Claude 3.5 Sonnet v2 and Claude 3.5 Haiku
October 2024 refresh shipped an upgraded Sonnet, a budget Haiku, and the first Claude model that could drive a desktop cursor.
October 22, 2024 was a double launch. A refreshed Claude 3.5 Sonnet with real gains on coding and tool use. A brand new Claude 3.5 Haiku for fast, cheap work. And one feature nobody saw coming: Computer Use, a mode where Claude could look at a screen, move a cursor, click buttons, and type like a person.
Key Specs
| Spec | Claude 3.5 Sonnet v2 | Claude 3.5 Haiku |
|---|---|---|
| API ID | claude-3-5-sonnet-20241022 | claude-3-5-haiku-20241022 |
| Context window | 200K tokens | 200K tokens |
| Input pricing | $3 / 1M tokens | $0.80 / 1M tokens |
| Output pricing | $15 / 1M tokens | $4 / 1M tokens |
| Max output tokens | 8,192 | 8,192 |
| Release date | October 22, 2024 | October 22, 2024 |
Computer Use (public beta)
This was the headline. Claude could look at a screenshot, read what was on screen, and decide where to click or what to type next. Not a browser macro. Not a scripted automation tool. The model was actually driving the desktop.
No major AI had shipped anything like it before. Computer Use opened up agent work that did not need a custom API for every tool. Point Claude at the interface and it could get the job done.
The feature ran in public beta through the API. Developers used it to build agents that:
- Move through web applications and fill out forms
- Move between desktop applications
- Extract information from visual interfaces
- Perform repetitive GUI-based tasks
- Test software by interacting with the actual UI
Claude needed screenshots in and coordinate outputs back. Anthropic wired safety limits in on top: the model could not take irreversible steps without confirmation, and rate limits kept it from running wild.
Upgraded Sonnet
The v2 was not a small patch. Coding, tool use, and following complex instructions all moved forward. Agent workflows that chained multiple tool calls, read the results, and adjusted on the fly got noticeably better. Developers using Claude as a coding assistant felt the bump right away.
Pricing stayed at $3/$15 per million tokens. The context window stayed at 200K. But the output quality was a real step up, and the "close but not quite" moments that plagued the original Sonnet got less frequent.
Claude 3.5 Haiku
The new Haiku pulled 3.5-generation smarts down to the budget tier. At $0.80/$4 per million tokens, it was a strong pick for classification, content moderation, chat products, and any high-volume job where cost matters more than peak reasoning.
Status
| Model | Status |
|---|---|
| Claude 3.5 Sonnet v2 | Superseded by Claude 3.7 Sonnet (February 2025) |
| Claude 3.5 Haiku | Still referenced in some deployments |
Both models have been passed by newer generations. But they mark the moment Claude stopped being a pure text generator. Computer Use was the proof point that opened the door to the more capable agent work that showed up in Claude 3.7 Sonnet and Claude 4.
Related Pages
- All Claude Models for the full model index
- Claude 3.5 Sonnet, the original June 2024 release
- Claude 3.7 Sonnet, the next generation with hybrid reasoning
Stop configuring. Start building.