the podcast, which mainly stresses efficiency, Manager Magazin shifts the responsibility to how the team is organized.
The counterpart to legal analysis is medicine, where the focus is documentation. In Diese Firma entlastet Ärzte mit KI bei der Dokumentation, the Handelsblatt reports that more than 10,000 healthcare facilities use the AI assistant from Tandem Health. Deutsche Telekom is now taking a stake, and company head Lukas Saari wants to build a European champion. The promise is to give doctors back a great deal of time. Documentation is a field in which a human can check the result before it counts. That distinguishes it from a diagnosis, where the judgment itself is at stake.
Beyond this week, what matters is that in the professions the question is shifting from “Can the AI do this?” to “Who checks what it delivers?”. The deeper tools are embedded in law firm databases and practice workflows, the more important become traceable tests on real professional tasks and a clear division of labor between machine and professional.
The per-task calculation is likely to last the longest. The halved introductory price of Gemini 4 Carbon and Anthropic’s adjusted margin of around ten percent will show what the systems really cost only in continuous operation. On security, it is worth watching whether patches arrive in days instead of weeks and whether access to Gemini 4 Carbon opens beyond the defenders as announced. Anyone introducing agents should clarify whether inventory, permissions and tests are in place before the pilot becomes an everyday tool. Leaderboards change within days; these questions remain.
Listen
Transcript of this episode
Female voice: A note up front: this review was produced with artificial intelligence and spoken by synthetic voices. Welcome to the weekly review. This time one pattern runs through all the topics: the benefit isn’t determined by the model alone, but by everything around it. That means updates, cost per task, agents’ permissions, and the question of who does the checking in the end. We begin with Black October.
Male voice: The name comes from a flaw in PHP. According to the KI-Beratung blog, it let anyone run foreign code without a password. It was a zero-day vulnerability, so when it became public there was no fix yet. It was reportedly discovered by the model GPT-5.6 Sol. Charité and the Bundeskriminalamt took systems offline as a precaution.
Female voice: Sounds like a total outage. Was it?
Male voice: Everlast AI takes a more sober view: around seventy percent of the internet wasn’t affected at all, and a lot was only switched off out of caution. KI-Beratung highlights a gap instead. The US government knew five days earlier, and according to the blog, Europe was barely warned.
Female voice: A five-day head start is huge. Were there real attacks, or did it stay with the vulnerabilities?
Male voice: Real ones. According to KI-Beratung, a Chinese hacker group used AI agents and stole the login credentials of more than two thousand two hundred fifty company accounts. So these systems don’t just find vulnerabilities, they also carry out the attack.
Female voice: The most tangible example for open models comes from Epoch AI. Open means anyone can download the weights. GLM 5.3 exploited vulnerabilities in six current Firefox versions. The trick is the patch gap, the time between a security update and its arrival at the user. The attack read out files and sent them away without anyone having to click anything.
Male voice: What does something like that cost the attacker?
Female voice: Around one thousand dollars per Firefox version. The researchers also found an unpublished vulnerability and combined it with the known ones. Mozilla has since shipped an update. Epoch reads this as a growing threat from open models.
Male voice: On the other side there’s the newsletter Interconnects. The author talks about Black October, names PHP2Shell and GitLab, and thinks the security scene overreacted. Open models, he says, had no catastrophic consequences. He sees the bigger risk in closed models and in the cyber industry itself.
Female voice: Wait, the Firefox case shows how low the entry barrier is. Almost anyone can come up with a thousand dollars.
Male voice: True. But the attack only works at all because updates arrive late. That’s a problem of distribution and patch discipline, not of the model alone. The two are arguing less about facts than about how to read them.
Female voice: And the defenders are arming up: Google DeepMind reports eighty-eight point two percent for Gemini 4 Carbon in vulnerability detection.
Male voice: Who’s allowed to use it?
Female voice: At first only trusted defenders in the Fairwind program, with a broader opening announced. For exploiting vulnerabilities, by the way, it’s at fifty-four point eight percent. So the provider decides who gets models with offensive capability.
