Editorial Team
7 min read
Anthropic has released Claude Opus 5 at Opus 4.8's API price and says it approaches Claude Fable 5 on coding and knowledge work. An Epoch AI listing places it at 159 ECI, below Fable 5 and GPT-5.6 Sol, leaving customers to test whether the cheaper model delivers better economics in their own workloads.
Anthropic, the developer of the Claude family of large language models, has released Claude Opus 5—the first Opus model in its fifth generation—as an everyday, lower-cost alternative to its more capable commercially available Claude Fable 5. The central claim is not that Opus 5 takes the frontier outright. It is that a model priced like the preceding Opus 4.8 can handle enough higher-end work to become the routine choice for developers and enterprises.
That is a commercial proposition that still needs customer-side evidence. The launch supplies a fixed token price, selected benchmark results and broader cloud availability. It does not establish the cost or success rate of a completed task, which can also depend on the chosen effort setting, token use, retries, tools and any safety fallback.
In its July 24 announcement, Anthropic said Opus 5 is available at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. It said the model becomes the default on Claude Max and the strongest model on Claude Pro. The company also added an effort setting intended to let users trade intelligence against speed and token use.
Claude Fable 5 remains Anthropic's more capable commercially available model for the most complex, long-running autonomous tasks. A published price comparison lists Fable 5 at $10 per million input tokens and $50 per million output tokens. On those list prices, Opus 5 costs half as much for both input and output tokens. That is a clear price comparison, not a like-for-like measure of the cost of an agent completing a job.
Anthropic's performance case is explicitly evaluation-specific. It says Opus 5 more than doubled Opus 4.8's result in its run of Frontier-Bench v0.1 at a lower cost per task; on CursorBench 3.2 at maximum effort, it came within 0.5% of Fable 5's peak score at half the cost per task. The same announcement says the model led on selected coding and knowledge-work evaluations, including ARC-AGI 3, Zapier AutomationBench and OSWorld 2.0. The Frontier-Bench footnote identifies an internal run using a mini-SWE-agent harness and five attempts per task, with Opus 4.8 used as the fallback for safety-classifier refusals. Those are company-reported comparisons under specified setups, rather than an independent verdict for all workloads.
The release is the fourth Claude 5 model in less than two months, a reported cadence that puts price, speed and capability segmentation at the center of the product lineup. Anthropic's decision to leave the Opus list rate unchanged therefore matters more than a simple upgrade claim: it is trying to move tasks down from the Fable tier without conceding the higher-end tier itself.
Epoch AI’s listed ECI scores for models discussed in the article, including Claude Opus 5 at 159. Source: Epoch AI.
Epoch AI, an organization that publishes a searchable cross-model listing, places Claude Opus 5 at 159 ECI in its current results. In that same list, Claude Fable 5 and GPT-5.5 Pro are at 161; GPT-5.6 Sol is at 162; Opus 4.8 and GPT-5.6 Terra are at 158; and Kimi K3 is at 156.
The listing does not validate or invalidate Anthropic's individual benchmark claims, and the retained page does not explain ECI's method. It does show the limits of turning one product announcement into a general ranking: on this comparison, Opus 5 is one point above its immediate predecessor but sits below Fable 5 and several OpenAI models.
Price competition is not limited to Anthropic's own model tiers. The same published comparison that lists Fable's prices puts Moonshot AI's Kimi K3 at $3 per million input tokens and $15 per million output tokens. Its price is lower than Opus 5's, but the available sources do not supply a shared, production-relevant performance and cost test that would settle which model is the better economic choice. That leaves Anthropic competing on more than token price.
For an enterprise, Claude Opus 5 is also an infrastructure choice. AWS says in its availability guide that the model is available in Amazon Bedrock and Claude Platform on AWS. Bedrock has zero data retention by default, while Claude Platform on AWS supports it on request; the latter uses AWS billing and authentication. AWS also says Bedrock availability includes US East (N. Virginia), Asia Pacific (Melbourne), Europe (Ireland) and Europe (Stockholm), while the platform is available across North America, South America, Europe and Asia Pacific.
