This week feels like one of those moments where the AI coding game quietly moves up a level.
Not because the model got a cooler name but because the shape of the work it can take on is changing.
π₯ The Big One
Anthropic just brought Mythos-level power to regular builders.
Anthropic released Claude Fable 5 and Claude Mythos 5 this week, and if you're a developer, this is one to pay attention to.
The short version: Fable 5 is the public version of Anthropic's Mythos-class model. Mythos 5 is still reserved for a smaller trusted-access group, especially around cybersecurity and life sciences. Fable 5 uses the same underlying model but adds stronger safeguards, including fallback routing for sensitive cyber and bio requests.
The part that matters for us is not just "new model dropped."
It is that Anthropic is positioning Fable for long-running, ambitious coding projects: large migrations, complex implementations, multi-stage agent work, and projects where the model plans, tests, checks its own work, and keeps going for longer than previous Claude models.
"Our next generation of intelligence for the hardest knowledge work and coding problems." β Anthropic
Read the full Anthropic announcement β
So what should devs actually do with it?
Do not start by giving Fable your whole codebase and saying "make it better."
That is still how you create chaos.
Start with bounded agent tasks where you can judge the result:
-
Refactor one messy module β ask for a plan first, then let it change one file or feature area.
-
Write tests for an existing feature β this is a perfect agent task because the output is easy to verify.
-
Migrate one dependency or API route β give it the docs, the current code, and a narrow success condition.
-
Build one internal tool β dashboards, scripts, admin panels, content helpers, QA utilities.
-
Review a PR like a senior engineer β ask it for risks, missing tests, edge cases, and simpler alternatives.
My personal rule would be simple:
If the task needs taste, product judgement, or security-sensitive decisions, keep Sonny in the loop.
If the task is repetitive, testable, and annoying, let the agent cook.
"The skill is not prompting anymore. The skill is scoping the work so the agent can actually succeed." β Sonny
What else is going on
1. Vercel says low-cost models are finally entering production
Vercel's June AI Gateway Production Index has a useful signal for builders: teams are not only testing cheaper models anymore, they are actually routing production traffic to them.
DeepSeek V4 went from almost nothing to 17% of AI Gateway token volume in May, while its cost share stayed around 1%. That tells me teams are getting more serious about model routing: frontier model for the hard stuff, cheaper model for the repeatable stuff.
That is exactly how I would build AI products now.
Do not marry one model. Build a routing layer.
"Teams are still increasing token budgets, but they are implementing smarter routing strategies." β Vercel
Read the full Vercel index β
2. GitHub is turning Copilot into a proper agent workspace
GitHub's new Copilot app is another sign that AI coding is moving away from "autocomplete inside the editor" and toward agent-native workspaces.
The big idea is simple: agents should work where your code, issues, PRs, terminals, and review flow already live.
For dev teams, this matters because the winner will not just be the smartest model. It will be the tool that fits into your real workflow with the least context switching.
"Agents can work the way you already work." β GitHub
3. Copilot usage-based pricing is now active
GitHub also switched Copilot plans over to usage-based pricing on June 1.
This is not the fun headline, but it is important.
The more agentic these tools become, the more developers need to understand what is happening under the hood: model choice, context size, cached tokens, retries, long-running tasks, and how often an agent loops before finishing.
AI coding is becoming a productivity tool and an infrastructure cost.
So the devs who win will not just ask "which model is best?"
They will ask:
Which model is good enough for this task, and how do I keep the workflow efficient?
"As of June 1, all Copilot plans bill based on GitHub AI Credits consumed." β GitHub
π§° Worth your time
-
Claude Fable 5 product page β The use cases are worth scanning, especially the sections on agents, coding, enterprise workflows, and safeguards.
-
Claude Mythos 5 product page β Useful if you want to understand why Anthropic is separating public access from trusted-access use cases.
-
Codex for every role, tool, and workflow β OpenAI is pushing the same broader trend: coding agents are becoming workflow agents.
My weekly message to YOU!
This week, do not just test a new model with a random prompt.
Give it a real engineering task with a clear finish line.
Something like:
"Review this feature, identify the top 5 risks, write the missing tests, and explain every change before editing."
That one prompt teaches you more about AI engineering than watching ten launch videos.
Because the future of coding is not just knowing which model came out.
It is knowing how to turn these models into reliable teammates.
Are you using AI agents for real coding work yet, or still mostly using chat / autocomplete?
Hit reply - I want to know what is actually working for you. I read every single one. Talk soon PAPAFAM!
Sonny ππΌ
ππ½ Don't forget to follow me across socials!
Responses