Grok 4.6 vs Sol vs Opus 5 vs Fable 5: Read the Tiers

xAI launched Grok 4.6 on August 12, 2026 with a ten-row benchmark table comparing it against GPT-5.6 Sol and Claude Fable 5 — and that table does something unusual for a vendor launch asset: every one of its 31 publicly checkable cells reproduces its cited source exactly. It is also a masterclass in why accurate numbers are not the same thing as a fair comparison.

Two levers shape what the table appears to say. The first is the effort tier: xAI benchmarked Grok 4.6 at its own default setting while quoting both rivals at their documented ceilings — an asymmetry that, counter to the cynical read, understates Grok 4.6 and even cost xAI a leaderboard win it could have claimed. The second is column selection: Claude Opus 5 leads seven of the eight benchmarks in this comparison that can be checked on independent boards, and it appears in none of the table’s four columns.

This analysis lays the three vendors’ effort ladders side by side, re-runs xAI’s rows against the public leaderboards they cite, separates the statistically real wins from the ones inside the error bars, and prices all four models by surface. If you route real budget on launch tables — as a follow-up to our previous-generation three-way comparison — this is the reading order that keeps you honest.

The launch table on x.ai/news/grok-4-6 lists ten benchmark rows across four model columns: Grok 4.6 (High), Grok 4.5 (High), GPT-5.6 Sol (Max) and Claude Fable 5 (Max), with the best score in each row set in bold. We checked every cell that has a public counterpart — 31 of them, spanning Terminal-Bench v3.0, DeepSWE v1.1, CursorBench v3.2, APEX-Agents, APEX-SWE, GDPval-AA v2 and AA-Briefcase — against each benchmark’s own public board: frontierbench.ai, deepswe.datacurve.ai, cursor.com/cursorbench, mercor.com/apex and artificialanalysis.ai/evaluations. Every checkable number reproduces its source exactly.

The table is not even arranged to hide defeats. Fable 5 wins five of the ten rows outright to Grok 4.6’s three, and xAI printed rows it loses badly — including Terminal-Bench v3.0 and DeepSWE v1.1. The launch mechanics, pricing and agentic results have their own write-up in our companion analysis of the Grok 4.6 launch; this post is about how the table frames those accurate numbers.

That footnote — together with its sibling line, “Competitor figures are drawn from the respective developers’ published system cards or benchmark leaderboards” — is the load-bearing caveat of the whole asset. It means the four columns were not run under one harness by one operator. Each cell is the best number available from somewhere, produced under whatever conditions that somewhere used. Accurate transcription, heterogeneous measurement.

Reasoning-effort tiers are the units of every 2026 benchmark number, and the three vendors build their ladders differently. Per xAI’s developer documentation, Grok 4.6’s reasoning_effort parameter has exactly four rungs — low, medium, high, xhigh — and the docs are explicit: “If not specified, reasoning_effort defaults to ‘high’. Reasoning cannot be disabled.” The top rung is documented as “xhigh — Maximum reasoning depth… available on grok-4.6.” There is no max rung, and Grok 4.5 does not support xhigh at all — requests sent to it with xhigh are silently treated as high.

OpenAI documents six effort rungs for GPT-5.6 Sol, from none up to a max ceiling, with medium as the default. Anthropic documents five rungs for Claude Opus 5 and Fable 5, from low up to max, with high as the default. Both rivals’ max ceilings are independently corroborated: Artificial Analysis’s board carries max-labelled runs for GPT-5.6 Sol and Claude Opus 5, and OpenAI’s live pricing page carries the tier. Lay the three ladders side by side — no vendor page does — and the launch table’s asymmetry becomes legible.

The asymmetry is not hypothetical — one public board measures it directly. Cursor’s CursorBench v3.2 leaderboard publishes scores by model and effort tier, so the same model appears at several rungs with a per-task cost attached. Grok 4.6 has no Max row there, which corroborates the four-rung ladder. What it does have is an Extra High row that xAI chose not to print for itself.

Read the top four bars. Grok 4.6 at Extra High scores 70.8% on Cursor’s private task suite — rank 1 on the board — above Fable 5 Max’s 70.5% at roughly a sixth of the per-task cost, $2.81 against $17.32. The figure xAI actually published is the 69.9% High row. By quoting its default instead of its ceiling, xAI gave up the rank-1 claim it could have made — the single clearest piece of evidence that the effort-tier asymmetry runs against xAI’s interest, not for it. How this board has evolved as a vendor-adjacent benchmark is a story we covered in our CursorBench v3.1 analysis; the version here is v3.2.

Artificial Analysis published its own independent Grok 4.6 score on launch day: 61 on the AA Intelligence Index, ranked #6 of the 183 models AA had evaluated (the median score is 34), labelled “Grok 4.6 (high).” That is genuine third-party corroboration, and it matches xAI’s rounded claim exactly. It also exposes what the four-column frame leaves out.

Claude Opus 5 at max tops Artificial Analysis’s entire board at 63.05 — above every model in every column of xAI’s table — and it is not in the table. The pattern repeats across the comparison’s independently checkable benchmarks: Opus 5 leads seven of the eight, including the AA Index (63.05), GDPval-AA v2 (an 1849 Elo rating against the benchmark’s human-expert anchor of 1000), AA-Briefcase (1715 Elo on the same style of scale), Terminal-Bench v3.0 (43.5% of its roughly 75 tasks), DeepSWE v1.1 (74% of its 113-task board), and the top position on both of Mercor’s public APEX leaderboards.

