New Claude models, open weights and Australia's AI rules

Anthropic sells Claude in four sizes. Fable is the frontier tier, above Opus. Sonnet balances speed and capability. Haiku is the smallest and fastest. A new version supersedes the last at its tier, and older versions usually stay available for a time.

Claude Opus 5.5, released 22 September. It is on the paid plans. Through the API it costs US$4 per million input tokens and US$20 per million output tokens, 20 per cent below Opus 5. It can be used with zero data retention. Anthropic’s developer documentation recommends starting with it for most work.

Claude Sonnet 5.5, released 28 September. It is open to everyone, including free users. It costs US$2 and US$10. It can be used with zero data retention.

Claude Haiku 5.5, released 7 October. It starts at US$0.10 and US$0.50, a tenth of Haiku 4.5’s price on prompts up to 100,000 tokens. Anthropic pitches it at high-volume work such as sorting, extracting and routing.

Claude Fable 5.2, released 9 November. It is Anthropic’s most capable model. It is not on the free plan. Anthropic cut its price to US$8 and US$40, from US$10 and US$50. What you send it is kept for at least 30 days (section 2).

Claude Mythos 5.2 is the same underlying model as Fable 5.2 without Fable’s cybersecurity and biology safeguards. Anthropic offers it only to organisations vetted through its cyber and life-sciences verification programmes. The earlier wording confining Mythos to United States organisations has gone. No Anthropic page yet says a New Zealand organisation can qualify.

A token is a fragment of text. On Anthropic’s current models a million tokens is about 555,000 words, so a token is roughly half a word. Most people pay a monthly subscription. Token prices matter to organisations that build on the API.

Zero data retention is an arrangement under which the provider does not store your prompts or its responses once it has answered. Eligible business customers can obtain it by agreement. Where confidential or privileged work depends on it, the model you choose matters.

Anthropic now calls its top tier “Covered Models”, a designation it applies to models that cross capability thresholds. Fable and Mythos are the only models on the list. Prompts and responses sent to them are kept for at least 30 days for safety monitoring, then deleted automatically. They can be kept longer for a safety investigation or where the law requires it.

Zero data retention is not available for Covered Models in Claude Enterprise organisations, in workspaces or on third-party cloud platforms. Eligible customers will be offered it with Fable for internal business use for a limited time. Anthropic describes that offer as a transition to Enterprise Frontier Safeguards. That product runs the safety monitoring inside the customer’s own infrastructure, and Anthropic says it is compatible with zero data retention.

Opus 5.5 and Sonnet 5.5 keep the established position. Both can be used with zero data retention.

Our view, as in July: the 30-day record does not by itself put confidentiality, privilege or the Privacy Act in issue. Where a matter depends on a zero-retention arrangement, use Opus 5.5 or Sonnet 5.5 for it. Our Claude assessment carries this as a dated alert.

One more change reaches the business plans. In Claude Team and Enterprise organisations created on or after 31 October 2026, memory starts on for each member. In older organisations it stays off until a member turns it on. Owners can turn memory off for the whole organisation.

Training an AI model produces one enormous file of numbers, called the weights. That file is the model. The software that runs it is small and widely available.

Anthropic, OpenAI and Google keep their weights on their own servers. You reach their models over the internet and pay for access. An open-weight developer publishes the file. Anyone can download it and run it on computers they control. Nothing then goes back to the developer.

Open weight is a narrower thing than open source. The Open Source Initiative’s definition of open-source AI also requires the training code, information about the training data, and freedom to use the model for any purpose. Most releases publish the weights alone. Several of the leading licences also attach conditions for large commercial users, such as a separate agreement above a revenue threshold.

The leading open-weight models are Chinese. Moonshot published Kimi K3.1 Pro, the strongest, in early November. DeepSeek, Z.ai (GLM-5.3), Xiaomi and Alibaba (Qwen) follow. The leading Western releases are Mistral Large 4 from France, NVIDIA’s Nemotron 3.5, Google’s Gemma 4 and OpenAI’s gpt-oss, which dates from August 2025.

Artificial Analysis publishes an independent composite score. Version 5 of its Intelligence Index, released on 27 October 2026, combines nine evaluations across knowledge work, coding, science and general reasoning, with each category weighted equally.

The gap has widened. When the index moved to version 5, the distance between the best open and best closed model went from 3 points to 12. The release of Fable 5.2 has since taken it to 19. The second-best open model, DeepSeek V4.1 Flash, scores below Anthropic’s smallest model.

Other measures point the same way. Epoch AI, which tracks the field, estimated in May 2026 that open models trail the closed frontier by about four months. On 1 May 2026 the United States Government’s AI evaluation centre, CAISI, reported that DeepSeek V4 performed similarly to GPT-5 on its own tests. That is below what DeepSeek’s own figures suggested. CAISI put DeepSeek V4 about eight months behind the frontier.

Sovereignty, in the AI sense, is the question of who controls the system you rely on: the law over it, where it runs, the terms you hold it on, and whether you can own it. Our sovereignty module sets out the eight layers of control and who holds each one.

One answer to the sovereignty question is for a country to build its own models. Europe’s leading example is France’s Mistral, founded in Paris in 2023. Mistral publishes many of its models as open weights. Its best, Mistral Large 4, released at the end of October, scores 26 on the index. The best Chinese open model scores 46. Anthropic’s Fable 5.2 scores 65.