Male voice: For companies that means in practice: if you run your own AI applications, penetration tests are part of the job, meaning controlled attacks on your own system. The Computerwoche reminds us that generative AI has long been drafting code and acting on its own through tools. What matters is whether patches arrive in days or only in weeks.
Female voice: Now to Anthropic. The IPO has been filed, the S-1 document is available, that is, the mandatory disclosure before listing. Ticker on the Nasdaq: ANT. Revenue in the first half of twenty twenty-six was sixteen point three billion dollars, after around one point two billion in the same period a year earlier.
Male voice: Heise calculates that as roughly fourteen-fold growth. Latent Space cites an increase of one thousand three hundred sixty-eight percent for the second quarter alone, and a valuation target of two trillion dollars.
Female voice: But then there’s a loss of eighty point seven billion. How does that fit together?
Male voice: According to Latent Space, the loss comes mainly from items where no cash goes out. In the second quarter there was even an operating profit of nine hundred seventy-four million. Actual cash burned so far is about nine point seven billion, against a cash balance of one hundred thirty-four billion.
Female voice: And what’s behind that?
Male voice: Compute costs. The Batch shows: in twenty twenty-four, spending on compute was still six and a half times revenue, in the first half of twenty twenty-six only about half. That yields an adjusted operating margin of around ten percent. So cost per revenue dollar says more than the book loss.
Female voice: Does Anthropic earn its money with the products or with the model?
Male voice: Both, depending on your angle. The Batch attributes the growth to agentic products, Claude Code and Claude Cowork, with around nine hundred percent customer retention among business customers. What exactly that metric measures isn’t in the material. Heise says: about eighty percent of revenue is usage fees for the models. The products drive consumption, but what’s billed is essentially the model.
Female voice: And according to Latent Space, forty-eight percent of revenue runs through cloud partners. That’s almost half through other companies’ platforms.
Male voice: Does the ten percent margin hold if that share stays this high?
Female voice: Only the first quarterly reports as a listed company will show that. The direction is interesting too: the filing states the goal of reaching more than one billion knowledge workers with virtual employees by twenty twenty-eight. Claude Duet fits that, a new agent for professionals and small businesses.
Male voice: And Heise has a report that is, however, unconfirmed. Anthropic is said to be consulting law firms on how agents could negotiate prices, salaries and contracts for their users.
Female voice: Agents that negotiate bindingly instead of just working. Who is liable, actually, if such an agent negotiates itself into a corner?
Male voice: The report says nothing on that. But it’s exactly the question that comes back later in the law firm segment.
Female voice: Now to the ranking. On October twenty-seventh, Artificial Analysis presented the fifth version of its Intelligence Index. New are FrontierCode 2.0 with real programming tasks and Terminal-Bench-Science with seventy research tasks. In addition, sixty percent of the tasks come from private datasets that the models aren’t supposed to know beforehand.
Male voice: Who led afterwards?
Female voice: Claude Opus 5.5, followed by GPT-6 Astra and GPT-6.1 Sol. One day later the top spot changed: Gemini 4 Carbon with fifty-nine points, strong in reasoning and science tasks, weaker in programming and knowledge. On the hallucination measure AA-Omniscience it leads with fifty-four.
Male voice: And the catch? With a jump like that there’s always one.
Female voice: Two, in fact. Only selected users have access, and there’s an introductory price with a fifty percent discount. That it’s cheaper per task than Opus 5.5 may only hold during this phase.
Male voice: You can see how thin the lead is at The Batch. On October twenty-ninth, they still had Opus 5.5 with fifty-eight and Astra with fifty-seven points in front, although Gemini had long since passed them. A leaderboard is a snapshot, and index version and date matter. And a one-point gap says little about everyday use.
Female voice: And Kimi K3.1 Pro? The fifty-five points are hard to place.
Male voice: Exactly. Latent Space reported that on October twenty-fourth, before version five, and which index version is behind it remains unclear. The model has open weights, around two point eight trillion parameters, of which one hundred four billion are active. The high-speed variant is two to three times faster and costs double. Reactions were mixed, over economics and how the benchmarks were presented.
Female voice: I find the cost side more interesting. GPT-6.1 Terra is at half or less of GPT-5.6’s price per token, the hallucination rate is halved, and in Terminal-Bench 4.0 it gains around twenty points. A token is the smallest billed unit of text.
