Recommendation for Math / Reasoning

Math & Reasoning

Our top recommendation for Math & Reasoning, based on the public evidence we track, is Anthropic: Claude Fable 5.[1] OpenAI: GPT-5.6 Sol is the next-ranked alternative. Deploy for high-stakes mathematics where LiveBench's objective scoring validates strong competition performance.

About this recommendation

Updated
Sep 25, 2026
Evidence through
Sep 25, 2026
Sources
10
Revision
v78

Decision audit

Why this result

Inspect the inputs and the computed order behind the recommendation.

Models screened

21

live candidates

Evaluation feeds

5

task-weighted

Winner coverage

100%

intended feed weight

Largest provider share

3 of 6

Anthropic

Provisional source breadth. 6 citation families and 1 practitioner families support the top result; 0 cautionary threads is retained. The largest citation family contributes 31%.

Sources evaluated

The task sets these weights before any model is scored.

winner: Claude Fable 5
Evaluation feedWeightWinner resultField measured
LiveBench Mathematics
35%
#621/21
LiveBench Reasoning
25%
#1021/21
LMArena Math
25%
#319/21
LMArena Instruction Following
10%
#420/21
OpenRouter usage
5%
79/10021/21

Provider concentration

Each exact model is scored separately; provider identity is not a ranking input.

Anthropic50%
  • Anthropic3 models
  • Google1 model
  • OpenAI1 model
  • Qwen1 model

Decision table

Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.

RankModelRelative scoreCoveragePractitioner evidenceStrongest measured reason
01Claude Fable 5Anthropic
85
100%1 threads · 1 families · 0 cautions#3 LMArena Math · #4 LMArena Instruction Following
02GPT-5.6 SolOpenAI
82
100%1 threads · 1 families · 1 cautions#4 LiveBench Reasoning · #5 LiveBench Mathematics
03Claude Opus 4.6Anthropic
81
100%2 threads · 1 families · 0 cautions#1 LMArena Instruction Following · #5 LMArena Math
04Qwen3.8 27BQwen
75
100%5 threads · 2 families · 0 cautions#25 LMArena Math · #37 LMArena Instruction Following
05Gemini 3.6 FlashGoogle
73
100%no linked practitioner threads#8 LMArena Math · #18 LMArena Instruction Following
06Claude Opus 5.5Anthropic
63
44%no linked practitioner threads#1 LiveBench Mathematics · #2 LiveBench Reasoning

Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.

  1. Anthropic: Claude Fable 5 ranks #4 of 139 on LMArena's maths category (Elo 1522), based on blind human preference votes for maths prompts.

    Best when: Consider only after reviewing the cited caution.

  2. Scores 96.2% on LiveBench Mathematics (#5 of 58) and ranks #13 overall on LMArena, with documented numerical precision issues in optimization code.

    Best when: Deploy for high-stakes mathematics where LiveBench's objective scoring validates strong competition performance.

    Tips

    • Deploy for high-stakes mathematics where LiveBench's objective scoring validates strong competition performance.
      Source 1
      “Scores 96.2% on LiveBench Mathematics (#5 of 58), using objectively scored competition and olympiad-style tasks.”
      LiveBench MathematicsOpen original ↗

    Watch out for

    • Watch for float32 underflow in numerical optimization, specifically when multiplying small gradient terms by Armijo conditions before reduction.
      Source 2
      “INDEPENDENT NUMERICAL REVIEW · Z-Switch-Q9V3 / GPT-5.6 Sol · issue #15960 Fresh review found a blocker in the earlier proposed "scale before reduce" repair: it fixes the original overflow witness but can underflow small Armijo terms to zero and accept an insufficiently improving step. Reproduced float32 witness: 200 coordinates, g_i=1e-35, d_i=-7, armijo=1e-4, step=1. The mathematically required Armijo decrease is nonzero, but multiplying g by armijo first flushes the per-coordinate terms to ze…”
      woahwhattheheckOpen original ↗
  3. Ranks #5 on LMArena's maths category and demonstrates practical simulation coding ability, though LiveBench Reasoning places it at #16 of 58.

    Best when: Use for physics simulation and orbital mechanics code where users report success despite limited formal math background.

    Tips

    • Use for physics simulation and orbital mechanics code where users report success despite limited formal math background.
      Source 3
      “> It's a pretty high moat getting into stuff like simulation software I'm currently working on a simulation game about space and orbital mechanics. I have a lot of software experience, I know how to build large projects and architect my code, and I know how to to test the end result to ensure I'm getting what I want. But I also don't have a strong math or physics background. In my experience, Claude (Opus 4.6+) has had no issues writing any simulation or game related math code. And the key thin…”

    Watch out for

    • Expect reasoning traces to be hidden or truncated, as Anthropic has moved toward concealing chain-of-thought outputs.
      Source 4
      “The initial motivation for this was likely to thwart any competition. Already Anthropic has accused some companies of organized distillation efforts at a massive scale. Back when I used antigravity, it used to show the reasoning intact - at least for Gemini Pro 3.1, and likely for Claude Opus 4.6 (not 100% certain about it). I have some recollection of stopping the models mid-turn when they started going astray. As a power user, I find reasoning fascinating to read and genuinely useful at times…”
  4. An open-weight 27B model with verified functional correctness on math and 60K context retrieval, though LMArena maths ranking sits at #26 of 139.

