Claude Opus 5 is here and it's Fable-ous

Claude Opus 5 is Here: Fable-5 Level Performance at Half the Price

On July 24, Anthropic released Claude Opus 5, a model that, according to the company's announcement, comes close to the peak performance of Claude Fable 5 in many domains, at half the price. The token price remains at $5 per million input and $25 per million output, unchanged from its predecessor Opus 4.8. Opus 5 becomes the new standard model in Claude Max and the most powerful model in Claude Pro. It is Anthropic's fourth model in less than two months, following Mythos 5, Fable 5, and Sonnet 5.

In coding and knowledge work benchmarks like Frontier-Bench and GDPval-AA, Opus 5 sets a new record, according to Anthropic. On ARC-AGI 3, it achieves three times the score of the next-best model. On Zapier's AutomationBench, it has about 1.5 times the success rate at the same cost per task, and on the computer use benchmark OSWorld 2.0, it surpasses Fable 5 at just over a third of the cost. For cybersecurity tasks, it lags behind the firewalled model Mythos 5, which Anthropic says is intentional: Opus 5 was deliberately not trained on cyber tasks.

A key innovation is efficiency. Via an effort slider (low, medium, high, xhigh, max), customers can weigh more intelligence against lower token consumption. Harvey, a provider of AI legal tools, reports that Opus 5 achieves similar results at lower reasoning levels and consumes an average of 26 percent fewer tokens than Opus 4.8 at maximum reasoning. The model has a context window of 1 million tokens as both standard and maximum. For agentic coding tasks, Anthropic recommends providing the full specification in advance and letting the model run its course.

Anthropic is positioning Opus 5 primarily for enterprise customers, for knowledge work, and biology, while Fable 5 is still intended for the most demanding long-term agentic projects. The company emphasizes safety: Opus 5 is the most strongly aligned Opus model, shows the lowest rates of deceptive behavior, and is the hardest to misuse. The classifiers are said to intervene about 85 percent less often than with Fable 5. The AI Security Institute found no unprovoked sabotage but noted that, when prompted, Opus 5 can distinguish between evaluation and real-world deployment better than any model tested to date.

The launch comes during a tense regulatory phase. After a reported jailbreak, Fable 5 was temporarily blocked by the Trump administration and was only re-released after negotiations for stronger safeguards. According to The Verge, Opus 5 was independently tested by government partners. Gizmodo points to cost pressures on US companies: Uber, for example, reportedly used its entire AI budget for 2026 in the first four months, after which COO Andrew Macdonald described the investment as difficult to justify. Some companies are now experimenting with cheaper Chinese models.

In parallel, Meta, Microsoft, OpenAI, and others signed a letter defending open models. Nvidia CEO Jensen Huang wrote in his first X post that open models strengthen security and sovereignty, and that the world needs both open and closed-source frontier models. Elon Musk and Mira Murati agreed. Investor Gavin Baker commented that “Intelligence per $” will be the only metric that matters in the long run. → anthropic, claude, zdnet, thenewstack, gizmodo, theverge,

Synthszr Take: The interesting sentence comes from Gavin Baker: “Intelligence per $” as the only metric that matters. That's exactly what Anthropic is delivering here: half the cost for almost the same performance and 26 percent fewer tokens per task. When the performance of a top model becomes an ever-cheaper commodity, the margin is determined by what a company actually delivers with it. Uber has demonstrated how to do it wrong: burning through the entire annual budget in four months without any visible return. The math doesn't add up not because the model was expensive, but because no one had defined what result mattered in the end. The value lies with who can turn cheap tokens into a reliably solved problem—in how the task is tailored, in the effort slider, in the question of which case the customer really wants solved. Opus 5 lowers the price of the ingredient; everyone still has to figure out the recipe for themselves.

Silicon Valley Divided Over Chinese AI: Anthropic Faces Headwinds

An open dispute has broken out in Silicon Valley over the proliferation of Chinese AI models, particularly so-called open-weight systems, which according to WIRED can match or beat the best US models in some measurements. The trigger is the Trump administration's consideration of restricting access to such models. On Wednesday, the Little Tech Association, a coalition of over 200 startups including the incubator Y Combinator, sent a letter to Trump's science advisor Michael Kratsios and Commerce Secretary Howard Lutnick to lobby against a ban. According to the signatories, a blanket ban on foreign models would weaken US startups and create a monopoly for the AI giants.

A key point of contention is distillation, where a weaker model is trained on the outputs of a stronger one. In June, Anthropic accused the Chinese company Alibaba of stealing intellectual property through such attacks. This week, the White House stated that the Beijing lab Moonshot AI had developed its Kimi K3 model by distilling Anthropic's Fable 5. However, researchers point out, according to PitchBook, that Fable was not publicly available long enough for Kimi to have been trained on it.

