The Claude 5.5 family is complete. Opus 5.5 landed September 22, Sonnet 5.5 on September 28, and today Anthropic shipped Claude Haiku 5.5.

It's the cheapest Claude model yet, and the jump is bigger than the version number suggests. There was never a Haiku 5. This goes straight from Haiku 4.5 to 5.5.

And if you read our Fable 5.5 issue, you'll remember the rumor said Fable would follow Haiku by about a week. That week starts now.

The price is the headline.

For requests under 100,000 tokens, that's 90 percent cheaper than Haiku 4.5. Anthropic says about 90 percent of Haiku 4.5 requests fell into that lower tier.

The specs:

1M token context window, with up to 128K output tokens, or 300K in batch, which is in beta.

Effort controls, a first for Haiku. It defaults to medium, with adaptive thinking on.

Computer use and browser use, with SDK support in beta.

Knowledge cutoff of June 2026.

It's live now as claude-haiku-5-5 on the Claude API, AWS, Google Cloud and Microsoft Azure.

Look at the price again. $0.10 input, $0.50 output. That's exactly what OpenAI charges for GPT-6 Luna, launched two weeks ago.

So the Claude 5.5 family now matches OpenAI's lineup tier for tier. Sonnet 5.5 costs the same as GPT-6.1 Sol, and Haiku 5.5 costs the same as Luna.

At the same price, here's how they compare on Anthropic's numbers:

Haiku 5.5 beats Luna across the board. Sonnet 5.5 is still clearly ahead of both.

Same warning as every launch: these are Anthropic's own numbers. Wait for independent runs before you move production traffic.

Skipping Haiku 5 shows in the scores:

Computer use: 15.7 percent → 72.4 percent

Terminal-Bench: 0 percent → 39.2 percent

Knowledge work (GDPval): 735 → 1,620

That's not an upgrade. Haiku 4.5 was a model for summarizing and sorting. Haiku 5.5 can operate a computer most of the time. Anthropic now pitches it for live customer support, browser use, database queries and other agent work where speed matters, not just classification and routing.

Four things worth knowing before you switch:

The new tokenizer counts about 30 percent more tokens for the same text. That's why Anthropic says average real-world savings are around 75 percent, not 90.

Big requests cost more. Go over 100K tokens and the price jumps to $0.50 input and $2.50 output.

The Terminal-Bench score is at maximum effort. At the default medium setting, it's closer to 20 percent.

You can't tune sampling. Setting temperature, top_p or top_k away from the defaults returns an error.

Sonnet 5.5 cache reads were cut in half, from $0.20 to $0.10 per million tokens. That matches GPT-6.1 Sol's cached input price, which was the one advantage Sol still had on cost.

Max and Team subscribers now get monthly API credits: $100 for Max 5x, $200 for Max 20x, and up to $500 shared across a Team plan.

In our Fable 5.5 leak issue, we said Haiku 5.5 was the tell. The rumor said Haiku would ship first and Fable 5.5 about a week later.

Haiku is out. If Fable 5.5 shows up in the next week or so, the leak was real. If it doesn't, the rumor joins Fable 5.2 and Opus 5.2 on the list of good guesses with the wrong name.

Anthropic's launch says nothing about Fable either way.

The small model just became an agent model.

For years the cheapest tier was for the easy, boring jobs, like tagging, summarizing and routing, while anything that took action went to a bigger, pricier model. Haiku 5.5 controls a computer 72 percent of the time on OSWorld at ten cents per million input tokens. A lot of agent work that ran on Sonnet last month can now run on Haiku.

It also locks in where pricing is headed. Both Anthropic and OpenAI now sell three tiers, and the bottom two are priced identically. Competition at each tier is now about quality, not price.

Practical move: if you run high-volume work on Haiku 4.5 or Luna, test Haiku 5.5 on a real sample this week. Recount your tokens with the new tokenizer, check how many of your requests go over 100K, and try medium effort before high.

The stories moving fast beyond today's headlines:

🧪 Claude Console Workbench is the fastest way to run your own prompts on Haiku 5.5, and compare effort levels side by side, before you switch any code.

📊 ccusage shows what you're actually spending per model in Claude Code. Run it before and after switching cheap tasks to Haiku 5.5.

🔀 LiteLLM lets you point the same code at Haiku 5.5 and GPT-6 Luna, so you can test them head to head on your own work at the same price.

Anthropic finished the Claude 5.5 family with Haiku 5.5. It costs $0.10 input and $0.50 output, matching GPT-6 Luna, and beats Luna on every benchmark Anthropic shared, with a 1M context and real computer use. The takeaway is that the small tier is now good enough for agent work, so the default question has flipped. It's no longer "is Haiku good enough?" but "does this job really need Sonnet?" Start by moving one high-volume workflow over, count tokens with the new tokenizer before you trust the savings, and keep an eye on the next week, because if Fable 5.5 is real, this is when it should show up.

Stay building. 🤖