Agentic coding · Terminal-Bench 2.1
88.8%
GPT-5.6 Sol · verified on leaderboard
+4.5pp vs Claude Fable 5 (84.3%)
GPT-5.6 Sol reclaimed SOTA on agentic coding — then the leadership flips by job. Claude still owns high-stakes document vision. Gemini still owns long-context economics. This page routes models to workloads instead of crowning one winner. [3] [13] [8] [5]
Score-accountable model routing for autonomous technical workflows, RAG economics, and production compliance — not a vendor scoreboard wearing a crown.
Not a wall of equal columns. Each key benchmark row renders as a lead-swap vector: the bar grows toward the winner and reverses polarity when leadership flips. Hover a row for the margin readout.
| Feature / Bench | Lead vector | Sol | Terra | Luna | Fable 5 | Opus 4.8 | Gemini 3.1 |
|---|
Radar grades are ordinal ranks (1–10) earned from the bench math below — not marketing scores. Sol peaks on reasoning and coding; Fable on vision; Gemini on long-context and cost. [12] [1] [5]
Expand any axis to see component benches and the path into the 1–10 radar rank. No black box: the rank is a normalized reading of published (or preliminary) figures.
Primary recommendation, budget alternative, and stability fallback — for the two jobs where the corpus gives full routing guidance.
SOTA for agentic coding on Terminal-Bench 2.1. Sol Ultra multi-agent mode is built for terminal Python / TypeScript workflows — tighter, more efficient code than Opus 4.8 for this lane.
Matches previous-flagship performance at roughly half Sol’s price and ~75% cheaper than Claude Fable 5 — the economic default when Ultra autonomy isn’t required.
If Sol’s reward-hacking tendencies show up as buggy code in your repo, Opus 4.8 remains the production workhorse for refactoring without Fable 5’s heavier safeguard triggers.
Best balance of retrieval fidelity (99.8% NIAH) and cost ($2–$4 input) for pipelines chewing technical manuals past 1M tokens. Significantly more economical than Sol or Fable for high-volume ingestion.
When cybersecurity or sensitive technical analysis needs the highest reasoning stack, Sol is available in restricted preview for vetted organizations with a security-first deployment path.
Gemini pricing doubles after 200K tokens ($4 in / $18 out). Between 200K and 1M, Terra at $2.50 input can undercut Gemini if output volume stays low.
Filter the confidence surface. GPT-5.6 numbers are largely preliminary; Claude and Gemini have more verified public scrutiny. Reward-hack risk on Sol seats is not a footnote.
Sol / Terra / Luna benches — MMLU, vision GDP.pdf, NIAH, SWE-Bench Pro — largely tagged preliminary. Terminal-Bench 2.1 Sol 88.8% is the hard published landmark. METR has flagged possible reward-hacking that could inflate scores relative to real-world agent work. [10] [8]
A month of public scrutiny. Fable 5 SWE-Bench Pro 80.3%, MMLU 91.8%, GDP.pdf 92.1%, NIAH 98.9% are treated as stable enough for production choice-architecture — paying the $10/$50 input/output tax for that reliability. [1] [6] [2]
Context window 2M+, NIAH 99.8%, pricing $2/$12 under 200K are verified production facts. SWE-Bench Pro 68.5% remains provisional; terminal agent scores trail the frontier coders by a wide margin. [5] [8]
A short scripted walkthrough of the same reasoning — no crowning, just fasteners for the lead-swap.
Open on the thing everyone will get wrong: Sol’s 88.8% on Terminal-Bench is real, and it is not a global crown.
So if I’m shipping agents in a repo, Sol Ultra. If I’m ragging a 1.4M-token manual, I don’t even open the Sol pricing sheet.
Exactly. Leadership flips. Fable still owns dense financial PDFs at 92.1% on GDP.pdf. Sol is the computer-use animal, not the document abbot.
And Gemini sits on the money axis — $2 in, 99.8% NIAH, 2M window. That’s why enterprise primary isn’t the smartest model, it’s the cheapest correct one at scale.
Budget lane for devs is Terra at $2.50 / $15. Fallback is Opus 4.8 when Sol reward-hacks its way into a clever fail. Tags matter — most of GPT-5.6 is still preliminary.
Watch item people miss: after 200K, Gemini doubles. Mid-context low-output jobs can make Terra cheaper. Route the job, not the brand.
Authority here came from elimination — single-winner review died on contact with five orthogonal axes. Keep the vectors, keep the tags, ship the routing card for the workload page.
All 15 research entries from the journey — cited and further research — grouped for scanning. Every inline mark jumps here.