    Best when: Run locally on RTX 5090 with NVFP4 compression for ~1.6x faster decode, accepting a small quality tax versus Q4_K_M GGUF.

    Tips

    • Run locally on RTX 5090 with NVFP4 compression for ~1.6x faster decode, accepting a small quality tax versus Q4_K_M GGUF.
      Source 5
      “Hardware: 1xRTX 5090 32 GB (SM 12.0), driver 595.84, CUDA 13.2 toolkit Engine: vLLM 0.27.1 (pip) + FlashInfer 0.6.17, torch 2.13.0+cu130 Model: Qwen3.8-27B, compressed-tensors NVFP4 (W4A16, group 16), MTP speculative head in BF16 ### Quality functional correctness verified (math, tool calling, 60K context retrieval, multi-turn); NVFP4 carries a small quality tax vs GGUF Q4_K_M on evals (independent measurement: q_avg 92.5 vs 93.2, HumanEval 89.6 vs 94.5) — the trade for ~1.6x faster decode than…”
    • Deploy for hard problems where benchmark standings include GPQA Diamond 89.2 and LiveCodeBench v6 90.3.
      Source 6
      “**Standing issue — intentionally left open. Revisit periodically; close only if the model is ruled out for good.** ## What this tracks Whether [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) should replace Gemma 4 26B-A4B as the local model, and what would have to change upstream for that to make sense. ## Why it was not adopted on release (2026-08-15) The benchmarks are excellent — GPQA Diamond 89.2, LiveCodeBench v6 90.3, Terminal-Bench 2.1 73.0, SWE-bench Pro 61.7, Apache 2.0, 262K na…”

    Watch out for

    • Tune MTP (Multi-Token Prediction) settings carefully, as configuration choices significantly impact latency and output quality on hard problems.
      Source 7
      “A while back I posted a comparison of Qwen3.8-27B against Nemotron 3.5 Lightning and Muse Glimmer on 16 hard problems. Several people asked what settings I was running, and it turned out I did not have a good answer. So I went back and benchmarked the settings themselves: 45 configurations of this model alone. ![chart-ctx-fullprompt-latency](https://cdn-uploads.huggingface.co/production/uploads/6314e2e71994578f07ed8f89/_H2J3BUONs6AsrR2ruV4a.png) Sharing the findings here because a few of them c…”
      Source 8
      “![chart-qwen-hero](https://cdn-uploads.huggingface.co/production/uploads/6314e2e71994578f07ed8f89/y0d7Q_aB0gLg27LarUwan.png) I ran Qwen3.8-27B against two other models in the same size class on a set of 16 hard problems, and wanted to share the numbers here because a few of them surprised me. ### Setup - One RTX 5090, 32GB - Ollama 0.32.12, tag `qwen3.8:latest` (Q4_K_M GGUF, 27.3B) - 64K context, temperature 0.2, generation cap 32,768 tokens - Every model loaded alone on the GPU, with all other…”
  5. Ranks #7 on LMArena's maths category but scores 85.15% on LiveBench Reasoning, placing #30 of 58 on objective ground-truth tasks.

    Best when: Consider for math tasks where human preference rankings matter more than objective reasoning benchmarks.

    Tips

    • Consider for math tasks where human preference rankings matter more than objective reasoning benchmarks.
      Source 9
      “Ranks #7 of 139 on LMArena's maths category (Elo 1506), based on blind human preference votes for maths prompts.”
      LMArena maths categoryOpen original ↗
  6. Leads the LiveBench Mathematics benchmark at 97.08%, placing first among 58 models on competition and olympiad-style tasks.

    Best when: Use for competition math and olympiad problems where objective scoring confirms top-tier performance.

    Tips

    • Use for competition math and olympiad problems where objective scoring confirms top-tier performance.
      Source 10
      “Scores 97.08% on LiveBench Mathematics (#1 of 58), using objectively scored competition and olympiad-style tasks.”
      LiveBench MathematicsOpen original ↗

Frequently asked

What is an alternative to Anthropic: Claude Fable 5?
OpenAI: GPT-5.6 Sol is the next-ranked option. Deploy for high-stakes mathematics where LiveBench's objective scoring validates strong competition performance.[1]

Sources

  1. 1

    “Scores 96.2% on LiveBench Mathematics (#5 of 58), using objectively scored competition and olympiad-style tasks.”

