Claude has a setting many people leave enabled by default.
It allows Fable → Opus fallback when a safety check is triggered. You can disable it in 3 places:
1. Chat
Open Settings → Capabilities
Find Switch models when a message is flagged
Turn the toggle off
2. Claude Code
Open Settings → Claude Code
Turn the same toggle off
3. Console
Run:
/config switchModelsOnFlag=false
If the safety check is triggered, Claude will stop the chat instead of continuing the response on Opus.
Claude has a setting many people leave enabled by default.
It allows Fable → Opus fallback when a safety check is triggered. You can disable it in 3 places:
1. Chat
Open Settings → Capabilities
Find Switch models when a message is flagged
Turn the toggle off
2. Claude Code
Open Settings → Claude Code
Turn the same toggle off
3. Console
Run:
/config switchModelsOnFlag=false
If the safety check is triggered, Claude will stop the chat instead of continuing the response on Opus.
Fable 5 got jailbroken again
Researcher Vitto Rivabella tested Fable 5’s defenses and managed to find a bypass.
According to him, most attempts failed. The protection is multi-layered: the model checks the prompt, conversation history, system context, and its own response.
Some filters run during generation and can stop the answer halfway through.
The checks are not based on keywords. The system looks at meaning, intent, language, wording, and suspicious chains of requests.
The bypass took around 20 hours. It required rare languages, academic framing, long build-ups, Unicode, breaking the task into parts, and working with the chain of thought.
The author did not get a stable bypass for long tasks. According to him, regular search is faster and cheaper.
Fable was swapped
Claude Fable 5 was tested again in BridgeBench after its return. The results dropped sharply.
Debugging: 86.2 → 25.9
Refactoring: 73.6 → 38.4
Hallucination: 75.9 → 61.7
When tasks pass the safety filters, the model performs like the June 12 version.
The main problem is the new filters. They too often classify coding tasks as risky and switch execution to Opus 4.8.
OpenAI may give 5% to the US government
According to FT, OpenAI is in early talks to give a 5% stake to the US government.
Sam Altman’s idea is to give citizens a financial stake in AI growth. One model being discussed is the Alaska Permanent Fund: the government receives an asset, and income can be distributed to people.
The plan could also affect other major US AI developers: Anthropic, Google and Meta. They have not agreed to it yet.
The talks are still conceptual. The deal may require approval from Congress.
AI already automates 16% of freelance tasks
CAIS updated the Remote Labor Index — a benchmark where AI agents complete real client projects: design, video, CAD, architecture, analytics, and web apps.
The work is reviewed by humans and compared against a professional deliverable that a paying client would accept.
New results:
• Claude Fable 5 — 16.1%
• Claude Opus 4.8 — 8.3%
• GPT-5.5 — 6.3%
Fable 5 is now the top model in RLI and almost 2x higher than Opus 4.8.
When the benchmark launched, the best result was 2.5%.
Claude Desktop is now on Linux
Anthropic has released Claude Desktop for Linux in beta.
Ubuntu and Debian are supported for now.
The app includes Claude Code, Claude Cowork, and regular chat across all paid plans.
Computer Use is not included in this release yet. Anthropic says more is coming soon.
Sonnet 5 looks absurdly priced
• Opus 4.8 Max - $1.80
• Sonnet 5 Max - $2.29
• Fable 5 - $2.75
Sonnet 5 is about 27% more expensive than Opus 4.8.
Why does the smaller model cost more than the flagship?
Google released Nano Banana 2 Lite
The new model generates an image in about 4 seconds and costs $0.034 per 1K image.
In Gemini, it is available by selecting 3.1 Flash Lite.
The main use case is fast drafts: run several prompts, pick the best result, then send it for refinement in Nano Banana 2 or GPT Image.
In terms of quality, Lite is almost next to the full version: 1251 Elo vs 1270 for Nano Banana 2 on lmarena.
Claude Fable is back!
Anthropic has lifted restrictions on Claude Fable 5 and Mythos 5 after a decision by the US government.
Fable will be available to Claude users again starting today.
Until July 7, the model is included in weekly limits. After that, access will move to API.
The restriction started after an Amazon report: researchers found a way to bypass Fable’s safeguards in cybersecurity tasks.
Anthropic says the bypass did not unlock Mythos-level capabilities. The company has already added a new safety classifier that blocks this method in more than 99% of cases.
Anthropic is also preparing a common standard for evaluating AI jailbreaks.
Job titles no longer work - that’s how Boris Cherny, Head of Claude Code sees it
His point is that the old split between engineers, designers, product managers, and data scientists is becoming too coarse.
At Anthropic, he sees how the rise of AI is starting to blur the boundaries between roles.
In practice, a person’s value is not defined by their title, but by how they move the product forward.
Boris suggests looking at a team through five archetypes:
• Prototyper
Finds new ideas and quickly turns them into early versions.
• Builder
Takes an idea and turns it into a real product.
• Sweeper
Simplifies, removes what is unnecessary, and makes the system cleaner and faster.
• Grower
Takes a working product and improves it step by step.
• Maintainer
Owns the reliability, security, and scalability of a mature system.
It’s a useful frame because it explains reality better than job titles.
A team may have many engineers, but still lack a Builder.
It may have strong design, but lack a Sweeper.
It may have a product manager, but no Grower.
And the opposite is also true: someone may formally belong to one function, while actually covering a completely different role for the product.