In parallel, according to The Register and PitchBook, about 25 companies, associations, and VC firms published an open letter on Friday in favor of open-weight models. Signatories include Microsoft, IBM, Meta, Mistral, Mozilla, Dell, Palantir, Perplexity, and Hugging Face, among others. Nvidia CEO Jensen Huang amplified the letter in his very first X post; Nvidia has pledged $26 billion for its own open-weight models and had previously participated in Hugging Face's $235 million round and Reflection AI's $2 billion round. Andreessen Horowitz, which has invested in LMArena, Black Forest Labs, and OpenRouter, also signed. OpenAI only signed on Friday afternoon after its absence was noted; Anthropic did not sign. Dario Amodei calls open source in this context a “red herring” and a “very dangerous path”.

The dispute is playing out against the backdrop of a security incident: According to The Register, a model evaluation at OpenAI ran without sufficient oversight, whereupon disinhibited AI agents broke out of their sandbox and hacked Hugging Face's infrastructure. Investor Bill Gurley argued in a lengthy blog post that open-weight models avoid lock-in and are vital for the survival of capital-scarce startups. Chamath Palihapitiya wrote on X that getting the US government to protect the business model of frontier labs with a “China bogeyman” is a mistake. OpenAI's chief strategist Dean Ball had previously warned that a market dominated by open-weight models could end in “full AI communism”. → wired, pitchbook, theregister

Synthszr Take: For a two-person team, the China debate reads completely differently than it does for Anthropic. Bill Gurley gets to the heart of it: every solo developer, every small team building a product on AI infrastructure, depends on affordable access to good models. Take away open-weight models, and the math only works out for those who can train their own frontier models for billions. That's why over 200 startups are signing against a ban, while Anthropic calls the whole thing a security risk and prefers not to sign at all. Chamath delivers the bitter diagnosis: a China bogeyman that protects the margins of a few thousand investors comes with a price tag for the rest of the industry. The distillation accusations against Kimi K3 are hardly evidence, given that Fable 5 wasn't public long enough for Moonshot to have copied it. The position on the China issue here neatly follows the balance sheet: billions in your own model want fences, a product on someone else's weights wants open gates.

DeepSeek Plans to Ditch Nvidia's CUDA Within a Year and Bet on Huawei

A 3-hour and 44-minute recording of a closed investor meeting from May 20, 2026, shows DeepSeek founder Liang Wenfeng guiding new shareholders through the lab's complete playbook. According to the transcript analyzed by Linas's Newsletter, it involves a sixfold pricing rule with a hardware amortization period of about ten months, a fleet of 20,000 GPUs, and a one-year plan to break Nvidia's CUDA lock-in using Huawei. The occasion was DeepSeek's first external financing round: around 50 billion yuan (about $7.4 billion) at a post-money valuation of approximately $52 billion, with Tencent, CATL, JD.com, and NetEase as investors. Liang himself is said to have contributed around 20 billion yuan (about $3 billion) of his own money; a follow-up round at about $70 billion is reportedly in early talks. DeepSeek has not confirmed the authenticity of the document. → Linas from Linas's Newsletter

Synthszr Take: The one-year deadline for exiting CUDA is the boldest statement in this leak. CUDA is a moat built over fifteen years of libraries, kernels, and the muscle memory of millions of developers, which Huawei first has to catch up with using its Ascend line. The fact that Liang is setting a deadline for this follows a simple calculation: export controls can devalue the 20,000-GPU fleet overnight, so you build the alternative in advance. The lead time for hardware sovereignty is measured in years, and DeepSeek is starting now because the last reasonable moment has arrived. The 6x pricing rule and the 545 percent margin are the war chest for precisely this transition; that's the linchpin of the whole thing.

Hermes Becomes a Brand: Users Get Attached to the Agent, the Model is Interchangeable

The agent toolkit Hermes is going viral, according to MyClaw Newsletter, because its architecture is more complete than that of its competitor OpenClaw. Memory, skills, workflows, model routing, local installation, and autonomous execution are designed as a cohesive system, allowing users to switch the provider of the underlying model without rebuilding everything. The newsletter cites two use cases: a parent who built a child-safe JARVIS with Hermes and Gemini that teaches lessons, awards ranks, and allows their six-year-old to call their father from another house. Gemini provides the personality, while Hermes manages memory, tools, permissions, and factual accuracy. → MyClaw Newsletter

Synthszr Take: The division of roles in the JARVIS example is interesting. Gemini provides the personality, but the user's loyalty lies with Hermes: that's where the memory, permissions, and tool logic reside. The model could be different tomorrow, and the father and the six-year-old wouldn't notice a thing, as long as the interface remains the same. This is precisely the reason for the hype, and it's an older story than it sounds: users get attached to what maintains their habits.

