Glance
2026-11-19 14 MODELS 9 LABS

Frontier LLM Picks & Real Task Costs

01

Quick Picks

Pick a workflow for default, budget, and max options

Default CursorBench #2

Gemini 4 Neon

high
Google
Per task (Cursor)
$6.54
$3/$15 Output 62k Hallucination 13.9%
Why Ranks 2nd on CursorBench at $6.54/task, 45% below Opus 5.5 max.
Caveat Trails Opus on pure terminal tasks (65.8%); peak SWE relies on official CLI.
Budget Budget CLI

GPT-6.2 Sol

max
OpenAI
Per task (AA)
$1.01
$2/$10 Output 40k Hallucination 55.4%
Why 61.5% on Terminal-Bench and $1.01/task AA spend via 40k output cap.
Caveat Not yet on CursorBench 4.1; 55.4% hallucination rate; 272K context tier step.
Max CursorBench #1

Gemini 4 Neon

xhigh
Google
Per task (Cursor)
$10.67
$3/$15 Output 142k Hallucination 14.4%
Why Top multi-file refactor score (65.4%) with 65.2% Terminal-Bench eval.
Caveat Costs 63% more per task than Neon high ($10.67 vs $6.54) for +1.7 pts.
Agent Synthesis

Key Changes

Read (3)
  • 01
    Frontier leaders

    Neon takes the top two CursorBench spots (65.4% / 63.7%), Astra holds AA spend to $4.00 (62.1), and Fable 5.2 leads AA at 64.6 ($26.54).

  • 02
    Open & budget

    Kimi K3.1 Pro tops open weights (46.5 AA, 59.7% Terminal-Bench), DeepSeek V4.1 Flash runs at $0.60/task (MIT), and Meta Muse Ember sets a $0.14 floor (36.0).

  • 03
    Pricing note

    Gemini 4 Carbon's $2/$10 promo ends Nov 30 (reverts to $4/$20 Dec 1; standard AA bill $7.76). Qwen4 Max (42.9, $3.27) is closed API only.

02

Cost & Capability

Hidden thinking tokens drive the real bill, not list prices

COST RANGE 190×
$0.14 → $26.54
Muse Ember max 36.0 Fable 5.2 max 64.6
RATE VS REAL BILL 3.1× cheaper
Sol max $2/$10 $1.01 40k
Hydrogen xhigh $1.5/$7.5 $3.16 110k
List ×1.3 · Bill $1.01 vs $3.16
OPEN VS CLOSED -18.1 AA
Fable 5.2 max Closed #1 64.6
Kimi K3.1 Pro Open #1 46.5
Open min · DeepSeek V4.1 Flash ($0.60)

Capability vs Cost

Closed Open Top-left = best value
Capability vs Cost <$1 $1–$5 $5–$15 $15+ 32 40 48 56 64 $0.1 $0.25 $0.5 $1 $2 $5 $10 $25 $50 Real Cost per Task (USD, log scale) Fable 5.2 max Astra max Neon xhigh Carbon high Neon high Opus 5.5 max Sol max Hydrogen xhigh Kimi K3.1 Pro Qwen4 Max Muse Ember max Grok 4.7 xhigh DeepSeek V4.1 Flash GLM-5.3 max

Pareto frontier: Muse Ember max ($0.14) → Sol max ($1.01) → Neon high ($3.74) → Astra max ($4.00) → Fable 5.2 max ($26.54).

Hover or tap node

Thinking Tokens & Bills

Thinking Answer
Direct Source · Artificial Analysis v5.0.1
Claude Fable 5.2 max
193k + 81k = 274k
Claude Opus 5.5 max
120k + 53k = 173k
Gemini 4 Carbon high
96k + 58k = 154k
Gemini 4 Neon xhigh
81k + 61k = 142k
Gemini 4 Hydrogen xhigh
68k + 42k = 110k
Gemini 4 Neon high
35k + 27k = 62k
GPT-6.2 Astra max
26k + 15k = 41k
GPT-6.2 Sol max
23k + 17k = 40k

AA v5.0.1 averages (thinking + answer); bills include input and cache.

03

Full Benchmarks

Raw scores and per-task spend, filterable and sortable

Direct Sources · 14 models · Click column to sort

↗
↗
↗
↗
↗
↗
↗
Claude Fable 5.2 ◈ max
$8 / $40
64.6
$26.54
61.3% $48.11
65.8%
— xhigh: 50.2% 59.3%
GPT-6.2 Astra ◈ max
$8 / $40 ≤272K
62.1
$4.00 —
62.1%
— 6.0: 44.0% 59.1%
Gemini 4 Neon xhigh
$3 / $15
61.4
$6.44
65.4% $10.67
65.2%
DeepMind eval
— 14.4%
Gemini 4 Carbon high
$2 / $10* Dec 1: $4/$20
60.0
$7.76
61.5% $8.52
64.5%
— 11.4%
Gemini 4 Neon ◈ high
$3 / $15
59.6
$3.74
63.7% $6.54
— xhigh 65.2% 46.2% 13.9%
Claude Opus 5.5 max
$4 / $20
59.1
$10.04
58.1% $11.88
65.8%
47.3% 50.3%
GPT-6.2 Sol ◈ max
$2 / $10 ≤272K
57.3
$1.01 —
61.5%
43.8% 55.4%
Gemini 4 Hydrogen xhigh
$1.5 / $7.5
53.9
$3.16
61.4% $8.09
56.4%
high
— 8.5%
Kimi K3.1 Pro default OPEN
$2 / $10
46.5
$3.68 —
59.7%
25.5% AA 2.0 —
Qwen4 Max default
$1.5 / $4.5
42.9
$3.27 — — 26.5% AA 2.0 —
Muse Ember ◈ max
$0.1 / $0.5
36.0
$0.14 — — xhigh 30.6% — 17.4%
Grok 4.7 xhigh
$2 / $6
35.5
$7.77
48.6% $10.19
38.8%
28.9% —
DeepSeek V4.1 Flash max OPEN
$0.3 / $1.2
32.6
$0.60 —
38.5%
— —
GLM-5.3 max OPEN
$1.4 / $4.4
31.7
$5.01 —
41.2%
21.6% AA 2.0 —

Past Snapshots

All snapshots →