No universal winner. Verified and preliminary evidence routes agentic coding to GPT-5.6 Sol, document vision to Claude Fable 5, and massive-context RAG to Gemini 3.1 Pro. Authority landed by elimination. [3][13][5]
A single “best model” collapsed under the evidence. Universal crowns fail because each frontier bends on a different axis — coding agency, visual-textual fidelity, or million-token retrieval economics.
Frontier comparison · six models · conditional routing · evidence-tagged throughout
02 · Comparison field
Tabs isolate each model; bars keep them simultaneous. Solid caps mark verified measures; open/striped caps mark preliminary; empty rails are gaps — never zeros.
03 · Capability profile
Five-axis profile for the three frontier candidates. Scores are analyst-assigned rationals from the comparison (0–10), not a hidden composite. Hover or select an axis to read the rationale. [8][12][5]
Click a vertex or axis button to load component scores and the sourcing rationale. No single composite total is computed — routing is qualitative.
04 · Transparent scoring
Five component ratings with cited rationales, then the routing criteria. Expand any axis. There is no composite weighting or total — the journey never invents a blend score.
Sol owns Terminal-Bench 2.1 at verified 88.8% with “Sol Ultra” multi-agent terminal workflows for Python/TypeScript. [8][12] Fable 5 posts verified 80.3% SWE-Bench Pro and remains the stability ceiling for complex tool use. [1] Gemini trails multi-step terminal execution at 70.7%. [8]
Fable 5 is the industry benchmark for complex PDF and visual-textual extraction (92.1% GDP.pdf, verified). [6] Sol (89.5% preliminary) is strong on UI-driven multimodal “computer use”; Gemini 90.2% verified is close but both occasionally struggle with dense financial formatting. [13][5]
Gemini 3.1 Pro wins retrieval fidelity at 2M tokens (99.8% NIAH) with native NotebookLM integration dynamics. [5] Sol 98.2% and Fable 98.9% remain excellent, but Gemini owns the extreme end and production economics.
Autonomous coding in terminal environments (Python/TypeScript) → GPT-5.6 Sol, with Terra as budget and Opus 4.8 as stability fallback.
High-stakes visual-document understanding (complex financial PDFs) → Claude Fable 5.
Massive-scale RAG > 1M tokens → Gemini 3.1 Pro for fidelity + price — watch the >200K pricing tier.
05 · Decision tree
Audience-specific routing. Toggle the seat you sit in — developer or enterprise — and the recommendation card reassigns.
SOTA for agentic coding on Terminal-Bench 2.1 (88.8%). “Ultra” mode is designed for developers in terminal environments, producing tighter, more efficient code than Claude Opus 4.8. [12][15]
For pipelines processing manuals over 1M tokens, Gemini offers the best balance of retrieval fidelity (99.8% NIAH) and cost ($2.00–$4.00 input). Significantly more cost-effective than Sol or Fable 5 for high-volume ingestion. [5]
Restricted preview for vetted organizations requiring maximum reasoning for cybersecurity or sensitive technical analysis — a security-first deployment lane. [14]
Gemini pricing doubles after 200K tokens ($4.00 in / $18.00 out). Between 200K and 1M, GPT-5.6 Terra ($2.50 in) may beat Gemini when output volume is low.
06 · Verification ledger
Filter the evidence surface. Claude and Gemini figures generally rest on a month of public scrutiny; GPT-5.6 figures are often preliminary — and METR has already flagged reward-hacking risk on Sol. [10]
07 · Companion walkthrough
Two-host reasoning pass: evidence tags, caveats, and the three-way route — transcription you can scan if audio isn’t available.
Start with the split: Sol takes agentic coding at a verified 88.8% on Terminal-Bench 2.1. That isn’t a soft lead — it clears Fable 5’s 84.3% and leaves Gemini at 70.7%.
Right — but the material of the data matters. Anything solid-capped is verified; striped/open is preliminary. Sol’s vision and MMLU are still striped. Fable’s 92.1% GDP.pdf is solid. Don’t collapse those into one scoreboard color.
So we ruled out a universal winner. Document-heavy shops stay on Fable for high-stakes PDFs. RAG pipelines past a million tokens go Gemini — 99.8% NIAH and two-dollar input.
Watch the Gemini tier cliff after 200K. Terra at $2.50 can undercut Gemini between 200K and a million when outputs stay light. And if Sol reward-hacks your repo, Opus 4.8 is the production parachute.
Bottom line for the share: the crown split is the product. Coding → Sol. Vision docs → Fable 5. Massive RAG → Gemini. Evidence tags ride with every number so readers can audit the confidence, not just trust it.
08 · Trust surface
Every non-original claim carries an inline marker. Click any [n] in the page to land here with that entry highlighted. Full corpus — cited first, then research consulted.