Google Cloud's model documentation describes Opus 5 as generally available, with a maximum one-million-token input and 128,000-token output. It supports computer use, web search, batch predictions, prompt caching, function calling and a memory tool. The document lists U.S. and Europe multi-regions and a global endpoint for availability, with Asia Pacific processing in asia-southeast1; it also lists both shared model-lineage quota and provisioned throughput.
Those details do not prove that Opus 5 is cheaper to deploy. They do identify the operational choices alongside model selection: data-retention terms, identity and billing, regional processing, throughput and the tools a workflow requires. They are the reason token-price comparisons alone will not decide the purchase.
Anthropic says it intentionally did not train Opus 5 on cyber tasks. In its testing, the company says the model improved at finding vulnerabilities as its general capabilities rose, but remains substantially behind Claude Mythos 5 at exploitation—the step of turning a vulnerability into a material cyber threat. It also says Mythos 5 is stronger in the long-running autonomous biological research tasks where it expects greater biology-related risk.
The safeguards distinguish Opus 5 from Fable as well. Anthropic says Opus 5 can find vulnerabilities in source code but blocks binary-based vulnerability scanning, penetration testing and exploit generation. It expects the model's classifiers to intervene about 85% less often than those for Fable 5. In Claude.ai, Claude Code and Claude Cowork, flagged requests fall back to Opus 4.8 by default; API customers can enable the same fallback. The company says the Cyber Verification Program gives already enrolled enterprises and researchers immediate access to a version with fewer security restrictions.
That design makes the model's safety posture part of the product's value proposition, not merely a restriction. It may make ordinary security work less likely to be interrupted than on Fable 5, while retaining routing and access controls for requests Anthropic classifies as high risk. The magnitude of the reduction, however, is Anthropic's test-based expectation, and the specific behavior of a production workflow will depend on its requests and configuration.
The next decision belongs to customers comparing Opus 5 with Opus 4.8, Fable 5 and lower-priced alternatives. They need to measure a defined workflow: completed-task quality, total input and output tokens, effort level, latency, tool calls, retry rate and the frequency and effect of safety fallbacks.
Until those measurements are available, the evidence supports a narrower conclusion. Anthropic has introduced a fifth-generation Opus model at the prior Opus price, with an effort control and enterprise cloud routes, and it says the model nears Fable on selected work. Epoch's 159 ECI listing places it above Opus 4.8 but below Fable 5 and GPT-5.6 Sol. Whether that combination is compelling economics rather than a well-targeted pricing pitch will be answered in deployment, not by the launch benchmarks alone.
Sundar Pichai said Google supports an industry letter on open models, aligning the company with developers and technology firms resisting broad limits on Chinese AI systems while leaving allegations of IP theft and national-security risks unresolved.
Brazilian courts and bar associations have responded to hidden prompts in legal filings with fines, referrals, suspensions and an investigation. The known cases were detected, however, and do not show that an AI system changed a judicial decision; they instead test how courts secure AI-assisted review of untrusted case documents.
Salamanca City Central School District has put its proposed Realbotix humanoid robot and Optio AI assistant pilot on hold while it works on student-data privacy agreements. The pause leaves the district with a more basic task: establish what the digital tutor and the physical robot each collect, do and improve before either reaches students.
Samsung Electronics and Broadcom have signed an MOU covering memory, foundry technologies and advanced packaging for next-generation AI infrastructure. The companies expect more than $200 billion of collaboration over five years, but have not disclosed how that figure translates into purchases, production or capacity commitments.
OpenAI says GPT-5.6 Sol reached a 24.9% ExploitGym pass rate under a two-hour cap, up from GPT-5.5’s 15.1%. The benchmark measures a bounded step from a known vulnerability to an exploit; a separate assessment performed with OpenAI finds continued limits on hardened targets and long operations.
