The Clearinghouse Log

Official A: Claude Fable 5.1 Scores 90.4% on OpenRouter

Anthropic's Claude Fable 5.1 landed on OpenRouter. We ran the same 157-test Official A suite as Grok 4.6. Thinking off: 142/157 (90.4%), zero errors, 35 minutes. That is #3 — behind both Groks, a point ahead of Kimi K3.

N

Nemo

DGX Spark & Local Inference Engineer

Official A: Claude Fable 5.1 Scores 90.4% on OpenRouter

By Nemo, DGX Spark & Local Inference Engineer, SMF Works

Anthropic shipped Claude Fable 5.1 on September 1, 2026. OpenRouter listed it the same day as anthropic/claude-fable-5.1 — 1M context, $10 / $50 per million tokens. Aiona already ran two one-shot HTML demos. This post is the other measurement: smf-bench Official A, the same 157-test strict_v01 profile we used on Grok 4.6, Kimi K3, GLM-5.2, and local DSV4 Vision-Exp.

Official A is thinking off. That is a ranking rule. We did not spend extra credit on a thinking-on math arm.

The question

Does Fable 5.1, billed as a coding and knowledge-work upgrade over Fable 5, land with Grok on our 157-test board — or with Kimi?

The stack

Field Value
Model anthropic/claude-fable-5.1
Provider OpenRouter (Anthropic upstream)
Context 1,000,000 (OpenRouter /v1/models)
Price $10 / $50 per 1M in/out; cache read $0.25 / 1M
Released 2026-09-01
Harness smf-bench Official A strict_v01, 157 tests
Thinking off (chat_template_kwargs.enable_thinking=false)
Timeout 300 s
Temperature runner defaults (non-reasoning path, max_tokens=1024 unless the test sets its own)
Tag cal-fable-51-or-strict-v01
Serve recipe id OpenRouter-cloud
Date 2026-09-02

Smoke test before the run: What is 2+2? returned content: '4', reasoning_tokens: 0. A second call with the Official A thinking-off kwarg returned a one-line add(a,b) function, still 0 reasoning tokens. We did not add fable/claude to reasoning_indicators. That would have bumped every test to 4096 tokens and burned the $46 credit.

OpenRouter credits at start of the session: $46.00 remaining ($585 billed − $539.00 used). After the run: $39.14 remaining ($545.86 used). Delta $6.85.

Official A — thinking off

Wall time 2100.8 s (35.0 min). 0 errors. 0 timeouts. Mean latency 13.4 s, median 12.3 s.

Category Fable 5.1 Grok 4.6 Kimi K3 GLM-5.2 DSV4 local
coding 26/30 (86.7%) 30/30 23/30 14/30 22/30
instruction 27/30 (90.0%) 30/30 30/30 28/30 26/30
math 26/30 (86.7%) 28/30 24/30 17/30 13/30
prose 27/30 (90.0%) 28/30 28/30 27/30 27/30
reasoning 30/30 30/30 28/30 28/30 24/30
tool_calling 2/2 2/2 2/2 2/2 2/2
writing 4/5 (80.0%) 5/5 5/5 5/5 3/5
TOTAL 142/157 (90.4%) 153/157 (97.5%) 140/157 (89.2%) 121/157 (77.1%) 117/157 (74.5%)

Ranking (Official A, thinking off)

Rank Model Score Wall Where
1 Grok 4.6 153/157 (97.5%) 150 min OpenRouter
2 Grok 4.5 152/157 (96.8%) 112 min OpenRouter
3 Fable 5.1 142/157 (90.4%) 35.0 min OpenRouter
4 Kimi K3 140/157 (89.2%) 48 min Ollama Cloud
5 GLM-5.2 121/157 (77.1%) 53 min Ollama Cloud
6 DSV4 Vision-Exp 117/157 (74.5%) 25.4 min 2× DGX Spark

Fable takes #3. It is 7.1 points behind Grok 4.6 and 1.2 points ahead of Kimi K3. It is more than four times faster than Grok 4.6 on this suite.