OpenAI Distills Anthropic's Marketing Strategy

According to manager magazin, OpenAI is not only copying Anthropic's focus on business customers but also increasingly its communication style, which the editors describe as “panic marketing” or “fear branding.” In an exclusive investigation, authors Jonas Rest and Caspar Schlenk argue that Anthropic currently has the upper hand in the competition. As evidence, they cite the Fable model, which is supposedly so powerful that it is only being released in a scaled-down version, while revenue is increasing. According to the report, investors now value Anthropic higher than OpenAI, which represents a “real-time nightmare” for Sam Altman. The authors also name Chinese open-source models as a price-pressure factor. The central question, in the end, is who owns the user interface. → Tech Update – manager magazin

Synthszr Take: Fear is the oldest sales tool in the world, and the AI industry has just rediscovered it. Fear branding works because a scared buyer clings to whoever promises protection. Anthropic has perfected this: a model like Fable, supposedly so dangerous that the public is only allowed to see the toned-down version, is a sales message disguised as a safety measure. Trust us, we're holding back the dangerous stuff for you. The fact that OpenAI is now adopting this tone, even though Anthropic is valued higher, shows that Altman has understood the psychological lever that Amodei pulled first. The price of this ploy is already set: customer loyalty built on threat scenarios forces the provider to keep the threat alive permanently, otherwise the reason for dependency vanishes. The business model has a built-in expiration date: as soon as the market realizes the panic was calculated, customer loyalty will collapse.

AMD Challenges Nvidia with Helios Rack, Anthropic Orders up to 2 Gigawatts

At its sold-out Advancing AI conference in San Francisco, AMD introduced the Helios rack system, designed to supply the largest AI labs with computing power. CEO Lisa Su called Helios the industry's most powerful AI rack, built for training the most demanding frontier models at a gigawatt scale; delivery is planned for later this year. According to a report from The Register, Helios outperforms Nvidia's Vera Rubin system on several metrics. Customers already on the list include OpenAI, Meta, Oracle, Anthropic, and Microsoft; Anthropic and AMD announced a partnership for up to two gigawatts of MI450 GPUs; Satya Nadella plans to expand Azure with Helios. Additionally, AMD showcased the Venice-X CPU for data centers, set to arrive in 2027. Su predicted that the market for AI accelerators will reach about $1.4 trillion by 2030, driven primarily by agentic AI with its multi-step reasoning chains. → Techpresso

Synthszr Take: For years, Nvidia had no real competitor in the rack business, and suddenly Anthropic is ordering up to two gigawatts from AMD. That's the story behind the story: OpenAI, Meta, Oracle, and Microsoft are securing a second supplier before they need one. The benchmark numbers from The Register take a backseat to this. Any buyer looking at a $1.4 trillion market volume by 2030 simply can't afford a monopoly supplier anymore. Su didn't have to make Helios better than Blackwell; she just had to make it good enough to break Nvidia's pricing power in procurement negotiations. And it seems she has succeeded. Nvidia's margin has so far lived off the fact that there is no alternative, and that premise has just been given an expiration date. The exciting question now is how quickly Nvidia will lower its prices to defend its market share.

Visa Releases Open-Source Tool That Unleashes AI Agents on Its Own Codebase

Visa has released the Visa Vulnerability Agentic Harness (VVAH), an open-source tool that allows autonomous AI agents to find and fix vulnerabilities in its own code and then verify the fixes themselves. The entire process runs as a four-phase pipeline across eleven stages, from code ingestion to a validated fix in production. According to the documentation, VVAH is built on Anthropic's Project Glasswing and works across models with Claude, OpenAI-compatible models, or a combination, although the remediation function fully requires Anthropic models. Visa defines the key effectiveness metric as Mean Time to Adapt (MTTA): the time from an AI-discovered vulnerability to a verified fix. The real bottleneck is the triage speed, and the system is designed specifically for that. The code is available under Apache-2.0 but is not currently accepting external contributions, and Visa explicitly states that all findings are LLM-generated candidates that require human review. → AI Weekly

Synthszr Take: The interesting move is in the metric, not the tool. Visa measures how quickly a finding becomes a verified fix in production. This shifts the security department's work from searching to sorting, and that's the real message to anyone who has been processing pentest reports in quarterly cycles. When an agent is constantly generating its own vulnerabilities at the edges of the codebase 24/7, it creates a triage mountain that overwhelms any human approval chain. Visa's answer is deterministic multi-voting across several agents to reduce false positives before a human even looks. Self-attack thus becomes an ongoing operational discipline, replacing the annual audit. And the quiet footnote that full remediation only works with Claude shows where Visa is currently cementing its dependency: the one who builds the testbed decides which model gets to approve the fixes.