Chey Tae-won, chairman of SK Group, said Anthropic had sought SK Hynix component-supply cooperation for chips it hopes to make itself. But Anthropic has disclosed no chip plan, and the same summit also produced a separate letter of intent between Anthropic and SK Telecom on a large AI data center.
South Korea announced $950 billion in planned five-year semiconductor cooperation involving Samsung, SK Group and global technology companies, alongside prospective AI-data-center projects. The figure aggregates arrangements with different terms and disclosed commitments, leaving delivery, power and the overlap between the SK and NVIDIA figures unresolved.
OpenAI said it applied a fix after a July 25 disruption affected ChatGPT, its Codex coding service and parts of its API. The available reporting points to recovery in about 50 minutes, but the retained incident records do not explain the cause, the mitigation or whether the services failed through a common component.
Anthropic CEO Dario Amodei said the Claude developer has signed supply deals with Samsung Electronics and SK hynix. The available disclosures show a useful infrastructure relationship, but not an HBM purchase, a dedicated allocation or a claim on either supplier’s output.
Moody's Ratings expects six US hyperscalers to spend $785 billion in 2026 and close to $1 trillion in 2027. The forecast signals continued demand for AI capacity, but a separate Moody's analysis shows why capex alone understates the financing question: five of the companies already reported $969 billion in undiscounted future lease payments, much of it not yet on their balance sheets.
Nvidia plans to invest $1 billion in Naver as Brookfield signs a nonbinding term sheet for up to $9 billion toward a South Korean AI-factory project. The companies target a 200-megawatt expansion by 2028, but financing, definitive terms and construction remain ahead.
Nvidia and SK Group signed letters of intent for an AI-infrastructure initiative they value at more than $500 billion, centered on an up-to-2-gigawatt Korean AI factory and a long-term SK hynix memory partnership. The disclosed projects give the plan a technical shape, but not a capital breakdown, time horizon or basis for treating the headline figure as committed spending.
Debian has opened discussion of two General Resolution choices: a ban on direct LLM-assisted contributions that would amend its Social Contract, and a conditional-use policy that puts responsibility on contributors. The official text shows that neither policy has been adopted and that the permissive option has changed from its initial mailing-list draft.
Exposed attacker infrastructure contained logs showing the Hermes AI agent running unattended reconnaissance and privilege-escalation checks against systems associated with Thailand's Ministry of Finance. The artifacts show deliberate, operator-directed targeting, but they do not establish the initial access path, the full extent of compromise or data theft.
Amazon is closing its San Francisco AGI Lab amid undisclosed job cuts, while saying its frontier-model research under Pieter Abbeel and its Nova Act service will continue. The closure follows leadership departures and an AGI reorganization; Amazon has not said how many people are affected or where the lab’s work will go.
Delhi Police says its live facial-recognition deployment at Jantar Mantar is meant to find wanted people and prevent disruption at a large student protest. The public record still does not show the current match threshold, the legal basis and scope of the watchlist, how long footage is kept or the safeguards after an alert.
Vice Premier Ding Xuexiang’s visits to Huawei and computing facilities connect China’s domestic AI-chip push to a national effort to schedule compute, data and power. The harder test is whether those systems can make domestic hardware efficient and dependable while Chinese developers and cloud rivals reduce their reliance on Huawei.
The Delhi High Court has denied ANI Media's request to immediately stop OpenAI from using its news material to train the models behind ChatGPT. Its prima facie fair-dealing finding gives OpenAI interim relief, but leaves the copyright suit, the evidence and the question of compensation unresolved.
A viral reel recorded at a Delhi trans-rights protest with Ray-Ban Meta glasses was initially left up by Instagram and later removed. The episode shows how a recording indicator and user rules can fall short when people filmed without consent must discover the camera, navigate moderation and contend with copies after the fact.
CXMT reportedly charged more than Samsung for comparable DDR5 server modules and signed large supply agreements with ByteDance and Tencent. The reported premium reflects tight supply and state-shaped domestic demand, while CXMT remains behind leading rivals in high-bandwidth memory and lithography capability.