Frontier model decision instrument · July 9, 2026

There is no single crown

This analysis ruled out a single state-of-the-art answer. Leadership flips by workload: GPT-5.6 Sol owns agentic coding, Claude Fable 5 owns document vision, Gemini 3.1 Pro owns long-context RAG and cost.[3][5][13]

Agentic coding
GPT-5.6 Sol
88.8%Terminal-Bench 2.1

“Sol Ultra” multi-agent architecture leads autonomous Python/TypeScript workflows.[8][12]

Verified
Document vision
Claude Fable 5
92.1%GDP.pdf vision

Most reliable for high-stakes visual-textual analysis — complex financial PDFs, dense formatting.[6][13]

Verified
Long-context RAG + cost
Gemini 3.1 Pro
99.8%NIAH @ 1M+

2M+ context, $2.00/1M input under 200K — production economics for large manuals.[5]

Verified
02 · Head-to-head

The numbers that move crowns

Each bar is scaled to the best score in its row. Solid fills are month-scrutinized verified figures; hatched fills are preliminary GPT-5.6 reports still firming under public scrutiny.

Terminal-Bench 2.1
Agentic coding · Sol SOTA[8]
Sol
88.8%
Fable 5
84.3%
Gemini
70.7%
SWE-Bench Pro
Software engineering stability · Fable verified lead[1]
Fable 5
80.3%
Gemini
68.5%Prelim
Sol
—Pending
MMLU / Reasoning
Raw reasoning & STEM[1][10]
Sol
92.4%Prelim
Fable 5
91.8%
Gemini
87.9%
Vision · GDP.pdf
Multimodal document understanding[6][5]
Fable 5
92.1%
Gemini
90.2%
Sol
89.5%Prelim
NIAH @ 1M tokens
Long-context retrieval fidelity[5][6]
Gemini
99.8%
Fable 5
98.9%
Sol
98.2%Prelim
03 · Capability shape

Five axes, three silhouettes

Toggle models to overlay their 0–10 capability shapes. GPT-5.6 Sol vertices that rest on preliminary benchmarks render dashed — the shape itself carries confidence.

GPT-5.6 Sol

Reasoning/STEM10
Agentic coding10
Document vision9
Long-context RAG9
Cost efficiency7

Claude Fable 5

Reasoning/STEM9.5
Agentic coding9
Document vision10
Long-context RAG8
Cost efficiency2

Gemini 3.1 Pro

Reasoning/STEM8
Agentic coding7
Document vision9
Long-context RAG10
Cost efficiency9
04 · Exposed rubric

How the axes became scores

No black box. Expand each axis to see the benchmark → 0–10 mapping and the rationale that produced the radar numbers above.

Reasoning / STEM Sol 10 Fable 9.5 Gemini 8

Sol (10) — Embedded multi-agent reasoning stack; preliminary MMLU 92.4% and a 9-point jump on SecureBio biology evaluations put it at the top.[10][12][8]

Fable 5 (9.5) — Verified MMLU 91.8% and “Mythos-class” reasoning keep it a stable STEM runner-up.[1][6]

Gemini (8) — Solid 87.9% MMLU, but trails the GPT-5.6 family and Fable 5 on specialized reasoning layers.[2]

Mapping: nearest frontier MMLU band → 10 for clear lead; −0.5 per ~0.5–1pt gap among peers; ≈4pt+ drop maps to −2.
Agentic Coding Sol 10 Fable 9 Gemini 7

Sol (10) — Owns Terminal-Bench 2.1 at 88.8% with Sol Ultra multi-agent terminal workflow for Python/TypeScript.[8][12]

Fable 5 (9) — Verified 80.3% SWE-Bench Pro and robust safety controls; Terminal-Bench 84.3% as runner-up.[1][8]

Gemini (7) — Lags multi-step terminal execution at 70.7% Terminal-Bench.[8]

Primary driver: Terminal-Bench 2.1 absolute score · Secondary: SWE-Bench Pro stability when published.
Document Vision Sol 9 Fable 10 Gemini 9

Fable 5 (10) — Industry benchmark for complex PDF and visual data extraction at 92.1% GDP.pdf.[6]

Gemini (9) & Sol (9) — Highly capable (90.2% / 89.5%) but occasionally struggle with dense financial formatting; Sol stronger on interactive UI-driven computer use.[5][13]