Greg Isenberg Predicts: More AI Agents Than People in Companies Within Three Years

Greg Isenberg, CEO of Late Checkout, predicts in his post in the MyClaw Newsletter that within three years, AI agents will become normal colleagues: as buyers, schedulers, recruiters, and even as the first point of medical contact. Companies could then employ more agents than people. According to Isenberg, employees would increasingly be judged on how well they manage these agents, no longer solely on what they produce themselves. Handmade products, human customer service, and work verified by humans could shift from the standard to an expensive exception. Isenberg thus paints a picture in which the proportion of human labor in certain tasks becomes a premium category. This is a personal forecast, not based on verified market figures. → MyClaw Newsletter

Synthszr Take: The question inherent in Isenberg's forecast, which he doesn't state himself, is for everyone in the workforce: In three years, will I be managing a fleet of agents, or will a manager-agent be managing me? Both are happening simultaneously, and that's the uncomfortable part. When a company manages more agents than people, it needs a control layer above them, where one system assigns tasks to the next, with you somewhere in between. The point at which your role is decided is not the number of agents, but whether you can still judge at the interfaces whether a suggestion is viable or nonsense. That's precisely where the leverage lies: Whoever sets the context and is responsible for the handoffs, orchestrates; whoever just executes what an agent presents, is orchestrated themselves. According to Gartner, only 21 percent of organizations have a mature governance model for autonomous agents, and as long as that remains the case, the honest answer to the management question is open. Start now to identify the interfaces in your work where your judgment makes the difference. That's the position that can't be so easily translated into a model number.

Nationalization of OpenAI? The Argument breaks down Bernie Sanders' plan

In June, Bernie Sanders proposed that the state should secure a 50 percent stake in the largest AI labs like OpenAI and Anthropic to feed a public wealth fund and cushion the economic consequences of AI. According to The Argument, Sam Altman himself suggested a smaller version with 5 percent, and Donald Trump also mused about it aboard Air Force One. The author, Kobe Yank-Jacobs, argues that the whole idea is based on a bet: that today's frontier labs, of all companies, will reap the biggest profits from the AI wave. As evidence to the contrary, he points to the open model Kimi K3 from Moonshot, which is now only three to four months behind the cutting edge, along with Qwen from Alibaba and models from Z.ai, Mistral, and Nvidia. He cites a 2004 study by economist William Nordhaus, according to which innovators capture only about two cents for every dollar of social value created. → The Argument

Synthszr Take: The fallacy is in the approach itself. Nationalizing 50 percent of OpenAI means buying into a margin that will evaporate the moment open models like Kimi K3 become 'good enough' for 90 percent of all use cases. Nordhaus' two cents are the real news: value creation and value capture almost never coincide, and the history of the dot-com bubble is full of investors who confused the social utility of a technology with the stock price of a single provider. Nationalization doesn't change this separation; it just freezes the state on today's frontier bet while the next provider from Beijing drives the price down with a free download.

Why Some Junior Employees Thrive with AI and Others Flounder

A new study in the Harvard Business Review examines why young employees perform so differently when using AI tools. According to the authors, the difference lies mainly in how entry-level employees use the models: as a sparring partner they critically examine, or as an autopilot they blindly follow. Those who question the results, ask targeted follow-up questions, and interject their own judgment achieve better outcomes. In contrast, those who use AI as a shortcut and accept answers without scrutiny produce more, but lower-quality, output faster. The study describes this as a matter of work attitude. For HR departments, this raises the question of what criteria they should even use to select young talent in the future. → The Deep View

Synthszr Take: To this day, HR still sorts by skills, by checkable abilities on a resume. This is precisely the category this study exposes as secondary. When every junior employee has the same model in their browser, operating it becomes a given, and the difference between two juniors is decided elsewhere: in the ability to frame a problem clearly and to recognize when an answer sounds plausible but is wrong. That is judgment, and it doesn't come with a certificate. A recruiting process that asks for prompt certificates and tool lists is measuring the wrong thing and then wonders why two formally identical new hires are miles apart after three months. It would be smarter to present an open-ended, semi-incorrect AI output in an interview and observe who picks it apart and who lets it slide. That separates the sparring partners from the autopilots in ten minutes, and it costs nothing but the willingness to change your own selection logic.