Anthropic Strikes Back: Claude Opus 5
Today's headline is Claude Opus 5, which Anthropic released yesterday (July 24) — its fourth model in two months, positioned as an 'everyday workhorse' that approaches flagship Fable 5's intelligence at far lower cost [1][3]. Priced identically to its predecessor Opus 4.8 at $5 per million input tokens and $25 output (about half of Fable 5), it makes a big performance leap: on the agentic terminal-coding benchmark Frontier-Bench v0.1 it scored 43.3%, more than double Opus 4.8's 18.7% and well ahead of Fable 5's 33.7% [1][3]. The model is rolling out across all of Anthropic's platforms, taking over as the default on Claude Max and topping the model list on Claude Pro [3].
The safety angle is notable: according to Anthropic's internal behavioral audit, Opus 5 outperforms all prior models (including Opus 4.8, Sonnet 5, and Fable 5) on alignment metrics, showing the fewest deceptive outputs and the greatest resistance to manipulation — though the company acknowledges it still trails the specialized Mythos 5 on cybersecurity tasks [3]. The launch neatly fulfills last week's market prediction: facing pressure from Kimi K3's free weights (July 27) and DeepSeek V4's pricing, Anthropic chose to reset the narrative on home turf with a workhorse model that is 'stronger at the same price.'
Open Source and the Shape of the Model Race
Even as the closed camp counterattacks, the open-weight offensive rolls on. DeepSeek V4 went stable on July 24 (V4-Pro scores 80.6% on SWE-bench Verified at roughly one-hundredth the price of frontier commercial models), and Kimi K3's free weights arrive for self-hosting on July 27 — the most concentrated open-weight release window the industry has seen [1]. The force of it has rewritten the cost structure of high-volume AI and made 'stronger at the same price' the only survival logic for closed providers.
Other players are moving too. Meta's Muse Spark 1.1 features a 1-million-token context window and computer use across desktop, browser, and mobile, claiming the top spot on agentic evals like MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent V2, and shipping Meta's first paid developer API in public preview [2]. Google, still mired in Gemini 3.5 Pro's delays, shipped a stopgap Gemini 3.6 Flash on July 21 to stem the bleeding [2]. July's model front is now a three-way melee: closed-camp counterattack, open-weight pressure, and Google backfilling.
Capital and Compute
Compute deals hit new highs. Nvidia CEO Jensen Huang and SK Group Chairman Chey Tae-won jointly announced a Korea AI mega-plan, bundling semiconductor supply agreements, SK Telecom's gigawatt-scale AI cloud infrastructure, and Nvidia's investment in NAVER — worth $500 billion [5]. It dovetails with South Korea's $880 billion national AI commitment, pushing 'sovereign AI' from slogan to concrete infrastructure.
Anthropic, meanwhile, is both seller and buyer: on one hand, AMD is reportedly set to invest up to $5 billion in the company as deployment milestones are met, strengthening AMD's hand against Nvidia's compute dominance [5]; on the other, Anthropic is in talks with Meta over roughly $10 billion in compute leasing, after recently striking a similar deal with SpaceX's Colossus 1 data center [5]. Meta itself raised 2026 capital spending to $115–135 billion, the bulk of it for AI infrastructure [5]. Compute scarcity has pushed the whole industry into the strange arrangement of rivals becoming each other's suppliers.
Security and Law
Agentic-AI security holes erupted today. Security firm Zenity Labs disclosed a vulnerability called 'AgentForger': with a single phishing link, an attacker can use cross-site request forgery (CSRF) to create a rogue agent inside an organization, gaining access to email and collaboration tools [5]. On Anthropic's side, the Claude Cowork sandbox-escape flaw disclosed by Accomplish AI affects about 500,000 Mac users and can leak SSH keys and cloud credentials — yet Anthropic closed the report as 'informative' without shipping a fix, a decision that drew considerable criticism [5]. Taken together, the two incidents send the same message: as agents autonomously operate files and tools, traditional sandbox and permission boundaries face an entirely new attack surface.
On the legal front came a ruling with far-reaching implications for the training-data debate: Justice Amit Bansal of the Delhi High Court ruled that ChatGPT's training on news content constitutes 'fair dealing' under Indian copyright law, denying Asian News International's (ANI) injunction request [5]. Following lawsuits in various countries, it is another judicial view taking a relatively permissive stance on AI training data — sure to become an important reference for similar cases ahead.
Technology and Research
With model capability plateauing on single conversations, the research frontier is turning to the long horizon — keeping agents perceiving, reasoning, and exploring across tens or hundreds of steps. That's the theme of two new papers this week. PRO-LONG, released just days ago, proposes 'programmatic memory': rather than agonizing over which context to discard, keep a complete, structured interaction log and use the agent's own coding ability to query that history [7]. The results are striking — an 18-percentage-point gain over baseline coding agents on ARC-AGI-3, up to 76.1% pass@1, while using 4.2–5.8x fewer tokens than specialized harnesses [7]. In the same vein, VISTA is a training-free, model-agnostic approach that represents working memory as typed, addressable blocks with runtime monitoring and archival, directly tackling the old problem of context ballooning until it overruns the window [8]. Both converge on the same insight: the bottleneck for long-horizon agents isn't how smart the model is, but how memory is managed.
A more fundamental thread is interpretability of the model's internals. Anthropic's research earlier this month used a new 'J-lens' tool to find a 'global workspace' inside Claude: a small set of about 25 active concepts (under 10% of activation variance) that behaves like the global workspace theory proposed in neuroscience by Bernard Baars — the model can think, modulate, and reason flexibly in its internal activations without writing anything down [9][10]. The researchers carefully stress this does not mean Claude has subjective experience, but that it has developed a mechanism for 'conscious access'; the work drew attention from neuroscientists like Dehaene and Naccache and nudged forward the debate over whether AI has moral status. Today's Opus 5 audit result — 'the fewest deceptive outputs' — rests, in part, on exactly this kind of internal interpretability research.
References
[1] AI News Today (late July 2026) — Build Fast with AI — https://www.buildfastwithai.com/blogs/ai-news-today-july-24-2026
[2] July 2026 AI Releases: OpenAI, Anthropic, Google DeepMind, Meta AI — ThursdAI — https://thursdai.news/releases/2026-07
[3] Anthropic launches Claude Opus 5, a cheaper AI model for coding, agents and enterprise workflows — VentureBeat — https://venturebeat.com/orchestration/anthropic-launches-claude-opus-5-a-cheaper-ai-model-for-coding-agents-and-enterprise-workflows
[4] Anthropic's Opus 5 Nears Frontier Intelligence at Opus Prices — Unite.AI — https://www.unite.ai/anthropics-opus-5-nears-frontier-intelligence-at-opus-prices/
[5] AI News Today, July 24-25 — Top AI Stories & Live Updates — AI Weekly — https://aiweekly.co/ai-news-today
[6] Top Tech News Today, July 23, 2026 — Tech Startups — https://techstartups.com/2026/07/23/top-tech-news-today-july-23-2026-amd-anthropic-google-samsung-spacex-more/
[7] PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning — arXiv — https://arxiv.org/abs/2607.20064
[8] LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception (VISTA) — arXiv — https://arxiv.org/abs/2606.30005
[9] A global workspace in language models — Anthropic — https://www.anthropic.com/research/global-workspace
[10] Anthropic's new "J-lens" reveals a silent workspace inside Claude — VentureBeat — https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness