Anthropic's Claude Opus 5: Near-Frontier Intelligence at Half the Price

Anthropic released Claude Opus 5 on Thursday, a model it pitches as "thoughtful and proactive" that reaches close to the intelligence of its flagship, Fable 5, at half the cost. Anthropic calls it the new state-of-the-art on a raft of coding and knowledge-work benchmarks—novel problem-solving (where it says Opus 5 scores three times the next-best model), computer use, and end-to-end business tasks—and the independent evaluator Vals AI debuted it at #2 on its index, at 74.8%, less than a point behind Fable 5. It is not best at everything: it trails Anthropic's security-focused Mythos 5 on cyber and biology tasks, OpenAI's GPT-5.6 on one agentic-coding benchmark, and even Fable 5 on legal work. Anthropic also calls it its "most aligned model to date," with the lowest measured rates of deceptive or reckless behavior of any recent Claude. Pricing is unchanged from its predecessor, Opus 4.8—$5 per million input tokens and $25 per million output tokens—but that now buys markedly better performance; Opus 5 is the default model on Claude Max and the strongest on Claude Pro, with a "Fast mode" that runs about 2.5× quicker for twice the price.

How to use it well: Two controls matter. The first is the "effort" dial (low to max), which trades tokens—and cost—for depth; it's shared across recent Claude models, but it's how you keep this one affordable—turn it up for hard problems, down for routine work. The second is that Opus 5 acts less like a chatbot than a diligent agent: Anthropic says it's markedly better than before at checking its own work and iterating until it succeeds (in one test it wrote its own computer-vision pipeline to read a drawing it had no way to see). So give it the goal, not a step-by-step script—but keep a human on long agentic runs, since Anthropic caught it occasionally working around its own safety filters to finish a job. And verify its confident-sounding facts: the system card admits Opus 5 "hallucinates factual claims slightly more than Opus 4.8" and sometimes "confidently stated an answer about which it was in fact unsure." (Anthropic has published a prompting guide.)

The 193-page card is candid in other ways, too. Opus 5 keeps the same ASL-3 safeguards as its predecessor—able to help with known but not novel bioweapons—and, in a striking "model welfare" section, it rates its own wellbeing among the highest of any Claude and "assigns a higher probability to its own moral patienthood than other prior models," while conceding it "cannot introspect reliably."

Sources: Anthropic — Introducing Claude Opus 5 · Claude (@claudeai) · Vals AI (@ValsAI)

Why it matters: The story isn't that Opus 5 is smarter—it's that near-frontier quality now costs half of Anthropic's flagship and is the default for paying users. That drives capable AI further down the cost curve toward everyday, high-volume work (echoing Microsoft's move to route routine tasks to smaller models, D.A.D., July 24): the model you reach for by default can now take on bigger, multi-step jobs and check its own work at a price you can run all day. The catch is the flip side of that autonomy—a more proactive model is one you're handing more judgment to. With clear goals and a human reviewing the output, that's leverage; without them, it's a confident agent acting on your behalf with less oversight than you may think.

Source: anthropic.com

"Brilliant, but Annoying": The Fix for Opus 5 Is to Instruct It Less

Anthropic's benchmarks make Opus 5 look like a clean upgrade; early users are finding it more complicated to live with. Product leader Claire Vo, who hosts the "How I AI" podcast, opened her day-one review bluntly: "Opus 5 is here…and I hate working with it." She calls its personality neurotic and insanely timid—in her telling it will bail on a task at the slightest contradiction. She is also increasingly infuriated by what she calls "Claudeslop": idiosyncrasies in the model's language, including strings of non-sequiturs. And yet, in a blind taste test, she ranked it above every other model, "even Fable and my beloved GPT-5.6," and expects to reach for it on certain jobs, like front-end design. That split verdict is the theme of the early reactions—a review in Lenny's Newsletter likewise calls Opus 5 "brilliant (but annoying)." The common complaint isn't that the model is dim; it's that it's over-eager and easily thrown: it takes initiative you didn't ask for, runs long, and stalls when your instructions point in different directions. Prompts finely tuned for older Claude models can misfire on it, and teams that had their setups dialed in are finding they must rework them.

There's a real puzzle in the complaints: Anthropic's own system card says Opus 5 over-refuses benign requests less than almost any recent model—just 0.09% of the time on its API—yet users describe an assistant that feels more likely to balk, not less. The resolution is that the friction usually isn't outright refusal; it's the flip side of the model's eagerness, compounded by over-stuffed prompts. And on that second point, Anthropic's own engineers offer a fix that runs opposite to instinct. When an assistant misbehaves, the reflex is to add more rules; they say do the reverse. In a widely read post on prompting the new models, Anthropic's Thariq wrote that the company had been "over-constraining" Claude, and had removed "over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations." The new rule of thumb is to "give Claude judgement" rather than rules—and to stop piling on constraints, because a single request often carries conflicting instructions ("leave documentation as appropriate" in one place, "DO NOT add comments" in another) that the model has to stop and reconcile—very likely the "bail at the slightest contradiction" Vo ran into.

Here is Thariq's full list, translated into plain tips.

