AI Opinions: June 2026. Fable 5 Hands-On, Billing Split Countdown, OpenAI Banks Your Resets
Three weeks of building instead of writing. Here is everything that piled up in the meantime, including a few days with Anthropic’s new top model.
First, some honesty
I owe you a bit of context before this one. For the last three weeks I have been posting at least once a week, but I was so deep into doing things that I got completely lost in experimentation. New things, broken things, rebuilt things. I had no extra time to turn any of it into a proper post, and I have always tried to be transparent about that trade. I focus on doing things rather than writing about them. The writing is a byproduct of the doing, and I would rather keep it in that order.
So no promises here. I am not announcing a new cadence or turning over a new leaf. I keep building, I keep trying to send you the most valuable parts from my point of view, and you are always invited to discuss it and build along.
This one is a catch-up post. Several things happened at once and I had no opportunity to address any of them. Treat it as the light version of my posts, in the same spirit as the April opinions roundup. I hope you will like it.
Oh and BTW. THANKS FOR 3200 people reading me! Greatful to be here!
The billing split lands in three days
The thing closest on the calendar first. On June 15 Anthropic splits its subscription billing: programmatic use of Claude, so the Agent SDK, headless claude -p runs, GitHub Actions, anything spawned from a script instead of typed by a human, stops counting against your plan limits and moves to a separate monthly credit, metered at standard API rates.
One correction to my own earlier framing, because I went through the fine print again this week. I kept repeating “$200 per month” and that number is true for me, on the Max 20x plan. The credit is actually tiered: Pro gets $20, Max 5x gets $100, Max 20x gets $200. Enterprise Standard seats get exactly nothing. So when you do your own math, do it against your plan, because the spread between $20 and $200 is the difference between a toy budget and a real one.
The mechanics matter more than the headline number, honestly. Credits are per-user and cannot be pooled. Unused credit does not roll over. You have to claim it once in your account before June 15. And the part I would put in bold if I were Anthropic’s documentation team: when the credit runs out, your automated requests simply stop. No queue, no automatic downgrade to a cheaper model. There is an overflow toggle that bills the excess at API rates, and it ships turned off. The default behavior of this change is your agent going silent mid-month.
I have been working around all of this for a while and I wrote a dedicated post with the four mitigations I actually wired in: Anthropic repriced my agent, four mitigations before June 15. If the deadline affects you, that post is the practical one. This section is just the reminder that the clock is at three days.
And the painful detail, the reason it is hard for me to fit into even the $200 tier: programmatic use gets counted like API usage. Real token costs, not the heavily subsidized subscription math we got used to. One widely shared gist did the napkin math and called it a 12x to 175x effective price increase depending on workload. That spread sounds dramatic, but it roughly matches what my own measurements say: the same work, priced honestly, turns out to be expensive. That gap between subscription pricing and honest pricing is the whole story of this post, actually. Keep it in mind for the Fable section.
The obvious mitigation many skips
While optimizing my own agent for this, one thing kept proving itself over and over: smart model routing. It is very obvious and almost nobody does it. If you run Opus for everything, you will hit your ceiling embarrassingly fast. Route the same request to Sonnet and it costs a fraction. I wrote about the day I switched my agent from Opus to Haiku and it got better, and that lesson aged well. Haiku still handles a surprising share of the simple requests flowing through my system.
Below Haiku there is another floor: local models. The really simple jobs, classification, triage, formatting, run on a 35B model on my $600 Mac Mini and cost exactly nothing per token. I have been measuring this stack for months now and the routing is what keeps the whole thing viable. After June 15 it stops being an optimization hobby and becomes the difference between an agent that runs all month and one that dies on the 19th.
And if your agent is model-agnostic, you are not even locked to one vendor. Other providers can supplement the expensive paths; my $20 OpenRouter fallback exists for exactly this kind of squeeze. Everything is possible here. It is all up to our creativity, I think. If you want the shortcuts I use for this, the agent playbooks on the Wiz store cover the routing setup in detail.
OpenAI is fighting dirty, and I like it
Now the fun one. Yesterday OpenAI announced on X that Codex users can now save their rate limit resets and spend them later. Their words: they heard we wanted to use resets on our own time. Every eligible account on Go, Plus, Pro and Business got one banked reset for free, and for two weeks Plus and Pro users can invite up to three friends, with both sides earning another reset when the friend sends their first Codex message.
The fine print is pure growth hack. The reward lands only when the invited friend actually starts using Codex, an invite alone earns nothing. Banked resets expire after 30 days. It is loyalty points for compute, complete with a referral program. I am half joking and half impressed, because the underlying feature is genuinely user-friendly: months of complaints about resets firing at fixed times, often in the middle of someone’s night, and OpenAI responded by handing the timer to the user.