This is an omission, not a fabrication — and the distinction is the post’s whole point. Bolding the best score in each row is only meaningful relative to the field the table defines, and this field excludes the model that would take most of the bold. Opus 5’s own launch numbers, and the rarer story of one of them being independently administered, are covered in our Opus 5 launch analysis and our piece on ARC Prize’s verified Opus 5 result.

Grok 4.6’s three row wins on its own table — GDPval-AA v2, AA-Briefcase and Harvey LAB — are also the comparison’s most fragile numbers. Artificial Analysis publishes 95% confidence intervals for its evaluations, so significance is checkable rather than a matter of opinion.

On GDPval-AA v2 — a Bradley-Terry Elo rating anchored to human experts at 1000, not an arena-style crowd vote — Grok 4.6’s 1753 carries an interval of 1733–1774, which overlaps Fable 5’s 1724–1757 and Sol’s 1711–1744. The win is inside the noise. Opus 5’s 1849, interval 1826–1871, overlaps none of the three: its 96-point lead over Grok 4.6 is the row’s one statistically significant gap, and it belongs to the model with no column. On AA-Briefcase the same shape repeats: Grok 4.6’s 1576.59 (1565–1588) against Fable 5’s 1573.78 (1564–1584) is a 2.8-point gap well inside overlapping intervals, while Opus 5’s 1714.59 (1704–1727) is 138 points clear with no overlap.

The third win, Harvey LAB, is a legal-agent benchmark of 120 held-out tasks — so a single task is worth 0.83 percentage points. Grok 4.6’s claimed 15.8% is roughly 19 tasks. In a two-proportion test that clears significance against Sol’s 2.5% (p=0.0003) but not against Grok 4.5’s 12.92% (p=0.52) or Fable 5’s 11.25% (p=0.30) as scored on Vals AI’s public board. And Grok 4.6 itself is absent from that public board, where two versions of Meta’s Muse Spark score higher than its claimed figure — 25.42% and 20.00% of the same 120 tasks. On the board xAI’s row implies it tops, the claimed score would rank third.

Effort rungs are one axis of selection. Benchmark versions and subsets are the other, and they are easier to miss because the names stay the same while the test changes underneath.

FrontierCode v1.1 defines three nested subsets: Extended is all 150 tasks, Main is the hardest 100, Diamond the hardest 50. xAI’s table reports the Extended subset — the easiest of the three — and Fable 5 led that row on xAI’s own table anyway. On Cognition’s public Main board, the best run per model scores: Fable 5 at 53.5%, Opus 5 at 53.4%, Grok 4.6 at 48.0% and Sol at 47.5% of the hardest 100 tasks. The ordering matches xAI’s row; the gap to Fable 5 stretches to 5.5 points on the harder subset. Scores are not comparable across subsets — which also means Anthropic’s self-published 53.5% (a Main figure) does not contradict xAI’s Extended row. They are different tests wearing one name.

Terminal-Bench is the sharpest version trap in the table. On v3.0 — a separate, roughly 75-task suite — the public board’s rival cells match xAI’s table exactly: Sol at 34.6%, Fable 5 at 34.1%. Grok 4.6’s 26% appears on no public v3.0 board — it is vendor-only. Each public row also runs under a different agent harness (Codex for Sol, Claude Code for Fable 5, Cursor CLI for Grok 4.5), and xAI discloses no harness for its own 26%. The board’s outright leader is, again, Opus 5 (max) at 43.5% — 17.5 points above Grok 4.6’s claimed score, and absent. Yet on Terminal-Bench v2.1, the version Artificial Analysis’s index actually uses, the story inverts into a near-tie: Sol xhigh 89.5%, Opus 5 max 89.1%, Grok 4.6 high 88.4%, Sol max 88.0%, Fable 5 84.6% of that suite’s tasks in AA’s runs. Any “crushed at terminal work” reading is v3.0-specific. State the version, every time.

Strip away the framing and a strong, defensible claim survives: at its default effort, Grok 4.6 statistically ties GPT-5.6 Sol at maximum effort on an independent index, at $0.84 per AA Index task against Sol max’s $1.23 — with Opus 5 max at $2.34 and Fable 5 at $3.14 on the same cost measure. At its own xhigh ceiling it takes rank 1 on CursorBench v3.2 for $2.81 a task against Fable 5 Max’s $17.32. The honest headline is not “new frontier leader” — it is Sol-class intelligence at roughly two-thirds of Sol’s per-task cost, while Anthropic’s best stays ahead on capability.

Per-task economics then depend entirely on which pricing surface your traffic hits — and none of these models is one price. Every figure below is an API list price per million tokens from the vendor’s own pricing page at the time of writing; how the two closest rivals price against each other is unpacked further in our Sol vs Fable 5 price-and-access comparison.

One more surface rule: reseller listings do not reliably mirror vendor lists. On OpenRouter, Grok 4.6 and GPT-5.6 Sol match their vendors’ standard list prices, but some of Sol’s smaller siblings list at rates that match other OpenAI surfaces instead — so check each model’s listing against the vendor page rather than generalizing from one. Catalog hygiene of exactly this kind — prices, dates and surfaces — is the subject of our companion guide to reading model catalogs.

The Grok 4.6 table is unusually instructive because it is honest at the cell level and shaped at every other level. That combination is the future of vendor benchmarking: as fact-checking gets easier, selection — of rungs, columns, versions and subsets — replaces fabrication as the lever. Four checks catch most of it.

The deeper skill — reading what a vendor chose to disclose, and what the disclosure pattern itself tells you — is one we apply across launches in our companion piece on reading disclosed losses in vendor tables. And if your team is choosing models with real budget attached, benchmark tables should only ever be the shortlist stage: run your own evals on your own tasks before you commit. That comparative eval discipline is exactly where our AI transformation engagements usually start.