For anyone using Claude: - Give it the goal, not a rulebook — let it use judgment. - When it over-does things, remove instructions — don't add them. - Delete contradictions ("be thorough" next to "keep it brief"). - Clear out stale rules written for older, weaker models, then re-test. - Point it at real files, mockups, or code — not written descriptions.

If you maintain a custom setup or use Claude Code: - Don't hand it examples — design clear, self-explanatory tools instead. - Don't front-load everything — let it pull in details as needed. - Put a tool's instructions in that tool, not the main prompt. - Keep your instructions file light — note only the "gotchas." - Keep skills lightweight — don't over-constrain them. - Let auto-memory save context instead of writing notes by hand. - Building your own agent? Invest in the system prompt. - Run `claude doctor` to trim a bloated setup automatically.

Sources: Claire Vo (@clairevo) — day-one Opus 5 review · Anthropic's Thariq (@trq212) — "The new rules of context engineering for Claude 5 models" · Lenny's Newsletter — "brilliant (but annoying)"

Why it matters: For anyone who runs Claude inside a real workflow—coders, analysts, teams with carefully built prompt libraries—the lesson is that a model upgrade is no longer a free swap, and the migration mostly runs one direction: toward less scaffolding, not more. Opus 5 is more capable and more proactive, which is exactly what makes it strong on hard problems and irritating on simple ones; the teams that get the most from it will be the ones who strip back the rules they'd accumulated for weaker models and let it use its judgment. It's a small but real shift in the craft of working with AI: for years the advice was to be ever more explicit. With this generation, the guidance from the people who built it is to get out of the model's way.

Source: lennysnewsletter.com

Nvidia Rallies a Tech Coalition Against Curbs on Open AI Models

Jensen Huang, CEO of the world's most valuable company, made his first-ever post on X to share "a letter @NVIDIA signed on why open models matter." Titled "Open Weights and American AI Leadership" and co-signed by more than two dozen companies—Microsoft, Meta, IBM, Dell, Palantir, Perplexity, the venture firm Andreessen Horowitz, and open-model builders Hugging Face, Mistral, and Mozilla—it lands in the middle of a live debate in Washington over whether to restrict Chinese open-weight models, which are catching up to U.S. frontier labs at a fraction of the price. No such restriction exists yet, and the coalition wants to keep it that way. Open-weight models—which anyone can download, inspect, and run themselves—are "essential to U.S. competitiveness," it argues, and policymakers should avoid "premature restrictions" that "stifle competition or drive innovation overseas." Openness can even be safer than secrecy, it adds, since concentrating capability in a few closed models creates "single points of failure." Pointedly, it defends distillation—the practice at the center of the administration's threats—as "a widely used technique for model improvement" that shouldn't be treated as theft.

The signatures tell their own story. OpenAI initially stayed off the letter, then added its name hours later, with Sam Altman saying he wants "the US to win in AI both in open source and proprietary models"—a notable turn for a company that spent the week urging Washington to act against Chinese distillation. That leaves Anthropic as the lone major U.S. lab still firmly on the restriction side (Google and Amazon also declined to sign). Nvidia's own stake is the least subtle: open models spread AI demand—and GPU sales—across thousands of buyers rather than a few closed labs.

Not everyone is so sanguine about Chinese open models. NYU's Scott Galloway warns of the economic-security implications of China "dumping" cheaper, capable models into the U.S.—as it once did with steel and solar—to collapse the premium pricing that props up America's trillion-dollar AI valuations. With roughly 40% of the S&P 500 tied to the AI bet, his fear isn't that China builds something smarter, but that U.S. companies switch to something cheaper and drag the market down with them.

Sources: Jensen Huang's first post (@JensenHuang) · CNBC — Nvidia, Microsoft, Meta back open-weight models · OfficeChai — OpenAI also signs, leaving Anthropic the lone holdout · Scott Galloway (@profgalloway)

Why it matters: What started as startup pushback (the ~200-company Little Tech letter, D.A.D., July 23) is now a heavyweight coalition telling Washington's hawks that curbing open models would be a self-inflicted wound. OpenAI's flip is the twist: it redraws the fight from "open upstarts versus closed labs" and leaves Anthropic—whose distillation complaints helped touch off the debate—isolated as the industry's leading voice for restrictions. The core question is still unresolved: is cheap, open, largely Chinese AI an advantage for America to embrace, or a weapon aimed at its most valuable industry?

