Following Opus and Sonnet, Anthropic is now also updating its smallest model to version 5.5. In several tests, Claude Haiku 5.5 comes very close to Sonnet's performance and costs only a fraction of its predecessor to run. At the same time, Sonnet 5.5 is becoming cheaper, and Max and Team subscribers will now receive monthly credits for the API.
Anthropic unveiled the Claude Haiku 5.5 on October 7th, the successor to the Haiku 4.5. It is the third model of the new generation, following the Claude Opus 5.5 on September 22nd and the Sonnet 5.5 six days later. This means the entire main series has been renewed within 15 days: Haiku as the smallest and most affordable model, Sonnet in the middle, and Opus as the largest and most expensive.
According to Anthropic's announcement, Haiku 5.5 is built for tasks that occur in large numbers and need to be inexpensive. Two further changes will accompany the launch: a price reduction for Sonnet 5.5 and a monthly credit for subscribers of the Max and Team plans.
Key Facts at a Glance
- The Claude Haiku 5.5 replaces the Haiku 4.5 and, according to Anthropic, is the company's fastest model to date at standard speed.
- For requests up to 100,000 tokens, the list prices drop by 90 percent; in practical operation, the model costs on average around 75 percent less.
- In knowledge work and computer control, Haiku 5.5 comes close to Sonnet 5.5, but in programming the gap remains large.
- Cached inputs now cost only half as much in Sonnet 5.5.
- Max and Team subscribers will receive a monthly credit for the Claude platform starting this week.
What Haiku 5.5 is intended for
Anthropic cites summaries, condensing long conversations, database queries, and sorting queries into categories as potential applications. In programming projects, Haiku 5.5 is intended to run as a helper agent alongside Opus 5.5 or Sonnet 5.5, handling subtasks while the larger model manages the main work. Due to its speed, Anthropic also envisions the model being used in real-time customer support and website navigation.
The speed specification has a limitation: it applies to the respective standard speed. In fast mode, the Opus models run faster than Haiku 5.5. For the first time, a Haiku model has an adjustable computation depth, allowing developers to balance lower costs with higher performance – this setting is already available on the larger models.
Haiku 5.5 is available now on the Claude platform as well as on Amazon Web Services, Google Cloud and Microsoft Azure.
The gap to Sonnet is shrinking significantly
Anthropic compares Haiku 5.5 to its predecessor, Sonnet 5.5, and OpenAI's GPT-6 Luna. A selection of the values from the announcement:
| Test | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| Knowledge work (GDPval-AA v2.1, points) | 1620 | 735 | 1437 | 1840 |
| Computer control (OSWorld 2.1) | 72,4 % | 15,7 % | 48,9 % | 83,9 % |
| Subject-matter expertise (Humanity's Last Exam, without tools) | 45,9 % | 10,2 % | – | 56,9 % |
| Agentic Programming (Terminal Bench 4.0) | 39,2 % | 0,0 % | 16,4 % | 70,6 % |
The biggest leap forward for the model is in computer control, specifically in the independent completion of multi-stage tasks on a real computer. Here, Haiku 5.5 achieves more than four times the score of its predecessor. In knowledge work, it is now only 220 points behind Sonnet 5.5, whereas Haiku 4.5 was over 1,100 points behind.
However, a significant gap remains in programming. Anthropic still recommends Sonnet 5.5 and Opus 5.5 for complex programming tasks and sees Haiku 5.5 as suitable for narrowly defined subtasks that would have been simply too expensive with earlier models.
Where the 75 percent come from
Haiku 5.5 has two price tiers, depending on the length of the request. The list prices in US dollars per million tokens are:
| Price per 1 million tokens | Haiku 5.5 (up to 100,000 tokens) | Haiku 5.5 (over 100,000 tokens) | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Input | 0,10 $ | 0,50 $ | 1,00 $ | 2,00 $ |
| Output | 0,50 $ | 2,50 $ | 5,00 $ | 10,00 $ |
| Cache read accesses | 0,01 $ | 0,05 $ | 0,10 $ | 0,10 $ |
| Cache write accesses | 0,125 $ | 0,625 $ | 1,25 $ | 2,50 $ |
Up to 100,000 tokens, the price is 90 percent lower than that of Haiku 4.5; above that, it's 50 percent lower. According to Anthropic, around 90 percent of all requests for Haiku 4.5 fell into the lower price tier. At this tier, a single request for Haiku 5.5 costs one-twentieth of what Sonnet 5.5 charges.
The fact that the savings end up being around 75 percent, and not 90 percent, is due to a revised tokenizer. Haiku 5.5 analyzes texts similarly to Sonnet 5.5 and Opus 5.5, and therefore requires slightly more tokens than its predecessor for the same task. The list prices alone thus exaggerate the effect.
Sonnet will become cheaper; Max and his team will receive API credits
With Sonnet 5.5, Anthropic is halving the price for cache read accesses from $0.20 to $0.10 per million tokens. Because such accesses account for a large portion of resource consumption, this should make Sonnet 5.5 approximately 20 percent cheaper for most agent tasks.
Users of Claude via a Max or Team subscription will receive monthly credits for the API starting this week. The amounts, according to Anthropic's help page, are as follows:
| Subscription | Monthly balance |
|---|---|
| Max 5x | 100 $ |
| Max 20x | 200 $ |
| Team, Standard Place | 20 $ per seat |
| Team, Premium Seat | 100 $ per seat |
With Team plans, the credit balance of all seats is pooled into a common fund, capped at $500 per month. Pro and Enterprise plans receive no credit. The credit applies to the API, Managed Agents, and the Agent SDK, but not to interactive sessions in Claude Code or Claude via Amazon, Google, or Microsoft. Unused credit expires at the end of each billing period.
Subscribers who purchased their subscription through the iPhone app via the App Store are also eligible. However, the credit can only be redeemed on claude.ai in a web browser by linking a Claude Console organization in the billing settings. This link can only be set once and can only be changed afterward via support. Activation takes several days, and new subscribers must also be on the plan for seven days.
Narrower limits on cybersecurity
According to Anthropic, Haiku 5.5 exhibited significantly fewer malfunctions in security tests than its predecessor and was more difficult to exploit. Its cybersecurity safeguards are stricter than those of Haiku 4.5, but somewhat less stringent than those of the most recent major models. Defensive tasks are permitted to a greater extent than in Sonnet 5.5, while penetration tests and similar attack techniques remain blocked.
A small model that takes work off Sonnet's hands
I believe the pricing structure is the real news, not the test results. A model that comes close to Sonnet's performance for knowledge work and costs a twentieth of that for short requests shifts the cost for anyone integrating Claude into their own tools. However, the test results are from Anthropic's own measurements, and the customer testimonials in the announcement come from partners who received the model in advance.
If you have a Max subscription in the iPhone app, you need to open claude.ai in your browser to redeem the credit – it cannot be redeemed within the app itself. Since it expires every month, linking the accounts is only worthwhile if you actually intend to build something using the interface.
Will you use the monthly API credit for your own tools or automations – or will Claude remain just a chat app for you? Tell us in the comments what you would build with it.
