Claude Cowork vs Claude API vs OpenClaw

Side by side

Pick…

• Cowork for knowledge work (documents, research, files, recurring summaries) with nothing to install and one subscription.
• The API when the agent is part of your product, or you need programmatic control and per-token billing.
• OpenClaw (or Hermes) when you want an always-on assistant in your chat apps, a choice of model, or everything on hardware you control.

Building on the API? Moving from Sonnet 5 to Sonnet 5.5

Claude Sonnet 5.5 (claude-sonnet-5-5, launched 2026-09-28) costs the same $2/$10 per million tokens as Sonnet 5, with a 1M-token context and 128K max output. It is also the default Sonnet in Claude Code from 2.1.284. Code written for Sonnet 5 can still break in five places; Anthropic's migration guide has the before-and-after requests:

• Thinking off: send thinking: {"type": "between_tools"} instead of "disabled", which now returns a 400 error. between_tools works at low, medium and high effort; at xhigh or max use adaptive thinking.
• Forced tool use: tool_choice of type any or tool returns a 400 error (use auto).
• Thinking blocks that Sonnet 5.5 produces work only in the account that produced them (or an account linked to it).
• Computer use: on the Claude API and Google Cloud, use the computer_toolset_20260801 toolset; computer_20251124 returns a 400 error.
• Advisor tool: a Sonnet 5.5 executor rejects Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5 and Sonnet 4.6 as advisors.

Still on Sonnet 4.5? Anthropic deprecated it on 2026-09-30 and plans to retire claude-sonnet-4-5-20250929 on the Claude API on 2026-11-30, recommending Sonnet 5.5.

Building on the API? Haiku 5.5

Claude Haiku 5.5 (claude-haiku-5-5, launched 2026-10-07) is Anthropic's small, fast model for high-volume work. It has a 1M-token context window and 128K max output, uses adaptive thinking with the effort parameter, and is priced by prompt length: $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 above that (cache reads $0.01 and $0.05). Prompt length counts every input token, including cache reads and writes, and each request is priced on its own. It succeeds Haiku 4.5 ($1/$5). In Claude Code the haiku alias resolves to it on the Anthropic API from v2.1.293. Code written for Haiku 4.5 changes in these places, per Anthropic's migration guide:

• Model ID: claude-haiku-5-5 is a fixed ID with no date suffix and no alias (Amazon Bedrock uses anthropic.claude-haiku-5-5).
• Token counts: the same text produces roughly 30% more tokens than on Haiku 4.5 (the newer tokenizer), so recheck max_tokens and cost estimates.
• Thinking: manual extended thinking with budget_tokens returns a 400 error; use {"type": "adaptive"} and effort. Adaptive thinking is on by default, so a response can start with thinking blocks; select content blocks by type.
• Request parameters: remove temperature, top_p and top_k, and end messages with a user turn instead of an assistant prefill.
• Refusals: handle stop_reason: "refusal"; Haiku 5.5 runs safety classifiers and has no server-side fallback.
• Computer use: replace computer_20250124 with the computer_toolset_20260801 toolset on the Claude API and Google Cloud.

Structured outputs aren't available for Haiku 5.5 on Amazon Bedrock. Max and Team plans now include monthly Claude API credits ($100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team) that cover the API, Managed Agents and the Agent SDK but not Claude Code; see Anthropic's page.

Cowork's risks, in Anthropic's words

• Isolation protects your computer, not your accounts. A cloud session can't reach your network, but it can still read your email, documents and websites through whatever access you grant, and act on them.
• Prompt injection needs two things at once: Claude reading content outside your trust boundary, and being able to take harmful actions. Remove either and the attack fails. Be sparing with write tools when Claude reads untrusted content.
• Computer use has no sandbox between Claude and your apps. Don't give it banking, healthcare or government apps.
• Scheduled tasks run unwatched. Start with summaries, and don't schedule anything that sends messages or makes purchases.
• Anthropic's safeguards (model training, content classifiers, Auto-mode action screening, deletion prompts, per-app computer-use permissions) reduce the risk but don't remove it.

Read the full guidance: Use Claude Cowork safely and Let Claude use your computer.