A month ago I wrote that GPT-6 Astra made me resubscribe to Codex. I meant it. I am an Anthropic guy, and I was still hitting the 5-hour limit often enough to plan my day around it. Some weeks I hit the weekly one too. I wanted things done, and the subscription kept telling me to wait.
Since then Anthropic shipped Opus 5.5 on September 22nd, Sonnet 5.5 on September 28th, and Haiku 5.5 on October 7th. Three models in fifteen days, plus a pricing change and a subscription perk I did not see coming. I have been living with all of it, and my honest read is short: this is the best Anthropic lineup I have used, and I stopped thinking about leaving.
Opus 5.5 is my daily driver now
Opus 5.5 does my agentic work. Building, fixing, orchestrating, long sessions in Claude Code. I run it at medium or high reasoning effort and I rarely feel the need to go higher, which is new for me. With Fable 5 I spent weeks learning that max effort was a trap, and I kept treating the effort dial as something I had to babysit.
It is also cheaper than the Opus before it. Opus 5.5 lists at $4 per million input tokens and $20 per million output. Opus 5 was $5 and $25.
The feeling is familiar. Almost a year ago, Opus 4.5 and then Opus 4.6 did something similar. Claude Code suddenly clicked. Coding and agentic work stopped feeling like a demo and started making sense as a way to work every day. I loved that period. Opus 5.5 gives me the same feeling again.
Fable 5.1 is the one I barely open
Fable 5.1 is still the strongest model in the family at some things. For design it sometimes lands a better result than Opus, and I still try it for that from time to time.
The price is $10 per million input tokens and $50 per million output, two and a half times Opus 5.5 on both sides. In my September notes I had already seen Fable 5.1 spend more tokens per job than its predecessor. When the quality gap on most of my tasks is small and the price gap is 2.5x, the math does itself. Since Opus 5.5 dropped, I have almost stopped using Fable.
Sonnet 5.5 keeps taking more of my work
I did not expect to write this after the previous Sonnet, which honestly felt like a mess next to Opus. Sonnet 5.5 is really good at a lot of things, and I use it far more often than I planned to.
Pricing helped. On October 7th Anthropic halved Sonnet 5.5's cache-read price, from $0.20 to $0.10 per million tokens. The headline input and output prices stay at $2 and $10. Cache reads sound like a footnote, but an agent rereads the same context all day, and Anthropic estimates the cut makes typical agentic tasks about 20% cheaper. For the long-running jobs my agent does, that is where the bill actually lives.
Opus is still the best for me. Sonnet is what I reach for when the best is more than the job needs.
I was wrong about Haiku
I have been joking about Haiku for a long time. I even built stale.jock.pl, a page that counts how long ago each model shipped and when it stopped reading, and Haiku 4.5 sat near the top of that shelf: released October 15th, 2025, training cutoff July 2025. Only Llama 4 looked older. (If you want to know why that matters even for models with a search button, I asked 16 of them who the king of Norway is.)
So I had a theory. Anthropic was pushing prices and models up. Fable would become the new top tier, Opus would become the workhorse, Sonnet would slide down into the slot Haiku used to fill, and Haiku would eventually be deprecated. Tidy theory. Wrong theory.
Haiku 5.5 is cheap, fast, and good, and Anthropic's launch numbers back that up:
$0.10 per million input tokens and $0.50 per million output, for prompts under 100k tokens. Haiku 4.5 was $1 and $5.
Above 100k tokens the price jumps 5x, to $0.50 and $2.50. Keep its context small and it stays absurdly cheap.
72.4% on OSWorld 2.1 (offline subset), a computer-use benchmark. Haiku 4.5 scored 15.7%. Sonnet 5.5 sits at 83.9%.
Adjustable effort, a first for a Haiku, with medium as the default.
Those are vendor numbers, so read them as a direction. Even as a direction, a jump from 15.7% to 72.4% is hard to argue with.
And yes, on a Max plan Haiku feels close to unlimited. Honestly, I care much more about what happens when you pair it with the big models.
Haiku as the hands, Opus as the head
I am not going to use it to write code, because writing the code is rarely the expensive part. The expensive part is gathering context, reading piles of material, verifying, categorizing. So I told Opus to orchestrate and to hand every subagent job to Haiku, and Opus.
That matches how I think about agent architecture. Back in February I moved parts of my agent from Opus to Haiku and it got better, because most agent work is plumbing, and plumbing does not need deep reasoning. Haiku 4.5 just was not good enough to push that idea far. Haiku 5.5 is.
So my suggestion, whether you run a framework or a custom setup like Wiz:
Opus (or Sonnet) plans, decides, and reviews.
Haiku executes: lookups, summaries, classification, file sweeps, compaction, routine tool calls.
Try it on computer use. Speed plus the biggest benchmark jump in the whole release.
Leave a little reasoning on. Theo mentions a developer who ran it with reasoning off and got a worse pass rate than OpenAI's small model. Anthropic's models are trained to think a bit first.
Keep Haiku's context under 100k. Past that line the price multiplies by five.
Your workflow gets faster and cheaper at the same time, which almost never happens.
The $200 I did not ask for
One more thing shipped with Haiku, and for me it might matter most.
I use OpenRouter for a lot of internal things. One place, nearly every model, one bill. I like it and I keep recommending it, because when you work through an API you do not want an account with every single provider. But it is a separate cost on top of my subscription, and I kept thinking: what if the subscription just came with some API credit?
Now it does. Max subscribers get monthly API credit: $100 on Max 5x, $200 on Max 20x. Team plans get a pooled amount, capped at $500. For me that is $200 every month to spend on any Claude model through the API.
Read the terms before you link anything. The credit works for the Claude API, the Agent SDK, Managed Agents and the Playground, through an API key from the Console organization you connect to the plan. Interactive Claude Code stays on your plan limits, and those limits do not go up. Unused credit disappears at the end of each billing cycle. You connect exactly one Console organization from your claude.ai billing settings, and changing it later means writing to support, so pick carefully.
Even with those limits, it changes my default. The next time I build something that needs an API, I will reach for this first, because $200 is already sitting there. With Haiku at ten cents per million input tokens, $200 goes a very long way.
I know this is lock-in, and I can see it clearly. I am happy to accept this kind, because it puts something on top of what I already pay. After May, when Anthropic repriced my agent with the Agent SDK billing split, this feels like the opposite move. Theo says Anthropic talked to developers before shipping it, and you can tell.
Codex stays, but as the second opinion
I am keeping the Codex subscription. Astra is still great, OpenAI's pricing still makes many of its models feel close to unlimited, and a model trained somewhere else catches things Opus misses. A month ago I wondered whether it would end up as my main tool. After the past few weeks of Anthropic releases, I am honestly impressed, and that question is closed for now.
My plan for this week: Opus 5.5 at medium effort as the head, Haiku 5.5 as the hands, and the $200 credit for the first small API project I would normally have sent to OpenRouter.
If you want the measuring side of this, the Agent Efficiency Kit shows where your Claude Code tokens actually go, with an optional Haiku audit built in.