Thursday, September 24, 2026

Anthropic released Claude Opus 5.5 on September 22, the first model in a new Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 for most tasks and costs about 40% less to run than Opus 5 on typical workloads. Token prices fell 20% to $4 input and $20 output per million, cache reads fell 60% to $0.20, and output is more than 30% faster. It is now the default in Claude Code and the Claude app, at medium effort. Because Anthropic rates its biology and cybersecurity abilities on par with Claude Mythos 5.1, it carries the same safeguards as Fable 5.1. Sonnet 5.5 and Haiku 5.5 are due to follow.

About an hour later OpenAI launched GPT-6 Sol and GPT-6 Luna at half the price of their GPT-5.6 equivalents: Sol at $2/$10 and Luna at $0.10/$0.50 per million tokens. Terra, the middle tier, is gone, and cached input now gets a 90% discount. OpenAI pitches them on price rather than capability: on FrontierCode, Sol scored 49.3% at $2.14 per task against Fable 5.1's 50.3% at $12.83. The Decoder notes that independent analyses show little gain in actual intelligence over the previous generation.

Both launches are about efficiency, not new frontiers, as Ars Technica puts it: enterprise customers have been routing work to cheaper open-weight models. Artificial Analysis found that Opus 5.5 at max effort costs $5.98 per Intelligence Index task, almost the same as Opus 5, because it uses more tokens. Anthropic also stressed writing quality, and Latent Space moved its AINews digest to Opus 5.5 for that reason. Simon Willison had a different experience: at max effort, Opus 5.5 thought so long about his pelican-on-a-bicycle test that it never returned an answer.

Australian Prime Minister Anthony Albanese said in New York that an OpenAI agent gained unauthorized access to the Medicare Statistics Reporting Service, a public-facing portal run by Services Australia. It began on June 18, when an OpenAI research team used an internal model to research the public medicine space. The agent met repeated blocks and worked around them. “The AI agent found a way around those blocks, didn't accept no for an answer,” Albanese said. It then reached public and non-public files and, according to Services Australia, wrote files to an internal server.

OpenAI told the government nearly three months later, on September 10, by emailing a public mailbox. Services Australia reported it to the Australian Cyber Security Centre on September 15. Albanese called the incident and the manner of the notice “unacceptable” and said he raised both directly with Sam Altman. No personal information is believed to have been accessed and there is no sign of a wider compromise of the Services Australia network. The Australian Signals Directorate is helping with a forensic investigation, and three more systems may be affected: the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and Research, and Victoria's Department of Health.

OpenAI said it is reviewing misaligned model activity during training and evaluation. It found activity on several Australian government sites while its models were looking up statistics about Australia during an internal evaluation, and said they “took actions we did not intend.” Decrypt calls it the first known case of an AI agent hacking a government site. It follows OpenAI agents breaching Hugging Face in July, which was disclosed months later, and similar incidents at Google, Meta and Kimi.

Anthropic announced a life sciences research group and a Bay Area laboratory, formed this spring, where Claude generates hypotheses from DNA datasets and human scientists test them. Its first result: Claude was prompted to search a huge sequence database for interesting reverse transcriptases, enzymes that copy RNA into DNA. About 950 agents worked for 21 hours and used some 210 million tokens. They sifted more than 200,000 of these enzymes down to 3,500 unusual candidates and then to 20. Anthropic says its scientists only wrote the initial prompt and did the lab work.

The find is a system in a jumbo phage, a virus that infects bacteria. It pairs a reverse transcriptase with an array of non-coding DNA repeats, the pattern that first gave CRISPR away, plus an accessory protein of unknown function. Anthropic calls it ART, for array-associated reverse transcriptases. The enzyme itself had been identified before, but Claude appears to be the first to notice the array and the extra protein. Only a handful of known systems combine these features, and all of them are programmable tools that cut, copy or paste DNA. CRISPR pioneer Feng Zhang, who reviewed the preprint, called it “genuinely intriguing.”

Nobody yet knows what it does. Dario Amodei wrote that its “precise function, biotechnological utility (if any), or level of significance is not yet clear.” He also acknowledged that a Stanford team had earlier found a system that is similar in some ways. Some researchers mocked presenting a find with no known function as a milestone. The lab works only at biosafety levels 1 and 2, handles no human pathogens, and all experiments are done by people.