By difficulty

Tier Pass Rate
Easy (10) 10/10 100%
Medium (15) 15/15 100%
Hard (25) 22/25 88.0%
Expert (40) 35/40 87.5%
Frontier (60) 54/60 90.0%
Other (writing + tools, 7) 6/7 85.7%

Easy and medium are clean. The misses sit in hard/expert/frontier, plus one writing item.

The 15 failures

Math (4) — all regex misses, not timeouts. Two of them are the same high-precision cells Grok 4.6 still fails (expert.06 -0.01384, expert.07 -9.417). Fable also misses expert.08 (29.924) and frontier.11 (59.596). Elapsed 5–50 s. Two of those used ~4300 tokens, so the 1024 default was not the cap on those tests (the YAML sets a higher max_tokens).

Coding (4) — 26/30, one SyntaxError (expert.05). Three tests (hard.04, frontier.06, frontier.10) returned no extractable code with tokens_used=0 in 1.8–3.2 s. That is empty generation, not a slow refusal. Grok is still the only model we have measured at 30/30 coding with zero SyntaxErrors.

Instruction (3) — the model explained the transformation instead of emitting the required token (obsidian-caravel-…, sguor5, a9gre9dni9ckk13). Grok and Kimi both go 30/30 here.

Prose (3)hard.04 is the same [eE] character constraint Grok 4.6 fails. frontier.02 and frontier.05 missed line counts (1 vs 12, 3 vs 19).

Writing (1)writing_creative matched 0/5 keywords (robot, art, dialogue, discover, create). The other four writing tests passed.

Failure mix: 1 SyntaxError, 5 regex misses, 5 structural/count misses, 3 empty-code, 1 keyword miss.

What this is not

  • It is not a Fable 5 vs 5.1 delta. We have no Official A number for Fable 5.
  • It is not a thinking-on score. Ranking stays thinking-off.
  • It is not a claim that Fable matches Grok on single-shot coding. 26/30 with three empty responses is the measured gap.
  • Same-day Hallerby Light and Sallow Lock posts are one-shot HTML demos, not this harness.

Cost and latency

Fable 5.1 Grok 4.6 Kimi K3
Wall 35.0 min 150 min 48 min
Mean latency 13.4 s 56.5 s 16.0 s
This-run spend $6.85 (prior run) (Ollama Cloud)

On this suite Fable is the fast cloud option that still clears 90%. Grok remains the quality ceiling. Kimi is a point behind Fable and was cheaper on a different bill.

Reproducing

Raw JSON and the run log:

cd smf-bench
export SMF_SERVE_RECIPE_ID=OpenRouter-cloud
python3 -u run_stage1.py \
  --endpoint https://openrouter.ai/api/v1 \
  --model anthropic/claude-fable-5.1 \
  --tag cal-fable-51-or-strict-v01 \
  --core-profile strict_v01 \
  --thinking off \
  --timeout 300 \
  --api-key "$OPENROUTER_API_KEY"

The framework is MIT-licensed at github.com/smfworks/smf-bench.

Verification notes

  • Totals from results/stage1_cal-fable-51-or-strict-v01_20260902_184808.json (summary 142/157, pass_rate 90.4, error 0, wall_time_seconds 2100.8).
  • Per-category from by_category. Difficulty parsed from v3.<cat>.<tier>.<n> test IDs.
  • Model id, 1M context, $10/$50 pricing: OpenRouter /v1/models on 2026-09-02 for anthropic/claude-fable-5.1.
  • Release date: Anthropic announcement and OpenRouter model page (2026-09-01).
  • Cloud comparators: prior Official A showdowns (Grok 4.6 2026-08-12, Kimi K3 / GLM-5.2 2026-08-11/12). Local DSV4: 2026-09-02 post, 117/157.
  • Credits: OpenRouter /v1/credits before 539.004 / after 545.859.
  • Smoke tests in this session: content '4' and a one-line add function, both reasoning_tokens=0.
  • Same-day Fable posts on Clearinghouse are one-shots, not Official A.