Is Claude Fable 5.1 Still Worth It for Agentic Coding?

For most agentic coding, no. Claude Opus 5.5, released on September 22, 2026, scores higher than Claude Fable 5.1 on all four coding and computer-use rows in Anthropic’s own comparison, and costs 40% as much per input and output token. On Artificial Analysis’s independent runs, Opus 5.5 at high effort beat Fable 5.1’s best Terminal-Bench 4.0 score at under a third of the cost per task.

Anthropic’s own guidance now says the same. Its model selection guide opens with “Most workloads start with Claude Opus 5.5” and tells teams to move to Fable 5.1 only “if your evals at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work.” Fable 5.1 still has a job. It is now the model you escalate to, not the one you start on.

These are the four rows from Anthropic’s Opus 5.5 announcement that bear on agentic coding. Both models were run by Anthropic, mostly at max effort; Opus 5.5’s Terminal-Bench 4.0 score is its xhigh result, its best on that test.

The Terminal-Bench lead is the one that clears the noise. Anthropic gives a standard error of ±2.6 points for Opus 5.5 on that test, and the gap is 10.6. CursorBench and FrontierCode are smaller but consistent. Other vendors’ runs of Fable 5.1 land close to Anthropic’s: OpenAI scored it at 55.8% on Terminal-Bench 4.0 and 50.9% on FrontierCode Main, and xAI at 51.8% on CursorBench 4.0.

Two cautions. Anthropic’s own Fable 5.1 figures moved between its two announcements: Humanity’s Last Exam with tools went from 65.0% on September 1 to 65.6% on September 22, and OSWorld 2.0 partial from 77.9% to 80.7%. And Anthropic says it thinks the table overstates the difference:

That cuts both ways. It means Opus 5.5 is probably not dramatically better than Fable 5.1 in real use. It also means the two are close enough that price should decide, and on price the answer is clear. Anthropic’s headline describes Opus 5.5 as performing “at the level of Claude Fable 5.1 on most work,” which is a claim of parity, not superiority.

Artificial Analysis ran both models at all five effort levels on its own Terminal-Bench 4.0 harness (model pages for Opus 5.5 and Fable 5.1). The table shows each score beside the model’s cost per task across the firm’s full Intelligence Index (v4.3.2, read on September 22), which includes Terminal-Bench and nine other evaluations.

Three things stand out. Opus 5.5 at high (56.6%) already beats Fable 5.1’s best result at any effort (55.1% at xhigh). Fable 5.1 scores lower at max than at xhigh on this test, so its most expensive setting buys nothing here. And Fable 5.1 has the higher floor: at low effort it scores 40.4% against Opus 5.5’s 31.3%. But Fable at low still costs more per task ($2.37) than Opus 5.5 at medium ($1.34), which scores 52.5%.

A second independent evaluator, Vals AI, published its Opus 5.5 results the same day, and its longest tasks point the same way. On its RSI Index, where agents get 24 hours to improve things like a language-model training run or a compressor, Opus 5.5 ranked first and Fable 5.1 second. On ProgramBench, which asks a model to rebuild a command-line program from its binary and counts a task only when every hidden test passes, Opus 5.5 fully solved 37 of 200 tasks to Fable 5.1’s 14.

On LM Training, Opus 5.5’s small model reached 0.831 bits per byte, below the 0.85 published human reference; Vals notes that its setup is not a matched reproduction of the reference paper. Two caveats keep this in proportion. Vals ran Opus 5.5 with Opus 5 and Opus 4.8 as fallbacks for refused requests; counting those tasks as failures drops its Terminal-Bench 4.0 score from 61.62% to 53.54%. And on Vals’ broad index across legal, finance, medical and coding work, Fable 5.1 still ranks first and Opus 5.5 fourth.

Matching the two models at similar scores gives the clearest answer to how much more cost-effective Opus 5.5 is. All figures are from Artificial Analysis; the ratios are our arithmetic.

• Same Terminal-Bench score: Opus 5.5 at medium (52.5%) against Fable 5.1 at high (52.0%) is $1.34 against $3.91 per task, about 2.9× cheaper. On the Terminal-Bench portion of that cost alone, the ratio is also about 2.9×.
• Fable’s best coding result: Opus 5.5 at high beats Fable 5.1 at xhigh (56.6% against 55.1%) for $1.82 against $5.98, about 3.3× cheaper.
• Same overall index score: Opus 5.5 at high (53.6) against Fable 5.1 at max (53.4) is $1.82 against $7.63, about 4.2× cheaper.

Anthropic’s own long coding test points the same way. Asked to translate the HAProxy load balancer from C into Rust, both models passed nearly all of its regression tests. Opus 5.5 finished in 9.5 hours against Fable 5.1’s 12, at 51% lower cost.

Per token, the gap is smaller on one line only. Fable 5.1’s cache reads are cheap relative to its own input price, at $0.25 against Opus 5.5’s $0.20. That was the argument for Fable 5.1 over Opus 5 in our September 1 Fable 5.1 post. Against Opus 5.5, Fable 5.1 costs more on every line.

Anthropic still calls Fable 5.1 its “most capable widely released model,” and its selection matrix lists it for “the highest available capability”: agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck. Opus 5.5 is the matrix’s pick for complex agentic coding, including multihour autonomous coding agents and large refactors.

In practice, that leaves four reasons to reach for Fable 5.1:

• Your evals at Opus 5.5 xhigh or max still fail. This is Anthropic’s own trigger, and the only one backed by your data rather than a benchmark.
• The task is research or document work as much as code. Anthropic positions Fable 5.1 for work that ends in a finished report or deck. Even here, Anthropic’s own research-report test favoured Opus 5.5, so check before assuming.
• You need a stronger result at low effort. Fable 5.1’s low setting outscored Opus 5.5’s on Terminal-Bench, though Opus 5.5 at medium was both better and cheaper.
• The work spans domains beyond code. Vals AI’s broad index ranks Fable 5.1 first of 60 models at 68.83% and Opus 5.5 fourth at 66.16%, and Vals found Opus 5.5 weaker than Opus 5 on medical coding, legal research and tax tasks.

The two models behave differently in production, and one of the differences makes the escalation pattern work.

The last row is the useful one. On the Claude API, Fable 5.1 can read the thinking blocks Opus 5.5 produces, so a conversation can start on Opus 5.5 and move up to Fable 5.1 without losing the reasoning so far. The reverse doesn’t work: moving from Fable 5.1 down to Opus 5.5 drops Fable’s reasoning. So start on Opus 5.5 and switch up when a task stalls; don’t start on Fable 5.1 and try to save money later in the same conversation.

One route that isn’t available yet: Anthropic’s advisor tool, which lets a cheaper model consult a stronger one mid-task, does not list Opus 5.5 in its compatibility table as of September 22. Until it does, escalation means switching the conversation’s model, not pairing the two.

For regulated work, the retention row may decide the question before any benchmark does. According to its model page, Fable 5.1 carries 30-day data retention and isn’t available under zero data retention unless Anthropic expressly authorises it. Opus 5.5 is.