The week in AI

The Anthropic launch and its suspension three days later took up most of the week. Fable 5 is the public version of Anthropic's top model, with safety classifiers that hand risky queries to an older model. Mythos 5 is the same model with those safeguards lifted, for a small set of partners. On 8 June, the day before the launch, Anthropic also published a study of how well its models turn security patches into working exploits.

Outside Anthropic, Z.ai gave its coding subscribers GLM-5.2 on 13 June. It is a 1M-token flagship that Z.ai says trails Claude Opus 4.8 by about 1% on FrontierSWE. Google DeepMind released DiffusionGemma on 10 June, an open model that writes blocks of text in parallel.

Anthropic calls Fable 5 its first public Mythos-class model and aims it at long-horizon agentic work. It has a 1M-token context window, up to 128K tokens of output, and adaptive thinking that is always on. The raw chain of thought, meaning the model's step-by-step reasoning text, is never returned to the user. Anthropic says the price is less than half that of Mythos Preview, the model's predecessor.

Fable 5 runs classifiers that watch for cyber, biological and chemical queries, and for attempts to distil the model. When one fires, the session falls back to Claude Opus 4.8. Anthropic reports this happens in under 5% of sessions on average and says it tuned the classifiers conservatively. An external bug bounty found no universal jailbreak in more than 1,000 hours of testing, by Anthropic's account, though the UK AI Security Institute made progress toward one in a brief early window. Business traffic on Mythos-class models now has a mandatory 30-day retention period.

Anthropic's capability figures are its own. With file-based memory, Fable 5 improved at Slay the Spire three times as much as Opus 4.8 did, and it reached the game's final act three times as often.

Mythos 5 replaces Mythos Preview and goes only to Project Glasswing partners, with vetted biology researchers to follow. Anthropic says its protein-design experts ran a drug-design process about 10 times faster with it. In blinded comparisons, its scientists preferred Mythos 5's molecular-biology hypotheses about 80% of the time over Opus-class models. Anthropic also says Mythos 5 built a single-cell genomics model 100 times smaller than one published in Science, and the smaller model outperformed it.

The directive covered all access by foreign nationals. Anthropic had no way to check a user's nationality in real time, so it turned both models off for everyone, US customers included. The trigger was a report from Amazon of a technique that bypassed the Fable 5 safeguards.

The models were still off when the week ended on 14 June. Anthropic gave no date for their return within the week.

The suspension follows a paper Anthropic published on 8 June about N-day exploits, meaning attacks on bugs that have been patched but not yet fixed on every machine. The model compares the published fix with the old code and works back to the bug. Mythos Preview, working on its own, built code-execution exploits from 8 of 18 Firefox patches. It also built full chains from a low-privilege user to SYSTEM for 8 of 21 Windows kernel patches, with no source code available. Public Claude models with safeguards turned off could also build exploits, though fewer. Anthropic told defenders to ship patches faster, because writing the exploit is no longer the step that needs scarce expertise.

GLM-5.2 is a mixture-of-experts (MoE) model, which sends each token through only some of its sub-networks. The parameter count comes from the Hugging Face model card. Z.ai says this is the first GLM to handle a full 1M tokens reliably. A method it calls IndexShare cuts compute per token by a factor of 2.9 at that length.

All scores are Z.ai's own. GLM-5.2 scores 62.1 on SWE-Bench Pro, up from 58.4 for GLM-5.1. It scores 91.2 on GPQA-Diamond and 40.5 on Humanity's Last Exam (HLE). On Terminal Bench 2.1 it gets 82.7 with its best harness. With the shared Terminus-2 harness it gets 81.0, against 85.0 for Claude Opus 4.8. Z.ai also says the model trails Opus 4.8 by about 1% on FrontierSWE.

Z.ai said the API and the weights, under the MIT license, would follow three days after the subscriber launch.

A normal language model writes one token at a time. DiffusionGemma starts each block of text as noise and refines all its tokens together over a few steps, the way image diffusion models work. Google built it by adding a diffusion head to a Gemma 4 base, using research from Gemini Diffusion.

Google reports up to 4 times faster inference on dedicated GPUs than standard Gemma 4 decoding. It labels the model experimental and aims it at interactive work on local machines. Google has not published how the speed trades against quality.

• Decart released Oasis 3 by API on 10 June. It is an interactive world model that generates photoreal driving environments from one prompt for closed-loop training of autonomous vehicles, three weeks after Decart's $300 million round led by Radical Ventures.
• Cognition published FrontierCode on 8 June, a benchmark built with more than 20 open-source maintainers that tests whether code is good enough to merge, and the best model scores 13.4% on its hardest tier.
• Moonshot AI released Kimi K2.7 Code on 12 June, a coding-only version of K2.6 that Moonshot says follows instructions better in long contexts and overthinks about 30% less.
• Google launched Gemini 3.5 Live Translate on 9 June, which translates speech to speech continuously across more than 70 languages and keeps the speaker's intonation, pacing and pitch.
• Luma Labs released Ray3.2 on 9 June with frame-level control over action, an API for its full set of controls, and HDR and EXR output for studios.
• Cohere released North Mini Code on 9 June, its first agentic coding model, a 30-billion-parameter MoE with 3 billion active parameters.