@anthropic/claude-fable-5.1 is live on Brainbase

Claude Fable 5.1 is now available on Brainbase as @anthropic/claude-fable-5.1. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

Anthropic released Fable 5.1 and Mythos 5.1 on September 1 — the first update to the line since Fable 5 shipped on June 9. The two are the same underlying model with different safeguard configurations: Fable 5.1 is generally available, while Mythos 5.1 is restricted to US organizations enrolled in Anthropic's Cyber Verification and Life Sciences Verification programs. Brainbase hosts Fable 5.1.

#Why Fable 5.1 matters for agents

Fable 5.1 is not a general intelligence jump so much as a targeted one. The gains concentrate in exactly the work long-running agents do: operating a terminal across many turns, driving a computer, following instructions without drifting, and finishing multi-step tasks without losing the plot.

The clearest example is Terminal-Bench-Science, where the model goes from 24.7% to 52.6% — more than double. AutomationBench, which measures business workflow execution, nearly doubles from 17.1% to 31.4%. Those are the benchmarks that most closely resemble a production agent trajectory, and they moved far more than the knowledge benchmarks did.

The other half of the release is behavioral. Anthropic reports 60% fewer false positives on cybersecurity safeguards and 85% fewer on harmless biology and medicine questions, plus a gain on instruction following. For anyone who has watched an agent refuse a legitimate security audit or a routine lab-notebook task halfway through a run, that is a reliability fix, not a footnote.

#Fable 5.1 benchmarks

On Anthropic's published evaluation set, Fable 5.1 leads Fable 5, Opus 5 and GPT-5.6 Sol on every reported benchmark. The margins vary widely: 23.6 points over Opus 5 on Terminal-Bench-Science, but only 29 Elo on GDPval-AA v2 and 1.4 points on Humanity's Last Exam with tools. Mythos 5.1, under lighter safeguards, reaches 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8%. Select a metric below to compare the reported results.

Independent testing from Artificial Analysis puts Fable 5.1 at the top of its Intelligence Index with a score of 66, ahead of Opus 5 at 63, GPT-5.6 Sol at 61 and Fable 5 at 60. That is the headline result — but the same evaluation carries the release's most important caveat, which is about cost rather than quality.

#Fable 5.1 pricing and context

Per-token pricing is unchanged from Fable 5. What moved is the cache: reads drop from $1.00 to $0.25 per million tokens, a 75% cut. Anthropic estimates that makes typical workloads roughly 25% cheaper and highly agentic ones up to 45% cheaper.

For comparison, Opus 5 runs $5 input and $25 output per million tokens — half the sticker price. Whether Fable 5.1 is actually cheaper than Opus 5 for your workload depends almost entirely on cache hit rate and the effort level you run it at.

#The cost story is about effort

Fable 5.1 exposes five effort levels, and they change the economics more than the price sheet does. At low and medium effort the model matches or beats Fable 5 while spending fewer tokens. At the top of the dial it spends considerably more.

Artificial Analysis measured this directly: at max effort Fable 5.1 costs about 20% more per Intelligence Index task than Fable 5 — $3.76 against Opus 5's $2.34 — because it emits roughly 1.7× the output tokens. Dropping to xhigh scores 65 at $2.72 per task. So the 25–45% savings Anthropic quotes are real, but they are a cache-hit story on agentic workloads, not a blanket discount. Turn the dial to max on every request and the bill goes up.

The practical read: default to medium or high, reserve xhigh and max for the small set of tasks that actually need them, and make sure your prompt prefixes are stable enough to hit the cache. That last point is where the 75% cache cut pays for itself.

#Three breaking changes to plan for

Fable 5.1 is not a drop-in replacement for every existing integration. Three changes will surface as errors or silent regressions rather than as quality differences.

• Forced tool use is rejected. Requests carrying tool_choice: {"type": "any"} or a named tool now return a 400. The model thinks before it acts, and forcing a call ahead of that step no longer does what callers expected.
• Thinking blocks are not portable. Only Mythos 5.1 can read Fable 5.1's thinking blocks. A fallback route to an older model drops them silently, and the model re-plans — paying for latency and tokens you already spent.
• Conversation history is validated. For accounts created after August 31, 2026, editing prior turns, the system prompt, or the tool array invalidates the thinking blocks that follow and raises an error. Trajectory-rewriting patterns need revisiting.

Anthropic's prompting guide also flags softer shifts worth testing for: less parallel tool calling, fewer progress updates mid-task, denser prose with fewer bullet points, and variable memory recall at low effort.

#Run Fable 5.1 on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Fable 5.1 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-fable-5.1"
    },
    "input": "Reproduce the flaky integration test, find the root cause, and open a fix."
  }'

#Where Fable 5.1 fits

Early adopters have been direct about the tradeoff. Cognition reported that Fable 5.1 matched or edged out Fable 5 in testing at a lower cost per task, with the new cache pricing finally making it economical. Jane Street said it solves more coding problems than Fable 5 or Opus 5 while holding readability across long, multi-step tasks. Millennium credited it with finding a rare crash nobody had explained in four to five years. Several teams, Cognition and Every among them, have moved Opus traffic over.

Fable 5.1 is the strongest default we have seen for agentic science, terminal-heavy engineering, computer use, and long knowledge-work sessions where instruction drift is the failure mode. It is not the cheapest way to answer a short question, and at max effort it is not even the cheapest way to answer a hard one.

Run it against your own traces before you route everything to it. Brainbase keeps the harness, sandbox, event stream, and observability layer constant so the model is the only variable — including the effort level you pick.