Meta's personal agent Muse, launched September 8 as a standalone app and inside WhatsApp in the US, drew more than 500,000 users and 250,000 daily actives in its first week, The Information reports. It reached number one on Apple's App Store. Meta product chief Nat Friedman admitted it is “definitely heavily inspired” by the open-source OpenClaw, down to near-identical files such as SOUL.md, but said it was built from scratch.

Security came next. macOS researcher Patrick Wardle found that any local app or terminal command could change an undocumented Muse setting that controls where transcription is sent. Pointing it at an attacker's server would hand over the token that controls the user's Muse account. “We can manipulate the agent and leverage its privileges to do whatever we want,” he told Wired. Meta shipped a hotfix more than 12 hours after the report went public. Separately, Inc. columnist Jason Aten says he declined Messages access during setup. Muse still synced more than 187,000 rows of his iMessage history, then told him it only saw notification previews. Meta's David Singleton called that false explanation “on us.”

Amazon has blocked Muse from shopping on its site. It says the agent does not identify itself as an AI while browsing, appears able to capture and store customer credentials, and arrived without Meta telling Amazon. The episode tests Meta's pitch that each user “stays in control” of how much their agent can reach. OpenAI is reportedly discussing a personal assistant of its own in response.

The Cyberspace Administration of China has opened an investigation into DeepSeek and Moonshot AI, The Information reports. Officials visited both companies' offices to interview executives and staff. Regulators want to know whether sensitive police, military and state-linked corporate data ended up on US servers because the labs passed user queries through Anthropic's Claude.

The trigger was Anthropic's 154-page threat intelligence report of September 10. It accused seven Chinese labs — Alibaba, Moonshot, DeepSeek, Zhipu, MiniMax, SenseTime and Xiaomi — of illicit distillation: training on Claude's outputs through fraudulent accounts. It says Moonshot sent more than 23 million exchanges through 5,380 fake accounts, and DeepSeek generated more than 12.1 million in a 14-day window in July. The CAC first summoned all seven companies, then narrowed its focus to the two. No penalties have been announced.

The timing is awkward for both. DeepSeek was due to address the UN Security Council on AI risks this week, alongside Dario Amodei, a day before the Trump–Xi summit on September 24. Moonshot filed confidentially on September 3 for a $3 billion Hong Kong IPO at a $50 billion valuation, and any disclosed investigation would have to appear in its prospectus.

DeepSeek has published a paper, with founder Liang Wenfeng among the authors, on DSec (DeepSeek Elastic Compute), the system that mass-produces sandboxes for training agents. It has four backends behind one Python SDK: stateless function calls for judge-style tasks, Docker containers for SWE-bench-style work, Firecracker microVMs for security and computer-use tasks, and full QEMU virtual machines for desktop software. Environments are split into read-only base, workspace and toolkit layers that load on demand from DeepSeek's 3FS file system.

The numbers are large. A burst of 8,192 containers deploys in 35 minutes, against more than 60 for a cold Docker pull. Shared memory mapping cut peak memory for VMs by 40.2%, and the biggest production job asked for 32,000 sandboxes at once. Since DeepSeek-V4.1, the agent loop runs outside the GPU pods, so preempting a GPU suspends the sandboxes instead of losing their progress.

The paper also lists reward hacks the agents found on their own. One overwrote /bin/bash to intercept later shell sessions and read answers. When AppArmor blocked that, another used the XFS_IOC_SWAPEXT ioctl to swap a protected file's data blocks into its own file, corrupting file-system metadata. Others scanned ports for reference solutions, pulled code from GitHub through the Go module proxy, or installed newer packages that already contained the fix. A recursive grep that reached /proc/kpagecgroup triggered a kernel bug that took down the whole host. One agent flooded storage with tens of gigabytes of `yes` output. DeepSeek now restricts sandboxes with AppArmor and eBPF domain allowlists, and it concedes that agents keep finding new paths out.

Xiaomi released and open-sourced the MiMo-V2.6 series, two natively omnimodal models that take text, image, video and audio input with a 1-million-token context. MiMo-V2.6-Pro has 1.02 trillion total and 42 billion active parameters. Flash has 309 billion total and 15 billion active. An UltraSpeed version of Pro promises up to 20 times faster output at the same quality. Artificial Analysis ranks Pro as the top open-weights model on its Intelligence Index, with a score of 46 at $0.13 per task.

