 |
VOOZH | about |
|
VOOZH | about |
We’re so glad you’re here. You can expect all the best TNS content to arrive Monday through Friday to keep you on top of the news and at the top of your game.
Check your inbox for a confirmation email where you can adjust your preferences and even join additional groups.
Follow TNS on your favorite social media networks.
Become a TNS follower on LinkedIn.
Check out the latest featured and trending stories while you wait for your first TNS newsletter.
Anthropic today launched the latest version of its Claude Haiku model, Haiku 4.5, the smallest, fastest and most affordable model in its family.
We’ve seen a flurry of new model releases from Anthropic in recent months, starting with the launch of Sonnet 4 and Opus 4 in May, followed by Opus 4.1 and, just a few weeks ago, Sonnet 4.5. What was missing from this latest generation of models was Haiku, which always offered a worthwhile trade-off between price, speed and model intelligence. The last generation of Haiku, version 3.5, launched at the end of last year, and it almost seemed like Anthropic had forgotten about that side of its model family.
What makes Haiku 4.5 stand out is that, in many respects, it’s up to par with Claude Sonnet 4, which was state-of-the-art when it launched six months ago. Anthropic argues that this makes Haiku 4.5 an especially interesting model for those who may want to use it in a multi-agent system where Sonnet farms out tasks to Haiku 4.5-based sub-agents.
The company is confident in the performance of Haiku 4.5 that it will make it the default for its Claude.ai service for a random selection of users. All users will be able to select Haiku 4.5 from the model selector, though.
That performance comes at a higher cost than Haiku 3.5, though. Pricing for Haiku 4.5 will be $1 per million input tokens and $5 per million output tokens. For Haiku 3.5, that was $0.80/$4 per million input/output tokens.
Claude Haiku 4.5 is now available on Claude.ai, as well as in the Anthropic API, on Amazon Bedrock and Google Cloud’s Vertex AI.
In the benchmarks released by Anthropic, Haiku 4.5 compares favorably with Sonnet 4 and, in many cases, even with Google’s Gemini 2.5 and OpenAI’s GPT-5. That’s especially true for areas where Anthropic’s models have traditionally been strong, including coding and tool use.
In some areas, including computer use and on the SWE benchmark, which tests how well the model can solve a set of GitHub issues, Haiku 4.5 actually surpasses Sonnet 4.
It helps that Haiku 4.5 is a hybrid reasoning model, with an optional ‘extended thinking mode,’ a first for the Haiku series.
The context window for Haiku 4.5 will remain at 200,000 tokens, but users on the Claude Developer Platform will be able to access a 1-million-token context window.
In real-world chat-based usage, however, Anthropic argues that most users won’t see any major difference between Sonnet and Haiku.
“Historically, models have sacrificed speed and cost for quality,” said Windsurf CEO Jeff Want. “Haiku 4.5 is blurring the lines on this trade-off: it’s a fast frontier model that keeps costs efficient and signals where this class of models is headed.”
When I asked Alex Albert, Anthropic’s head of developer relations, what had taken so long for a new model to be released, he noted that the company’s focus over the last year had been on improving its frontier models.
“One thing Anthropic does really well is to stay laser focused on the goal at hand and not try to do too many things all at once and really deliver on what our stated goal is,” Albert said. “But now, I think, things have shifted somewhat. [With] Sonnet 4.5, we delivered another leap in the frontier, and now it’s about, hey, how do we actually unlock some more use cases as well for different types of applications that pair really nicely with the frontier.”
The use case that Anthropic is homing in on here is the use of Haiku for building fast, task-specific sub-agents in multi-agent systems. Indeed, that’s what Claude Code, Anthropic’s popular coding agent, will start doing. For these systems, Sonnet 4.5 can provide guidance, with Haiku 4.5 handling the execution (though Haiku 4.5 itself is also quite good at tool calling, Albert stressed). Since the tasks are clearly scoped and directed (and speed matters since many of these sub-agents can run in parallel), there isn’t much of a performance tradeoff here.
“I’m just really excited about this concept of thinking about intelligence levels. So if you think about a Sonnet 4 intelligence level, Haiku, 4.5 is right at that level or even slightly above in some ways. And the price has now dropped 3x in five months,” Alber told me, referring to the price difference between Sonnet 4 and Haiku 4.5. “I think that’s a really great story here around how models in certain intelligence levels are getting cheaper over time. That’s been the promise of AI over the past few years. To see that in action in a really tangible way is pretty cool to me, and I think developers are really going to like that.”
Claude Opus 4.1, which used to be considered as Anthropic’s flagship model, is currently listed as a “legacy brainstorming model” in Claude.ai’s model picker. Albert said that some users prefer it, but he now recommends that everybody use Sonnet 4.5.
That may change, of course, when Anthropic launches the next version of the Opus model.
