Claude Haiku 5.5 – Anthropic’s cheapest model, with a million-token window
The new Haiku took almost a year to arrive: Anthropic released Claude Haiku 5.5 on Wednesday 7 October, and the previous model, Haiku 4.5, on 15 October 2025. Claude Code has supported the new model since version 2.1.293, released the same day, and you switch to it with /model haiku. Anthropic calls Haiku 5.5 the cheapest and fastest small model it has released, and pitches it at large volumes of simple, repetitive work: summaries, sorting support tickets, database queries. Compared with Haiku 4.5, the new model has a context window of a million tokens instead of two hundred thousand, and it’s the first Haiku that lets you set an effort level – medium by default.
The pricing is the interesting part if you pay for tokens through the API rather than for a subscription. A request with a prompt of up to a hundred thousand tokens costs $0.10 per million input tokens and $0.50 per million output tokens, ten times less than on Haiku 4.5. Past that threshold the rates are five times higher, $0.50 and $2.50, and they apply to every token in the request, not just the ones above the line. Our guess is that Anthropic wants Haiku to work a bit like Jev, a model from TypeSafe AI that doesn’t write text but makes simple decisions fast: it classifies, routes a case onward, picks one of several paths. Short queries like that cost next to nothing on Haiku 5.5: five hundred requests of two thousand tokens each add up to a million input tokens, which is ten cents in the API, plus the rate for the answers.
Anthropic itself recommends Haiku 5.5 as a subagent working for Opus 5.5 and Sonnet 5.5. We read this as the direction Anthropic’s new models are heading: a main model, Opus for instance, runs the work and hands narrow tasks to subagents on Sonnet and Haiku. In our view, a small model for subagent work like this was missing from Anthropic’s line-up until now. A subagent whose definition already says just the alias haiku has been running on Haiku 5.5 since version 2.1.293. Claude Code doesn’t move any other subagent to Haiku, and on a subscription the built-in Explore subagent runs on Opus. If you want a particular subagent to work on Haiku, ask Claude Code in the chat to set its model to haiku.
Alongside Haiku 5.5, Anthropic halved Sonnet 5.5’s API rate for cache reads: from $0.20 to $0.10 per million tokens. Until now a cache read cost the same on Sonnet 5.5 as on Opus 5.5, which stays at $0.20.
How to give a subagent its own model, and why the bare alias haiku moves on to newer versions over time, is covered in the lesson Skill and subagent – splitting the work. When to reach for Haiku and when for Sonnet or Opus is the subject of our tip Haiku, Sonnet, Opus, Fable – which model for what.
A Claude Code subagent at the effort level you ask for
Since version 2.1.292, released on Tuesday 6 October, Claude Code runs a subagent at the effort level you ask for in the chat. Before that, the only way to give a subagent its own effort level was to set it in its definition, for good. Now you just add something like “run it at high effort” to your request, and Claude passes that level to the subagent for this one run, including when the subagent later resumes its work. The definition stays as it was.
Effort for a single run comes in handy with a subagent that normally does something simple and now and then gets a harder job – say, one that lists your new emails every day, and today has to compare two versions of a price list and point out what changed. It fits nicely with Haiku 5.5: the subagent works on a cheap model, and when something harder comes up, you raise its effort for that one time only.
A monthly API credit on the Max plan – not for working in Claude Code
Max subscribers now get a monthly credit from Anthropic to spend on the API: $100 on Max 5x and $200 on Max 20x. Anthropic started handing it out on 7 October, gradually, over several days, and new subscribers can claim it after seven days on the plan. The Pro plan doesn’t get the credit. The credit goes to a single organisation in the Claude Console – the one you link to your plan under Settings → Billing, in the API credits section.
The credit doesn’t raise your limit and doesn’t cover ordinary use of Claude Code – not in the terminal, not in VS Code, not in the desktop app or in the browser. Nor does it cover extra usage beyond your limit. You can spend it on your own automations through the API, and also on running Claude Code with claude -p using an API key from that linked organisation. Any credit you don’t use expires at the end of the billing period.
It’s good that Anthropic spelled out straight away what the credit doesn’t cover – “$100 on Max” is easy to read as more Claude Code. If you already run something on the API or plan to try it, this is money to get started. As a reason to move from Pro to Max, though, the credit doesn’t hold up: we only recommend Max once you regularly hit the Pro limit in the middle of your work.
The VS Code extension: a permission prompt before acting in Chrome, and message times on by default
Since version 2.1.296, released on Friday 9 October, the Claude Code extension for VS Code asks for permission in every session before it acts in the browser while Claude is driving Chrome – just as Claude Code does in the terminal. When an action concerns a site you haven’t approved yet, you can allow actions on that site for the rest of the session, and Claude won’t ask about the next ones. You’re signed in to your own accounts in that browser, so a few extra clicks seem a fair price to us.
Since version 2.1.290 from Monday 5 October, the extension also shows the time above every message by default, plus a date line where a new day begins. The feature arrived in version 2.1.284 but was off by default, which is how we described it in our News from 5 October. If the times get in your way, you can turn them off in the VS Code settings – the option is called Claude Code: Show Message Timestamps.
To finish: ArtCraft – seven open-source apps modelled on Adobe’s programs, built with Opus 5.5
Something to start the week: ArtCraft, seven free, open-source apps modelled on Adobe’s programs. Brandon Thomas published them on GitHub at the turn of September and October. PhotoCraft is modelled on Photoshop, VectorCraft on Illustrator, FilmCraft on Premiere, LightCraft on Lightroom, PrintCraft on Acrobat Pro, EffectCraft on After Effects and DesignCraft on InDesign. They’re all written in Rust, and each also has a version that runs in the browser. Thomas writes that he builds them with Opus 5.5 in plain Claude Code, with no add-ons, and that he doesn’t decompile Adobe’s software but recreates it from scratch.
Brandon Thomas is an experienced programmer. According to Ars Technica, he has around a decade of work on large Rust projects behind him; he himself mentions a fifteen-year career. Since February he hasn’t written code by hand – he works with AI. He also has a team, most of whom, he writes, are artists who use programs like these every day.
The apps are in an early alpha, and the author makes no secret of it. VICE tested them, reported plenty of bugs and concluded that they work – well, more or less. For us it’s above all a demonstration of what Opus 5.5 can do: a programmer with a small team of artists can now take on, with AI, seven programs Adobe has spent decades on. Thomas says his apps will reach full parity with Adobe’s within a month. The code is in open repositories, so anyone can check, commit by commit, how close PhotoCraft is getting.