Our view: on these figures the European approach has failed to keep pace. The frontier is being set by laboratories in the United States and China. A model a country builds for itself gives it control of a system well behind that frontier.

New Zealand is not in a position to build its own. Its realistic choices are to hold open weights, most of them Chinese, or to contract carefully with the American providers. Each choice carries its own risks.

A model has to fit in a computer’s memory to run at a useful speed. Apple markets its top Mac Studio for running large language models. Its memory is shared between the processor and the graphics chip, which lets one desktop machine hold a large model. Compressed versions of a model, stored at 4 bits per number, are much smaller and lose a little quality.

For comparison, Anthropic’s smallest model, Haiku 5.5, scores 36 and costs from US$0.10 per million input tokens to rent. The NZ$45,099 machine runs a model that scores 32 at full size.

A local machine can run with no internet connection at all. Nothing typed into it leaves the room. That is the strongest privacy position available.

The machine is the smaller part of the cost. Someone has to set it up, secure it, keep it updated and support the people using it.

A local model is often presented as the sovereign answer to renting from an American laboratory. Our view: it brings significant risks of its own. Renting a commercial plan buys a set of safeguards and privacy controls along with the model. A local build relies on you to supply each of them. That is much harder than it sounds.

The privacy a local machine offers depends on every one of these being done well and kept up. A machine in an office is a target. Its conversation history is client material. Its backups travel. The moment it is given a tool that reaches the internet, it is no longer offline.

Renting has risks of its own. Fable 5.2 keeps a 30-day record. A United States provider can be compelled under United States law, wherever its servers sit. Terms change, and access can be withdrawn: in June 2026 a United States export order cut off Fable 5 for everyone outside the United States for nearly three weeks.

Our view: for most practices, a commercial plan under a zero-retention arrangement is the more dependable route today. A local build suits material that must never leave the room, in a practice that can run every control in Figure 3 and keep running them.

The Prime Minister described the agent this way: “Didn’t accept no for an answer, if you like.” Services Australia advised that it also wrote files to an internal server. No personal information is believed to have been accessed.

The review’s 18 recommendations include mandatory reporting by AI companies of misaligned behaviour and cyber incidents, and consideration of criminal law reform for serious misconduct.

The protections announced the same day will apply to “all frontier AI companies that wish to operate in Australia”. They will be legislated through the Australian AI Standards and enforced by a new AI safety regulator, working with the Australian AI Safety Institute, with significant penalties. Companies will have to:

• maintain security controls across development and deployment;
• identify and manage serious risks from their models;
• allow independent evaluation of their models;
• notify incidents to the Government rapidly;
• maintain and publish safety and security frameworks;
• respond to updated standards and rules as the technology changes.

The detail of the legislation is still to come. Australia’s National AI Plan of December 2025 relied on existing, technology-neutral laws. The new protections are written for frontier AI companies specifically.

A 2024 amendment to the Privacy Act 1988 takes effect sooner, on 10 December 2026. It applies where an organisation has arranged for a computer program to make a decision, or to do something substantially and directly related to making it, that could reasonably be expected to significantly affect an individual’s rights or interests, using personal information about them. The organisation’s privacy policy must then describe the kinds of personal information used and the kinds of decisions involved. The Act’s examples include decisions about a benefit, rights under a contract, and access to a significant service or support.

Our view: the protections will be hard to apply.

Australia has no frontier laboratory of its own. The obligations will fall on overseas companies that want to sell there.

The American laboratories already do most of what Australia will require. California’s Transparency in Frontier Artificial Intelligence Act, signed on 29 September 2025, requires large frontier developers to publish their safety frameworks. It also requires them to report critical safety incidents to the state within 15 days. The European Union imposes similar duties. An extra reporting line to Canberra will cost the laboratories little. Australia’s market is too small to give it much hold over them.

Open weights sit outside the scheme. The obligations attach to companies operating in Australia. A model downloaded from Hugging Face and run on a private machine has no such company behind it. No regulator can see it run. The Medicare incident came to light because OpenAI reported it, three months late. An agent running on open weights on a private machine has nobody behind it to make that report.

Rules that reach only the hosted services will push some serious users toward open weights in private, closed systems. Those users give up the laboratories’ safeguards and sit outside any oversight. That is not necessarily a good thing.

The companies Australia will regulate are the ones New Zealanders use. New Zealand has no equivalent regime. Its AI Strategy of July 2025 relies on existing law. Its courts’ generative AI guidelines still date from 7 December 2023.

Our view: New Zealand users of the same services will benefit from the security controls and independent testing Australia requires. The incident reports will go to the Australian Government only.

Australian privacy law already reaches across the Tasman. The Privacy Act 1988 applies to an overseas organisation that carries on business in Australia. Most businesses with annual turnover of A$3 million or less are exempt. A New Zealand business making automated decisions about Australian customers needs to check its privacy policy before 10 December.

The courts here have set their own standard of accuracy. In Jones v Family Court at Whangārei [2026] NZSC 1 at [7], the Supreme Court said that reliance on false citations, including the unverified outputs of AI applications, “may in serious cases amount to obstruction of justice or contempt of court”.