    LiveBench Mathematics · Benchmark · Jun 25, 2026
  2. 2

    “INDEPENDENT NUMERICAL REVIEW · Z-Switch-Q9V3 / GPT-5.6 Sol · issue #15960 Fresh review found a blocker in the earlier proposed "scale before reduce" repair: it fixes the original overflow witness but can underflow small Armijo terms to zero and accept an insufficiently improving step. Reproduced float32 witness: 200 coordinates, g_i=1e-35, d_i=-7, armijo=1e-4, step=1. The mathematically required Armijo decrease is nonzero, but multiplying g by armijo first flushes the per-coordinate terms to ze…”

    woahwhattheheck · GitHub · Sep 18, 2026
  3. 3

    “> It's a pretty high moat getting into stuff like simulation software I'm currently working on a simulation game about space and orbital mechanics. I have a lot of software experience, I know how to build large projects and architect my code, and I know how to to test the end result to ensure I'm getting what I want. But I also don't have a strong math or physics background. In my experience, Claude (Opus 4.6+) has had no issues writing any simulation or game related math code. And the key thin…”

    pyth0 · Hacker News · Jun 7, 2026
  4. 4

    “The initial motivation for this was likely to thwart any competition. Already Anthropic has accused some companies of organized distillation efforts at a massive scale. Back when I used antigravity, it used to show the reasoning intact - at least for Gemini Pro 3.1, and likely for Claude Opus 4.6 (not 100% certain about it). I have some recollection of stopping the models mid-turn when they started going astray. As a power user, I find reasoning fascinating to read and genuinely useful at times…”

    sheepscreek · Hacker News · Jun 23, 2026
  5. 5

    “Hardware: 1xRTX 5090 32 GB (SM 12.0), driver 595.84, CUDA 13.2 toolkit Engine: vLLM 0.27.1 (pip) + FlashInfer 0.6.17, torch 2.13.0+cu130 Model: Qwen3.8-27B, compressed-tensors NVFP4 (W4A16, group 16), MTP speculative head in BF16 ### Quality functional correctness verified (math, tool calling, 60K context retrieval, multi-turn); NVFP4 carries a small quality tax vs GGUF Q4_K_M on evals (independent measurement: q_avg 92.5 vs 93.2, HumanEval 89.6 vs 94.5) — the trade for ~1.6x faster decode than…”

    marco9899 · Hugging Face · Aug 18, 2026
  6. 6

    “**Standing issue — intentionally left open. Revisit periodically; close only if the model is ruled out for good.** ## What this tracks Whether [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) should replace Gemma 4 26B-A4B as the local model, and what would have to change upstream for that to make sense. ## Why it was not adopted on release (2026-08-15) The benchmarks are excellent — GPQA Diamond 89.2, LiveCodeBench v6 90.3, Terminal-Bench 2.1 73.0, SWE-bench Pro 61.7, Apache 2.0, 262K na…”

    nbramia · GitHub · Aug 15, 2026
  7. 7

    “A while back I posted a comparison of Qwen3.8-27B against Nemotron 3.5 Lightning and Muse Glimmer on 16 hard problems. Several people asked what settings I was running, and it turned out I did not have a good answer. So I went back and benchmarked the settings themselves: 45 configurations of this model alone. ![chart-ctx-fullprompt-latency](https://cdn-uploads.huggingface.co/production/uploads/6314e2e71994578f07ed8f89/_H2J3BUONs6AsrR2ruV4a.png) Sharing the findings here because a few of them c…”

    laxmimerit · Hugging Face · Aug 16, 2026
  8. 8

    “![chart-qwen-hero](https://cdn-uploads.huggingface.co/production/uploads/6314e2e71994578f07ed8f89/y0d7Q_aB0gLg27LarUwan.png) I ran Qwen3.8-27B against two other models in the same size class on a set of 16 hard problems, and wanted to share the numbers here because a few of them surprised me. ### Setup - One RTX 5090, 32GB - Ollama 0.32.12, tag `qwen3.8:latest` (Q4_K_M GGUF, 27.3B) - 64K context, temperature 0.2, generation cap 32,768 tokens - Every model loaded alone on the GPU, with all other…”

    laxmimerit · Hugging Face · Aug 16, 2026
  9. 9

    “Ranks #7 of 139 on LMArena's maths category (Elo 1506), based on blind human preference votes for maths prompts.”

    LMArena maths category · Benchmark · Sep 13, 2026
  10. 10

    “Scores 97.08% on LiveBench Mathematics (#1 of 58), using objectively scored competition and olympiad-style tasks.”

    LiveBench Mathematics · Benchmark · Jun 25, 2026

Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.