@anthropic/claude-opus-5.5 is live on Brainbase

Claude Opus 5.5 is now available on Brainbase as @anthropic/claude-opus-5.5. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

Anthropic released Opus 5.5 on September 22 as the first model in the Claude 5.5 family. It keeps the 1M-token context window and 128K output limit of Opus 5, accepts text and images, runs adaptive thinking that cannot be turned off, defaults to medium effort, and has a June 2026 knowledge cutoff. Anthropic's own line on it is that it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.

#Why Opus 5.5 matters for agents

The headline is efficiency rather than a new ceiling. Opus 5.5 costs 20% less per token than Opus 5 and uses fewer tokens per task, which is where the 40% figure comes from, and it generates output more than 30% faster. Anthropic's early testers report the same shape: Box saw about a third as many tokens as Opus 5 with answers 40% shorter and no loss of accuracy, and GitHub saw more terminal tasks completed in less than half the steps inside Copilot.

The second thing is where the gains land. The largest improvements over Opus 5 are on agentic coding, computer use and long-horizon professional work, the three things an internal agent spends its time on. Anthropic is candid that at this level benchmark margins are a weaker guide than they used to be, and that in its own use the gap to Fable 5.1 is narrower than the scores suggest. The scores are still the best public evidence there is, so here they are.

#Opus 5.5 benchmarks

On Anthropic's published set, Opus 5.5 leads every compared model on Terminal-Bench 4.0, CursorBench 4.0, FrontierCode, GDPval, Humanity's Last Exam and OSWorld, and improves on Opus 5 on every row. GPT-6 Astra keeps the lead on AutomationBench and Terminal-Bench-Science. Select a metric below to compare the reported results.

Anthropic ran Opus 5.5 at max effort with its production safeguards on, except Terminal-Bench 4.0 at xhigh; the GPT figures are as reported by OpenAI, and AutomationBench was run by Zapier without fallback models. Independent testing from Artificial Analysis puts Opus 5.5 at 58 on the Intelligence Index, the highest score they have measured and several points clear of the field, with leading results on six of the ten evaluations, including Humanity's Last Exam, SciCode, GDPval-AA and AA-Briefcase. Their Terminal-Bench 4.0 run scores it 59.6%, level with GPT-6 Astra and eleven points over Opus 5.

#Opus 5.5 pricing and context

Every rate is lower than Opus 5, and the one that matters most for agents fell furthest: cache reads, which are most of the cost of a long agentic session, are 60% cheaper. There is no long-context surcharge across the 1M window. A fast mode at up to 2.5× speed is available in Claude Code and the Claude Platform at double the standard rate.

These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage. For an agent that replays a long prefix on every step, the cache read price sets the bill, which is why Anthropic's 40% figure is plausible on real workloads even though the sticker cut is 20%.

#What changed in the API

Four breaking changes carry over from Fable 5.1 for code written against Opus 5: thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation, and the older computer-use tool is not accepted. A fifth change alters the response shape without failing anything: text between tool calls now arrives in thinking blocks that are empty at the default display setting, so an application that streams that text as progress goes quiet between calls until it sets a display value. Brainbase handles the harness side of this; if you read the raw event stream, check it.

One more thing to plan for. Opus 5.5 ships with safeguards for cybersecurity, biology and frontier-AI development, and when they intervene the request is served by an older model: Opus 4.8 for most cyber work, Opus 5 for the others. In a multi-step agent that means individual steps can be answered by a model with different capabilities. Observability shows you which; permissions decide what those steps may touch.

#Run Opus 5.5 on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Opus 5.5 inside the Claude Code harness on a managed sandbox:

curl https://api.brainbaselabs.com/v2/threads \
  -H "Authorization: Bearer $BRAINBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": {
      "harness": "claude_code",
      "model": "@anthropic/claude-opus-5.5"
    },
    "input": "Audit this repository, fix the highest-impact reliability issue, and run the tests."
  }'

#Where Opus 5.5 fits

Opus 5.5 is the new default for terminal-heavy and repository-scale work, computer use, and the multi-hour professional tasks that end in a document. It beats GPT-6 Astra on FrontierCode at about a fifth of the cost per task and matches it on Terminal-Bench at about 40%, on Anthropic's numbers, and it costs 60% less per token than Fable 5.1 while scoring higher on most of this set.

The practical move is to start at the default medium effort, which is where most of the efficiency lives, and only raise it where your evals say it pays. Brainbase keeps the harness, sandbox, event stream and observability layer constant so the model and its effort setting are the only variables.