Male voice: So is second or third place worth it too?
Female voice: Often, yes. In the Coding Agent Index, Terra reaches forty-eight points, above GPT-6 Luna, below Sol. Not being first doesn’t rule a model out for many tasks.
Male voice: It gets even clearer with GitHub Copilot Fusion. The system distributes tasks across models from Anthropic and OpenAI. Seventy-one points in Terminal-Bench 4.0, at five dollars fifty-nine per task. Claude Code with Opus 5.5 costs eleven dollars forty-two. Fusion needs seven point two million tokens, about half as many.
Female voice: And what do you pay for that?
Male voice: In Coding Agent Index v2.0, Fusion sits in third place in the FrontierCode ranking with fifty-one points, behind Claude Code and Devin Fusion. You give up part of the peak performance and save roughly half.
Female voice: Language adds to that. The Batch talks about the tokenizer tax: some languages use significantly more tokens for the same content. You notice it on the bill if you work a lot in such a language. The Computerwoche even considers the cost block more opaque than the cloud bills that companies built FinOps departments for.
Male voice: So the question is no longer which model is the best, but what a solved task costs. And that is exactly the calculation Artificial Analysis now includes systematically.
Female voice: The sharpest prediction of the week comes from Emanuel Böminghaus, the head of AvenDATA. According to KI-Beratung, he considers it possible to replace SAP as the central ERP system with AI software. ERP stands for Enterprise Resource Planning, the software for finance, purchasing, production and HR.
Male voice: His argument: many companies use only a fraction of the functionality, and a custom build wouldn’t have to do everything. It would get dangerous for SAP if retail chains like Lidl or Aldi switch. He also predicts that call centers will be replaced by AI assistants, and that in ten years only a tenth of office jobs will remain.
Female voice: A tenth? Where does that number come from?
Male voice: Not from the material, at any rate. It’s one company head’s estimate, with no derivation. He frames it with a stagnating German economy, high taxes, expensive energy, and labor costs far above those in China.
Female voice: SAP tells it differently. At TechEd, board member for technology Phillip Herzig fleshed out the Autonomous Enterprise, meaning rethinking every business process end to end with AI. At the center is Joule Work, plus five autonomous domains. The upheaval happens inside the house, while Böminghaus sees the whole house as superfluous.
Male voice: And Oracle?
Female voice: The Computerwoche headlines it there as the end of ERP. CEO Mike Sicilia told around twenty-five thousand conference guests that AI should give people time to be human. What the end looks like in concrete terms stays open in the report.
Male voice: More concrete is Nuclos 5 from Novabit, reported by The Decoder. It’s an open-source low-code platform where you configure applications rather than program them. The agent Nuclosian takes ERP requests via chat and configures directly, five to ten times faster according to the vendor.
Female voice: Independently verified?
Male voice: No, that’s a vendor claim. And everything generated stays manually editable, usable even without AI. That doesn’t replace the ERP, it speeds up setup and keeps people in the process.
Female voice: On office jobs, the Deutschlandfunk podcast KI verstehen is more sober: fifty-seven percent of large companies in Germany use AI, but only three percent exploit the potential heavily. Nvidia CEO Jensen Huang says AI creates jobs, journalists remain skeptical, partly because of Microsoft’s announced cut of six thousand positions.
Male voice: Is the gap between fifty-seven and three percent then the real measure of the pace?
Female voice: That’s how I read it. The sources support a shift rather than a replacement. What’s documented is that systems can be configured via chat and that vendors are rebuilding their products around agents. And whoever proclaims the end usually wants to sell it as their own product. What remains open is who will own the process logic in the future: the software house or the company that builds it itself.
Male voice: On to regular operation. The pilot with a few users usually runs smoothly, says the Computerwoche. The real test comes when thousands work with different devices and agents trigger orders or send emails themselves.
Female voice: What does governance mean here, concretely?
Male voice: A binding framework of responsibilities and rules. Without it you get isolated AI islands, hard to steer, with rising costs and compliance risks. Who leads, steers and is liable has to be settled at the start.
