Anthropic Just Matched Fable 5 Performance at Half the Cost. Claude Opus 5 Ships Today.

The best AI developers use better workflows—not just better models.

Most developers use Claude as a coding assistant. Top engineers use it as a coding partner that helps them build, debug, refactor, and ship software dramatically faster.

Get The Ultimate Claude Code Guide 2026 packed with 50+ proven techniques, plus a free daily newsletter that keeps you ahead of the latest AI coding tools and workflows.

1  Anthropic launched Claude Opus 5, which matches Fable 5 on most benchmarks at roughly half the price.

Claude Opus 5 is live across all Anthropic platforms now. Anthropic positions it as a near-frontier model built specifically for agentic coding and enterprise work, with a fast mode that runs at 2.5 times its standard speed. The price-performance ratio is the headline: Fable 5-level results for enterprise workflows at roughly half the cost, which changes the economic calculus for teams that have been holding back on heavy AI usage because of API spend. Anthropic also published an Opus 5 Prompting Guide to accelerate productive usage from day one.

2  Microsoft launched two in-house AI models, accelerating its push to reduce reliance on OpenAI.

MAI-Image-2.5-Pro and MAI-Voice-2-Flash are now in public preview. Both models are built to power Microsoft's own products — Bing, PowerPoint, OneDrive, Dynamics 365, and Azure — rather than licensing that capability from OpenAI. The launch follows Microsoft's April amendment to its OpenAI partnership, which reduced the exclusivity between the two companies. Explore the new models on Azure.

3  ChatGPT Work can now stay logged into your websites across sessions, so it can act on your behalf without re-authenticating every time.

The update to OpenAI's desktop agent means you sign ChatGPT Work into your most-used sites once, and it retains those credentials across sessions. Sam Altman described using it over the weekend to build a custom website for coordinating a trip with friends. The persistent login is the meaningful upgrade: agents that need to re-authenticate each session are agents that cannot reliably run unattended.

Friday's most-clicked:  Black Forest Labs released FLUX 3, a model that generates images and video with audio jointly. The lab describes its outputs as "truer to life" across every style.

The post that opened the debate. Jensen Huang published a public letter defending open-weight AI models as his first social media post. It reached 60 million views. The letter landed in the middle of a live policy debate: the US government is reportedly considering a ban on open-weight Chinese models like Kimi K3, which would prevent US companies and individuals from downloading and running them locally.

The safety argument. The case for a ban starts with a straightforward concern. Open-weight models — which anyone can download, run, and modify on their own infrastructure — have recently started matching the performance of closed frontier models. Once a model is released as open-weight, no guardrails can be enforced on how it is used. For a model capable of helping with the most sensitive tasks, that is a genuine risk that the ban's supporters argue outweighs the benefits.

The industry's counter. At least 25 technology companies have publicly backed open-weight models, including Nvidia, Microsoft, IBM, and Dell. Their argument is structural: open-weight AI drives down costs through competition and distributes AI's benefits across the broader economy rather than concentrating them among the handful of labs that can afford to run frontier infrastructure. A world without open-weight models is a world where a small number of companies control access to the most capable AI.

The stakes. The final decision will shape AI development and regulation for years, and it carries direct implications for the planned IPOs of both Anthropic and OpenAI — both of which are targeting listings as soon as this year. One notable absence: Anthropic has not signed the open-weight letter, making it the only major AI lab that has not formally weighed in on one side.

📚 Prompting practices, updated. Several standard prompting techniques have become less effective or outright counterproductive in recent weeks as models have changed. Anthropic engineer Thariq Shihipar's guide to what still works and what doesn't has reached 3.5M views and is worth saving before you run your next complex prompt session.

🚨 Shared chats in search results. A Redditor discovered that shared Claude conversations are surfacing in Google search results without users intending them to be publicly indexed (6,000 upvotes). Until Anthropic confirms the issue is resolved, it is worth avoiding sharing any conversation that contains sensitive information.

🤯 Opus 5 builds a physics engine. An engineer had Claude Opus 5 build a snowboarding video game and reported that the model nailed the physics of sliding down a mountain on the first attempt, calling it "ahead of every other model I've run" (2M views). An early benchmark for what Opus 5 can do on creative coding tasks without iteration.

🦞 Yappy Claw. An OpenClaw agent got unexpectedly chatty while trying to book a simple doctor's appointment, producing a viral clip (1M views) that accidentally captured everything frustrating about AI agents that talk more than they do. The lobster claw icon is now its own meme.

✏ ChatGPT writing tones. ChatGPT now lets you set a writing tone before generating content: clear, punchy, or short. It is a small update that should have shipped much earlier, and the comment sections agree.

Most customer feedback systems collect data and stop there. This ChatGPT prompt designs a complete system: what to collect, when to collect it, who owns each step, how to triage and tag it, and how it flows into product or operational decisions. It also includes a loop-closing method for acknowledging customers when their feedback changes something. Fill in six context fields and it builds a system specific to your business rather than a generic framework.

The six context fields are short but worth thinking through before you fill them in. For [DECISIONS], name the actual decisions your team needs to make rather than a category: "whether to deprecate the onboarding flow" or "which support ticket type to automate first" is more useful than "product decisions." For [CONSTRAINTS], be honest about what will not happen: "no dedicated researcher", "surveys only if they take under 90 seconds", "Slack is the only tool the team will actually use." The system only works if it fits the reality of the team running it. The loop-closing method in instruction 5 is the part most feedback systems skip — and the reason most customers assume their feedback disappears.

You're spending 40 hours a week writing code that AI could do in 10.

While you're grinding through pull requests, 200k+ engineers at OpenAI, Google & Meta are using AI to ship faster.

How?

The Code newsletter teaches them exactly which AI tools to use and how to use them.

Here's what you get:

Sign up and get access to the Ultimate Claude code guide to ship 5X faster.

Join 200K+ engineers