This week in AI: Sonnet 5, Fable, and the politics of AI

Good morning,

Anthropic had a monster week by launching Claude Sonnet 5, rescuing Fable 5 from an export freeze, and dropping Claude Science, while OpenAI floated giving the U.S. government a 5% stake alongside a rigorous new biology benchmark, Google aggressively undercut media generation pricing, and Meta laid the groundwork to launch its own AI cloud business.

Let's dive in!

Anthropic drops Claude Sonnet 5

Anthropic’s newest mid-tier model is here, targeting heavy-duty coding, agent workflows, and professional tool use.

The strategy is clear. They want to migrate the complex, agentic capabilities that used to require massive frontier models down to a cheaper, faster tier.

• Sonnet 5 is now the default model for Claude Free and Pro users.
• It is live across Claude Code and through the Claude API.
• Introductory API pricing sits at $2 per million input tokens and $10 per million output tokens until August 31.
• After that, it moves to $3 input and $15 output per million tokens.
• Anthropic claims major leaps over Sonnet 4.6 in multi-step planning, browser navigation, and long-running execution, alongside tighter security against prompt injections.

The takeaway is about balancing capability with production costs. While teams prefer flagship models for R&D, running high-volume agent loops at that price is not sustainable. By hitting the agentic sweet spot at a mid-tier price, Sonnet 5 becomes the practical workhorse for scaling production.

Fable 5 is back online after export freeze

Following an abrupt U.S. government export control directive that forced Anthropic to suspend Fable 5 and Mythos 5 on June 12, the restrictions have been lifted.

The controls were removed on June 30, and Fable 5 rolled back out globally on July 1 across all major Claude platforms.

• Pro, Max, Team, and select Enterprise accounts get access to Fable 5 for up to 50% of their weekly usage limits through July 7.
• Beyond that, it shifts to a standard usage-credit model.
• Mythos 5 access remains restricted, returning only for vetted U.S. organizations.
• Anthropic is collaborating with Amazon, Microsoft, and Google to build a unified system for grading AI jailbreaks.

Claude Science brings an operating system to the lab

Anthropic also introduced Claude Science in beta. This is not a new underlying model, but rather a dedicated interface built to fit into actual scientific workflows.

Instead of typing prompts into a generic chat box, researchers get an environment that interacts directly with the tools they already use.

• It runs locally on macOS or Linux and connects to remote clusters via SSH or HPC login nodes.
• It ships with over 60 built-in scientific tools and database connectors spanning genomics, structural biology, and cheminformatics.
• It features specialized reviewer agents tasked with double-checking citations, tracking reproducibility, and vetting calculations.

By positioning this as a workbench rather than a biology chatbot, Anthropic is focusing on the unglamorous part of science. They are targeting messy data, traceable logs, and ensuring steps can actually be replicated.

OpenAI’s GeneBench-Pro targets "research taste"

OpenAI released GeneBench-Pro, a benchmark designed to test whether AI agents possess actual scientific judgment rather than just memorized facts.

The evaluation consists of 129 complex problems in quantitative biology, translational medicine, and genomics. Instead of multiple-choice questions, models are given raw datasets and ambiguous contexts. To succeed, the agent has to choose an analytical path, spot flawed data, pivot when assumptions fail, and reach a defensible decision.

The benchmarks show how much ground frontier models still have to cover.

• OpenAI’s top-tier model, GPT-5.6 Sol, scored a 28.7% pass rate on the highest reasoning tier.
• Turning on Pro mode only bumped that score to 31.5%.

Human experts typically spend 20 to 40 hours solving just one of these problems. The low AI scores highlight the current bottleneck in AI-for-science. Models are great at generating text, but they still lack the taste required to navigate complex, ambiguous research pipelines autonomously.

Google undercuts the media generation market

On the creative and media side, Google quietly launched two pragmatic, low-cost models via the Gemini API and Google AI Studio.

• Nano Banana 2 Lite: A fast, hyper-cheap image generation model that delivers images in under four seconds at $0.034 per 1,000 images.
• Gemini Omni Flash: A public-preview video generation and editing model. It handles up to 10-second clips using mixed text, image, or video prompts, priced at $0.10 per second.

Both models bake in SynthID watermarking by default. While this lacks the drama of a major LLM launch, it positions Google as the aggressive price leader for developers looking to scale generative media features without breaking the bank.

Quick hits 🗞️

• OpenAI eyes a Washington alliance. OpenAI reportedly floated giving the U.S. government a 5% equity stake. It is a clear sign that frontier labs are increasingly aligning themselves with national security interests. Source
• OpenAI prepares GPT-5.6 for broad release. Fresh signals in OpenAI's Codex app suggest public access to the GPT-5.6 family (Sol, Terra, and Luna) could drop as early as next week. App leaks show a new "reasoning effort" slider that will let developers dynamically balance speed against compute depth. Source
• Meta weighs entering the cloud market. Bloomberg reports that Meta is exploring "Meta Compute" to rent out spare AI infrastructure and host models like Muse Spark, introducing a massive new competitor to the AI cloud race. Source
• New Gemini Flash spotted on LMSYS. An unannounced Gemini Flash checkpoint appeared on LM Arena. Google hasn't confirmed it yet, but it matters because Flash handles a massive share of high-volume, everyday production workflows. Source
• Cursor puts coding agents on mobile. A new iOS app lets developers monitor cloud agents, review diffs, and approve code changes from their phones. Coding is rapidly shifting from manual execution to high-level supervision. Source
• Anthropic looks to custom silicon. The company is reportedly in early talks with Samsung to design custom AI chips as frontier labs rush to secure their own supply chains and curb soaring hardware costs. Source
• Arena hits a $100M ARR milestone. The evaluation platform's surging run rate proves that model benchmarking has officially graduated from casual public leaderboards into critical corporate infrastructure. Source

See you next week!

If this helped you get up to speed on AI fast, forward it to someone who'd find it useful. They can subscribe here.