Ali Asif

The AI log, frontier · open weights · agents and coding

Covers 8 Jun 2026 to 14 Jun 2026

Claude Fable 5 and Mythos 5 arrive, plus Kimi K2.7 Code

Anthropic launches a Mythos-class model for general use at $10 and $50, Moonshot releases an open coding model, and agent exams multiply.

By , published

What shipped

  1. Anthropic

    Claude Fable 5 and Claude Mythos 5

    A Mythos-class model made safe for general use at $10 and $50 per million tokens, with the more cyber-capable Mythos 5 limited to Glasswing partners.

  2. Moonshot AI

    Kimi K2.7 Code

    An open coding-focused follow-up to K2.6 under a modified MIT license, released June 12.

  3. OpenAI

    Codex app 26.608 and Import from Claude Code

    June 9 added flows that import setups from Claude Code and Claude Cowork, with a revamped plugins screen.

  4. Anthropic

    Scheduled deployments and vault credentials for Managed Agents

    Hosted agents can now run on a schedule and read credentials from environment variables stored in vaults.

  5. arXiv

    Agents' Last Exam

    A new hard benchmark for agents, first on the week's Hugging Face papers list.

  6. arXiv

    Workflow-GYM

    Long-horizon computer-use tasks drawn from real professional fields.

  7. arXiv

    MiniMax Sparse Attention

    The attention method MiniMax built its M3 model around.

Anthropic took the week on Tuesday, and then a banner on its own launch page complicated things on Friday.

Fable 5 and Mythos 5

On June 9 Anthropic launched Claude Fable 5 and Claude Mythos 5, both described as Mythos-class, a tier above Opus. Pricing is $10 per million input tokens and $50 per million output, which Anthropic says is under half the price of Mythos Preview. Fable 5 is available through the API as claude-fable-5 and across the apps. Mythos 5 goes to Project Glasswing partners with cyber safeguards lifted, and soon to select biology researchers.

The safeguard design is the unusual part. Classifiers route cybersecurity, biology and chemistry, and distillation requests to Opus 4.8, and users are told when that happens. Anthropic says fallbacks fire in under 5% of sessions on average. Fable 5 requires 30-day data retention and is not available under zero data retention, and the model uses a tokenizer that produces roughly 30% more tokens than models before Opus 4.7.

The evidence in the post is mostly customer stories. Stripe says a 50-million-line Ruby migration finished in a day, work it estimated at over two months by hand. Hebbia reports its highest finance-benchmark score. Treat those as partner claims. The page also carries a banner saying access to both models was suspended on June 12. The post does not explain why, so I will not guess.

Kimi K2.7 Code

Moonshot released Kimi K2.7 Code on June 12 as an open coding model under a modified MIT license, a follow-up to K2.6. The company reports a gain over K2.6 on its own Kimi Code Bench v2 with fewer reasoning tokens. I would wait for independent runs.

Tools and tests

Codex app 26.608 added flows that import setups from Claude Code and Claude Cowork on June 9. Anthropic added scheduled deployments and vault-stored credentials to Managed Agents the same day. Three papers stood out: Agents' Last Exam, Workflow-GYM for long-horizon computer-use tasks in professional fields, and MiniMax Sparse Attention.

What I'd watch next: whether access to Fable 5 comes back quickly, and whether the fallback rate stays under 5% for people doing security work.

Sources

Topics: Anthropic, OpenAI, arXiv, Codex, Kimi, Moonshot AI, Fable 5