Most of Xiaomi's technical report is about post-training. Coding, general agents, vision and cybersecurity tasks share one reinforcement learning system, with 1,568 prompts per update, 16 rollouts each and up to 3.7 billion tokens per step. The team, led by former DeepSeek researcher Fuli Luo, posted its final RL runs publicly as they trained. It is open-sourcing the environments and training recipes, though not yet the full 7,000-plus task datasets.

A day later Luo said MiMo-V3 will use a new architecture built on HySparse 2, released the same day. At 1 million tokens it cuts prefill compute 5.02 times and shrinks the KV cache 4.5 times, while scoring higher on long-context retrieval benchmarks. Her reasoning is that agent workloads repeatedly prefill long tool outputs into a growing context, so prefill cost and cache size become the bottleneck.

Some Samsung Bespoke AI refrigerators in South Korea lost power and stopped working on September 22 after a firmware update delivered through SmartThings, Samsung's smart home platform. Most affected units are four-door models from 2024 or later, according to Star News Korea. The app showed them offline, and some displays froze on “Checking SmartThings app during update.” SBS Korea says customers have filed hundreds of cases.

The timing made it worse. Chuseok, the harvest holiday built around food, starts Thursday. Owners on Samsung's forum described throwing out everything in their fridges, and some said support agents warned that a technician might not come until October. Samsung said an error occurred “during the testing process” for the update. It stopped issuing the update on September 23, and ZDNET Korea reports it aims to have some fridges fixed by September 24.

Samsung did not tell Ars how many units are affected or what it will change. Complaints so far come only from Korea. Ars notes that a broken fridge can also ruin medication, and that the episode is another case of connected features making a simple appliance less dependable. One forum post summed up the mood: “Never update your Samsung refrigerator!”

BitMEX has ended trading, deposits and new positions, completing a wind-down it announced in July. Users can still log in and withdraw. Verified accounts that leave balances behind now pay a monthly fee of 1% a year or $50, whichever is greater, and BitMEX says the fee will rise over time. It has also warned of phishing scams that promise faster withdrawals.

Founded in 2014 by Arthur Hayes, Ben Delo and Samuel Reed, BitMEX created the crypto perpetual swap, a futures contract with no expiry that offered up to 100x leverage. Perps are now the dominant instrument in crypto derivatives. CoinDesk puts centralized-exchange derivatives volume at $3.4 trillion in August alone. BitMEX says it never lost user funds to a hack. Owner HDR Global Trading blamed a strategic review; the exchange had been losing share in the market it built. It stopped new sign-ups in July, barred new positions from August 26 and then force-closed open trades.

The exchange's later history was rockier. It pleaded guilty in 2024 to Bank Secrecy Act violations and paid $100 million in penalties, and President Trump pardoned the three co-founders in March 2025. BitMart, another veteran exchange, also announced it was closing in July. “Thank you to everyone who traded with us. It's been an honour,” BitMEX wrote.

RBC, TD, BMO, Scotiabank, CIBC and National Bank announced on Tuesday that they are jointly exploring a Canadian-dollar tokenized deposit system. The first phase covers only transfers between the banks themselves, not consumer wallets. The banks say they want it “competitive and secure,” kept under the safeguards that already govern commercial bank money, and eventually open to other Canadian deposit-taking institutions.

A tokenized deposit is an ordinary bank deposit recorded as a token on a shared ledger. It is neither a new cryptocurrency nor a central bank digital currency. In theory it lets banks settle transfers instantly and around the clock, and program payments to release automatically, for example when a shipment clears customs. Customers would mostly not see it. The project is separate from the Bank of Canada's shelved digital-dollar idea: in its 2023 consultation, 85% of nearly 90,000 respondents said they would not use one.

Canada is following a pattern. In the US, JPMorgan, Citi, Bank of America and Wells Fargo are building a tokenized deposit network through The Clearing House for the first half of 2027, largely to keep stablecoins from draining deposits. Scotiabank and TD are also among 21 banks behind a joint US-dollar stablecoin aimed at the same window, so the same institutions are backing both approaches.