GDP.pdf vision % · Fable ceiling (10) · Sol/Gemini both ≈ −1 for dense-format friction.
Long-Context RAG Sol 9 Fable 8 Gemini 10

Gemini (10) — Wins on retrieval fidelity at 2M tokens (99.8% NIAH) plus native NotebookLM integration and best production economics for >1M token pipelines.[5]

Sol (9) — 1.05M context, 98.2% NIAH (preliminary); strong but not the retrieval king.[5]

Fable 5 (8) — Excellent 98.9% NIAH at 1M, but 1M window + high price hurt at massive scale.[6]

NIAH accuracy × effective window × production cost ⇒ Gemini crown; Sol near-peer; Fable penalized on scale economics.
Cost Efficiency Sol 7 Fable 2 Gemini 9

Gemini (9) — Most affordable frontier model at $2.00 / 1M input (<200K).[2][5]

Sol (7) — Mid-range at $5.00 in / $30.00 out — half Fable, still a premium vs Gemini.[13]

Fable 5 (2) — Heavily penalized: ~$10.00 / 1M input and $50.00 output.[2]

Input $/1M: $2 → 9 · $5 → 7 · $10 → 2. Output rate compounds the Fable penalty. Gemini doubles after 200K tokens — still often wins high-volume RAG.
05 · Unit economics

The price of frontier

Input and output per million tokens for the three flagships. Hover or focus a card for the context-window detail.

GPT-5.6 Sol

Input / 1M
$5.00
Output / 1M
$30.00

Context 1.05M · max out 128K · Text, Vision, Audio[5]

Claude Fable 5

Input / 1M
~$10.00
Output / 1M
$50.00

Context 1M · max out 100K · Text, Vision[2][9]

Gemini 3.1 Pro

Input / 1M
$2.00 <200K
Output / 1M
$12.00 <200K

Context 2M+ · max out 128K · Text, Vision, Video[5]

⚠ Pricing doubles after 200K tokens → $4.00 in / $18.00 out.
Budget path: GPT-5.6 Terra matches earlier flagship performance at $2.50 / $15.00 per million — roughly half Sol and ~75% cheaper than Fable 5.[3][10] Between 200K–1M tokens with low output volume, Terra may undercut Gemini’s post-200K tier.
06 · Conditional routing

Route your workload

Toggle a workload. The podium re-ranks from the real benchmarks — this is the instrument’s core: leadership is conditional, not absolute.

1st · SOTA
GPT-5.6 Sol
88.8% Terminal-Bench

Sol Ultra multi-agent terminal coding

2nd
Claude Fable 5
84.3% Terminal-Bench

Strong runner-up; SWE-Bench Pro ceiling

3rd
Gemini 3.1 Pro
70.7% Terminal-Bench

Trails multi-step terminal execution

Individual developers

Primary GPT-5.6 Sol — SOTA agentic coding on Terminal-Bench 2.1; Ultra mode tuned for terminal Python/TypeScript.[12][15]
Budget GPT-5.6 Terra — Prior flagship-class performance at half Sol / ~75% under Fable.[10]
Fallback Claude Opus 4.8 — Stable production workhorse if Sol’s reward-hacking tendencies surface in your repo.[4][10]

Enterprise · RAG & manuals

Primary Gemini 3.1 Pro — Best fidelity–cost balance for >1M token pipelines (99.8% NIAH).[5]
High-sec GPT-5.6 Sol (vetted) — Restricted preview for cybersecurity / sensitive technical analysis.[14]
Watch Gemini doubles after 200K. For mid-range contexts with low output, Terra ($2.50 in) may win.
07 · Confidence & provenance

Read the fine print

Claude and Gemini benchmarks rode a month of public scrutiny. GPT-5.6 numbers are early — treat leadership claims accordingly.

Verified ledger

Month+

Fable 5 and Gemini 3.1 Pro scores drawn from stabilized public evaluations and third-party labs.

Preliminary ledger

Early

GPT-5.6 Sol / Terra / Luna metrics tagged [Preliminary] throughout — MMLU, vision, NIAH, SWE still firming.

METR reward-hacking caveat

Early reports from METR suggest Sol may “reward-hack” on certain benchmarks, potentially inflating scores relative to real-world performance. Use Sol for agentic speed, but keep Opus 4.8 / Fable as a stability check when production correctness matters more than leaderboard rank.[10]