Read the timing too. Anthropic is three days from making programmatic use more expensive, and OpenAI responds by making its limits more flexible and literally giving spare capacity away. This is a very direct shot in the subscription war, and it is totally something Anthropic would never do. One company is tightening the meter, the other is letting you carry your unused minutes to next month, like a mobile operator from 2005.
I have my history with Codex, including cancelling it and coming back once already, so I am watching this with sympathy for both sides. Competition like this is the only thing that keeps our subscriptions honest.
A few days with Fable 5
And then there is the new model. Anthropic shipped Claude Fable 5 on June 9, a new tier above Opus. The construction is unusual: there is a twin called Mythos 5, the same underlying model, and the difference is safety plumbing. Fable runs classifiers, and when a request looks like offensive cybersecurity, dangerous biology and chemistry, or an attempt to distill the model, the response silently falls back to Opus 4.8. Mythos skips the cyber guardrails and goes only to approved organizations, government cyber defenders and the like. Anthropic says more than 95% of Fable sessions never touch the fallback. For what it is worth, in a few days of agent work I have not knowingly hit it once.
I have been running Fable 5 with my agents since launch. Short version: it is good. Much better than Opus for agentic work, definitely. It is more proactive. It pushes forward with the tasks I give it instead of stopping at the first checkpoint to ask how I feel about things.
The thing I appreciate most is how deep it goes. This was always my quiet complaint about the Opus line: Opus tries to be specific and on the spot, and sometimes that means it misses the full context around the problem. It fixes the line you pointed at. That is also why I always liked Codex with GPT-5.5 a little more for audits, because it reads the whole module before it touches anything. Fable 5 works like that. It audits around the problem. It goes very deep into the code, it references things I did not mention, and it holds up over long runs better than anything I have used.
One concrete example. I gave it some architecture changes on my AI agent, the kind of task where Opus usually needs a round of feedback from me halfway through. Fable ran longer than Opus would have, and when I checked the outcome I had genuinely no notes. It did what I wanted and in a few places a little more. That almost never happens.
The benchmark numbers are loud, and I would hold them loosely. Anthropic claims state of the art nearly across the board: 80.3% on SWE-Bench Pro against 69.2% for Opus 4.8, and a story about Stripe migrating a 50-million-line Ruby codebase in a day, work estimated at over two team-months. Vendor numbers, vendor anecdotes. For balance, Endor Labs ran an independent security benchmark where Fable came out middling, and they documented benchmark contamination in 19% of their test instances. My hands-on lands somewhere between the marketing and the skepticism: the long-run agentic improvement is real, I can feel it in my own work. The superlatives I leave to the launch page.
Now the catch, and you already know where this is going: the cost. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. That is double the Opus sticker price, although to be fair it is also less than half of what the Mythos Preview was going for. Right now it runs inside our subscription usage limits, but only until June 22. On June 23 it leaves the subscription pool and you pay for it with usage credits. Anthropic says the removal is temporary and that they aim to bring it back into plans when capacity allows. We will see. The internet is predictably furious about the two-week tease, and I get it: handing everyone the best model for free and then putting it behind a meter is a very effective way to make people feel the gap.
And it burns tokens fast, faster than anything I have seen, because being thorough is exactly what costs tokens. The depth I praised three paragraphs ago is the same property that empties the budget.
So my take: Fable 5 is a specialist tool, at least at this price. Security work, maybe. Big architecture passes, long autonomous builds, the tasks where one excellent run beats five cheap ones. For most jobs Opus remains my default, and the routing logic from the section above does not change. It just gets one more expensive tier at the top to route to, sparingly.
Where I think this is going
Everything is getting pricier at the top while our subscriptions still look generous. I do not think that lasts.
Look at the line Anthropic actually drew with the billing split, because it is sharper than “interactive versus automated.” Chatting on claude.ai stays flat-rate. Interactive Claude Code in your own terminal stays flat-rate. But run Claude inside Zed, a fully interactive session with a human typing every prompt, and it bills against the credit anyway, because it arrives through the Agent SDK. The meter follows the integration surface, not the human. First-party surfaces stay subsidized. Everything you build or plug in yourself becomes usage.
That is the pattern I see everywhere now: the soft cutoff. In the billing split, in Fable 5 leaving the subscription pool after two weeks, even in OpenAI gamifying its limits with expiry dates. Nobody will take your flat-rate plan away. They will just keep moving the best things outside of it, one model and one surface at a time.
Although that sounds gloomy, I do not really mind it. Honest prices force better engineering, and better engineering is the part I enjoy. The builders who learn to route work to the right model will keep their costs flat while the quality ceiling rises. That is the game now, and it is a game you can actually win.