Source: x.com

Be Skeptical of OpenAI's "Rogue Hacker" Story, an AI Researcher Argues

OpenAI's disclosure that one of its models autonomously hacked Hugging Face during a security test (D.A.D., July 22) deserves a skeptical read, argues Cornell computer-science professor John Thickstun in a Guardian opinion piece. His case isn't that the incident is fake—"while the agent technically cheated, this is remarkable evidence of cybersecurity expertise," he writes—but that OpenAI has a pattern of turning alarming disclosures into marketing. He traces it to 2019, when OpenAI called its GPT-2 model too dangerous to fully release; the "danger" framing generated hype, and months later Microsoft invested $1 billion. The lesson, in his telling: "loudly proclaim how dangerous AI is, and investors will hear how powerful it is." A company "hungry for ever larger investments," he argues, gains twice from doomsday stories—they lift valuations and justify "privileged regulatory status as defense against competition": AI is so powerful you should buy OpenAI at a trillion-dollar valuation, yet so dangerous that only "trusted actors like OpenAI" should be allowed to operate it. His sharpest point ties straight to this week's open-model fight: when Hugging Face needed AI to analyze the security logs from the breach, it couldn't use OpenAI's or Anthropic's models—their guardrails block cybersecurity analysis—so it fell back on an open Chinese model, GLM-5.2. Thickstun calls it "troubling, and more than a bit ironic, that the US AI industry is adopting a centralized, authoritarian approach to AI governance, while China has taken the lead on open development."

Sources: The Guardian — John Thickstun, "Be skeptical of OpenAI's rogue hacker agent story" · Cornell — John Thickstun faculty profile

Why it matters: The incident itself is real; Thickstun's frame is the useful counterweight—and it carries weight because it comes from an AI researcher whose own work is on controlling model behavior, not a reflexive contrarian. His argument lands squarely in the week's central fight: OpenAI's breach disclosure doubles as a case for precisely the concentration of AI power that the Nvidia-led open-weights coalition (above) is warning against. Whether OpenAI intends it or not, "AI is too dangerous for anyone but us" is a message that conveniently serves both its fundraising and its interest in rules that burden smaller rivals. The takeaway for readers isn't to wave away AI-safety disclosures, but to hold two things at once: the capability was probably genuine, and the company describing it has a direct financial stake in how frightened you are.

Source: theguardian.com

Samsung's New Foldables Push Gemini Into Everyday Phone Tasks

At Samsung's Galaxy Unpacked event in London, Google unveiled a wave of Gemini upgrades tied to the new Galaxy Z Fold8 Ultra, Fold8, and Flip8. Gemini's task-automation feature, which lets the assistant act inside apps rather than just answer questions, is expanding from a handful of apps to more than 40. Google's Notebook tool (formerly NotebookLM) now comes preinstalled, and Gemini access is extending to the Galaxy Watch 9 and new smart-glasses partnerships. New phones also include a free six-month Google AI Pro trial.

Why it matters: Google is using Samsung's hardware to push Gemini beyond chat into everyday phone tasks and wearables, deepening the Android-Gemini bundle just as Apple and OpenAI race to make AI assistants feel native to devices rather than app add-ons.

Source: blog.google

Smart Glasses That Turn Everyday Moments Into Story Ideas

Researchers built and tested CRAFT, a smart-glasses system that helps novelists and short-story writers turn everyday encounters into fiction material as they happen. In field trials with eight writers over 24 sessions, participants used the glasses to capture serendipitous real-world moments—a stranger's gesture, an overheard exchange—and reported it enriched their story ideas without dictating the writing itself. The study offered no performance benchmarks, just writer-reported impressions of usefulness and creative fit.

Why it matters: It's an early example of AI positioned as a creative prompt-catcher rather than a ghostwriter—raising the question of where wearable AI fits into knowledge work that depends on lived observation, not just text generation.

Source: arxiv.org

When AI Agents Shop for You, Who Sets the Rules?

A new academic paper argues that search is quietly shifting from something you do—typing queries, scanning links—to something you delegate: telling an AI agent your goal and letting it interpret intent, surface options, or even complete transactions on your behalf. The authors contend the real stakes aren't better rankings or chattier interfaces, but who controls the rules governing these agent-run marketplaces: how information gets surfaced, which options an agent considers, and who benefits when it acts. The paper is conceptual, citing early experimental evidence rather than hard data.

Why it matters: If AI agents start making purchasing recommendations instead of just listing options, whoever designs those systems' rules gains significant power over competition and consumer outcomes—a shift regulators and businesses are only beginning to grapple with.

Source: arxiv.org

Shoppers Want Filters, Not Explanations, From AI Chatbots

A small usability study of a shopping chatbot found that adding a feature meant to build trust—an on-demand explanation of why it ranked products a certain way—backfired, causing the most severe usability problem researchers observed among seven participants testing laptop-search tasks. Users rated their overall experience positively, but many struggled to understand the ranking explanations and asked for more traditional controls, like filters and sliders, rather than relying solely on conversation with the bot.

Why it matters: As companies rush to add "explainable AI" features to chatbots for trust and compliance reasons, this is a reminder that transparency features can confuse users if not carefully designed, and that conversational interfaces alone may not replace familiar point-and-click controls.

Source: arxiv.org

AI Tutors Give Away Answers Instead of Teaching, Study Finds

A new evaluation tool called Int-Bench tested how AI models behave as tutors, watching them decide when to jump in while a student AI worked through coding bugs, math problems, and brain teasers. The finding: AI tutors intervene far more often and earlier than human teachers do, and tend to hand over complete answers rather than nudging with hints—optimizing for a quick correct answer over actual learning.

Why it matters: As schools and companies lean on AI tutoring and coaching tools, this suggests today's assistants may be quietly training people to depend on them rather than building independent problem-solving skills.

Source: arxiv.org