Female voice: Christoph Magnussen takes the opposite approach: many treat AI like a classic computer and prescribe every step. Better to treat it like an employee. Define the goal clearly, leave the path open.
Male voice: That clashes with the governance view. Freedom in the how here, hard limits on access there. Can the two go together?
Female voice: You can read them together: set the goal openly, limit the scope of action technically. But that’s our interpretation, neither source says it that way.
Male voice: On architecture, according to The Decoder, Anthropic advises against multi-agent architectures in which many subagents take on partial tasks. Better is a main agent that loads context only when needed. At one customer, speed doubled, and a migration kit is on GitHub.
Female voice: Is one customer case enough as evidence?
Male voice: No, that’s a single case. The Latent Space podcast fits here: separate the agent’s brain and hands, logic and execution in different environments, plus ongoing evaluations, meaning systematic tests. Few agents, a clear dividing line: fewer moving parts you have to secure.
Female voice: Vendors are adding management on top. At the AI Summit in Munich, ServiceNow wants to register, monitor and control agents. At TechXchange, IBM showed a whole bundle of tools and points to Gartner: by twenty twenty-eight, an average Fortune 500 company will have more than one hundred fifty thousand agents.
Male voice: And where does it get stuck?
Female voice: With persistent credentials, meaning long-lived or over-privileged access. First you need an inventory: which agents exist at all, and what permissions do they have?
Male voice: Gartner analyst Annette Zimmermann says the market is still early, with many experimenting and testing the waters carefully. So leadership, architecture and permissions have to be decided together, not one after another.
Female voice: Last block: law firm and medical practice, where reliability counts more than speed. Anthropic has extended Claude for Legal with interactive overviews and tables. They’re meant to ease the start for legal teams and speed up contract reviews. Through more than sixty interfaces, you can connect your own databases.
Male voice: So less a smarter model than a faster start?
Female voice: That’s how I see it. The podcast Prompt mich mal goes further: AI is moving from text generator to problem solver. As an example it names the Model Context Protocol, a standard for access to data sources and tools. With it, AI could work locally in law firm databases without putting sensitive data at risk.
Male voice: How is that protection implemented technically?
Female voice: The material doesn’t say, and a law firm would have to clarify that before use. For professionals bound by confidentiality, the question of where the data flows often counts for more than the quality of the answer.
Male voice: On top of that, the choice is getting harder to survey. The FAZ counts one thousand four hundred legal AI offerings, twice as many as a year ago. Microsoft, OpenAI, Google and Anthropic are getting in, and the small vendors that so far won customers on trust face powerful competition.
Female voice: Then name recognition isn’t enough as a criterion. How do you check whether a tool is any good?
Male voice: With tests on real professional tasks. On Hacker News there was Dayjob: Tax, fifty professional tax tasks for AI agents, one hundred ninety-eight points, one hundred twenty-four comments. Results aren’t in the material, so we don’t know where the agents failed. What’s convincing is the design: professional tasks instead of exam questions.
Female voice: And in everyday work, the Manager Magazin warns about the self-assured tone. AI answers in seconds and sounds sure while doing it. Teams need to know when to trust it and when their own judgment applies. The danger is less the obvious mistake than a tone that makes checking seem unnecessary. The responsibility here lies with the team’s organization.
Male voice: Where is it easier to check?
Female voice: In medical documentation. According to the Handelsblatt, more than ten thousand healthcare institutions use the assistant from Tandem Health, Deutsche Telekom is coming in, and CEO Lukas Saari wants to build a European champion. A person can check the result before it counts. With a diagnosis, the judgment itself is at stake.
Male voice: So the question shifts from what AI can do to checking its results. For the coming week, three things are worth watching: whether patches arrive in days instead of weeks, whether Gemini 4 Carbon opens up beyond the defenders, and what the cost per task shows after the introductory price discount ends. And if you’re introducing agents, settle inventory, permissions and tests first, before the pilot becomes an everyday tool.
Female voice: That was the AI week. If you don’t just want to listen but build something yourself: on foundic.org, under KI-Projekte, there are step-by-step guides, including on how this podcast is built.
Male voice: We’ll be back next Saturday. Until then.