claude-fable-5.1
Meant to excel at
-
Built for long-running agentic coding, multistep research, and document, spreadsheet and slide work over extended sessions.
Vendor's own site anthropic.com -
Big gains over Fable 5 on terminal and agentic science tasks; Terminal-Bench 4.0 reported at 55.8 percent.
Vendor's own site anthropic.com -
Cheaper cache reads make long agent runs roughly 25 to 45 percent cheaper than Fable 5, by Anthropic's estimate.
Vendor's own site anthropic.com -
Ranked first on Vals AI's Excel modeling benchmark at 76.7 percent accuracy, narrowly ahead of Opus 5.5.
Independent evaluator vals.ai -
Plans architecture first and finishes long coding tasks unprompted; preferred over GPT-6 Astra for finishing software projects in one comparison.
Independent evaluator hierarchy.ch -
Prose reads with fewer AI tells and beat GPT-5.6 Sol on judgment-heavy work like decks and persona answers.
Independent evaluator every.to -
Legal testers found it strongest on general legal knowledge and on separating what a document says from what it infers (vendor-run test).
Practitioner report litera.com
Watch out for
-
Overdelivers: exceeded explicit word, theme and quote limits, and some quoted passages were not in the source text.
Practitioner report every.to -
At extra-high effort it can spawn unneeded subagents and keep going after interruption; set a budget first.
Practitioner report theneuron.ai -
Can burn through subscription allowances fast on fleet-style multi-agent jobs, stalling work until reset.
Practitioner report ainativeproductmanager.com -
Penetration testing, exploit work and life-science R&D requests are rerouted to Opus models by safeguards.
Vendor's own site anthropic.com -
In legal testing it was the slowest and most expensive, wrote the longest answers, and gave fully complete answers less often than Opus 5.5 (vendor-run test).
Practitioner report litera.com
Benchmarks
Coding percentile 0.83
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| lmarena code arena webdev | 1743 | 0.83 | Independent evaluator arena.ai | 2026-10-19 |
Agentic coding percentile 0.55
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| cursorbench 4.0 | 51.8 | 0.33 | Vendor's own site anthropic.com | 2026-09-22 |
| swe-bench pro | 81.2 | 0.92 | Vendor claim, relayed computingforgeeks.com | 2026-09-01 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 85 | 0.33 | Independent evaluator anotherwrapper.com | 2026-10-19 |
| terminal-bench 4.0 | 55.8 | 0.63 | Vendor claim, relayed computingforgeeks.com | 2026-09-01 |
Knowledge percentile 0.92
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| humanity's last exam | 59.1 | 0.85 | Independent evaluator deeplearning.ai | 2026-09-11 |
| vals excel modeling benchmark (emb v1.0) | 76.67 | 1.00 | Independent evaluator vals.ai | 2026-10-18 |
Overall percentile 0.76
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 53 | 0.70 | Independent evaluator artificialanalysis.ai | 2026-10-19 |
| gdpval-aa v2 | 1853 | 1.00 | Vendor claim, relayed the-decoder.com | 2026-09-01 |
| gdpval-aa v2.1 | 1735 | 0.50 | Vendor's own site anthropic.com | 2026-09-22 |
| vals index | 68.83 | 0.83 | Independent evaluator deeplearning.ai | 2026-09-11 |
Vision percentile 0.67
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| chartography | 88.4 | 0.67 | Vendor's own site anthropic.com | 2026-09-22 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +/-
In an Endor Labs run of 179 coding tasks under Claude Code it passed 87.2% of functional tests versus 68.7% for Opus 5.5 (memorization removed) but only 37.4% of hidden security tests; it took about 9.5 min per task and the run cost $672 versus $116 for Opus 5.5.
Task: bug-fix | Compared with claude-opus-5.5 | 2026-09-24
Practitioner report endorlabs.com - +
Community default in launch-week threads: Opus for scoped build-and-fix, Fable for open-ended or knowledge-heavy problems (secondhand Reddit digest).
Compared with claude-opus-5.5 | 2026-09-24
Practitioner report buttondown.com - +
Every's reviewers still prefer Fable 5.1 over Opus 5.5 for the hardest, biggest problems and for deliverables that must be right first time; it also made a better branded deck (Opus 5.5 used wrong logo and colours).
Task: writing-content | Compared with claude-opus-5.5 | 2026-09-22
Practitioner report every.to - +
Day-zero users kept Fable for the hardest problems, deadline-critical deliverables and prose, while Opus 5.5 covered everyday coding at lower cost.
Compared with claude-opus-5.5 | 2026-09-22
Practitioner report every.to - -
Testers found Astra won most head-to-heads on coding and reasoning and needed less re-explaining, though raw benchmarks were roughly level.
Compared with gpt-6-astra | 2026-09-13
Practitioner report mindstudio.ai - +/-
Artificial Analysis found it leads on long-running tool use but trails GPT-6 Astra on overall knowledge, computer use and multistep terminal work, at higher cost and time per task; at max effort it used about 1.7x the output tokens of Fable 5.
Task: agentic-long-running | 2026-09-11
Practitioner report deeplearning.ai - +
Author would hand a whole software project to Fable 5.1 over Astra, though Astra clearly won Blender and 3D work; Fable over-delivered and spawned unrequested subagents.
Compared with gpt-6-astra | 2026-09-08
Practitioner report hierarchy.ch - +
Reviewer found Fable 5.1 needed about two PR corrections versus six for Astra, understood intent better, and produced polished one-shot front ends; Astra won on computer use and 3D.
Compared with gpt-6-astra | 2026-09-07
Practitioner report finance.biggo.com - +
Reviewer said Astra trailed Fable 5.1 by a wide margin on every coding task, especially UI polish, while Astra's computer-use work was strong.
Compared with gpt-6-astra | 2026-09-05
Practitioner report medium.com - +
Author gave it one prompt to write an image codec in Go and got something that beat WebP on compression by 1-9% and ran faster, though tested on under 50 images and not production-ready.
Task: feature | 2026-09-04
Practitioner report thomas-wiegold.com - +/-
Same reviewer could not confirm the final output is better than Fable 5, only that it feels more direct and less verbose; suggests defaulting to Opus 5 and escalating to Fable 5.1 only for hard or high-stakes work.
Task: feature | Compared with claude-opus-5 | 2026-09-04
Practitioner report thomas-wiegold.com - +
Jane Street's Craig Falls (customer quote in Anthropic launch coverage) said outputs stay readable across long multi-step tasks.
Task: agentic-long-running | 2026-09-02
Vendor claim, relayed sdtimes.com - +/-
Rewrite ran with few nudges and resumed cleanly after a crash, but cost estimates ran low and the top-tier five-hour allowance ran out in 43 minutes.
Compared with claude-fable-5 | 2026-09-02
Practitioner report ainativeproductmanager.com - +/-
Week of team testing: matched Fable 5 on long delegated coding, wrote cleaner prose, but broke explicit count and format limits that Opus 5 respected.
Compared with claude-opus-5 | 2026-09-01
Practitioner report every.to - +/-
On subscription Max plans it burns weekly limits faster than other Claude models; Reddit users reported five-hour windows gone in 15-20 minutes with subagent fan-outs, while token-billed API cost per turn was about 43% lower than Fable 5 in the author's transcripts. Full-file rewrites rose to 44% of file operations from 16%.
Task: agentic-long-running | Compared with claude-fable-5 undated
Practitioner report paddo.dev - +/-
Commenter noted Claude Code hits usage limits mainly when running Fable for everything, and writes its own monitor script otherwise.
undated
Practitioner report news.ycombinator.com