Frontier AI benchmarks: summary and insights, week of 2026-11-13

Written from tables for 2026-11-13 (data collected 2026-11-13). If that tables file is collected again, check these insights still match.

Leaders at a glance · week of 2026-11-13

8 benchmarks · 4 publishers

Each tile is the model's rank on that benchmark. Ranks, not scores, because scores from different benchmarks are not comparable.

	SWE Atlas	SWE-Bench Pro V2	Terminal-Bench*	MCP-Atlas	GDPval-AA*	GDPx*	Agent Arena	WebDev Arena
Claude Fable 5.2	1	–	4	–	5 ≈	1	1	1
Claude Opus 5.5	2 ≈	–	4	5	1	2	2	2 ≈
Gemini 4 Carbon	4 ≈	–	2	–	9	3	3	4
Claude Sonnet 5.5	7	7	3	–	4 ≈	–	4	3
Claude Fable 5.1	3 ≈	2 ≈	11	3 ≈	7	–	5	7
GPT-6 Astra	6	5	6	–	–	–	6	5

Stronger tile = higher rank, in the lab's colour · ≈ too close to call with #1 · – not in the rows the table shows · right of the dashed line: public votes · hover a tile for details · * no rank published, ranked by score order

Who is really tied?

All tables

The top models on each benchmark: the dot is the score, the line its 95% interval.

Independent tests · each chart has its own scale

SWE Atlas (Overall)

95% interval shown · Scale Labs

Show all 12

SWE-Bench Pro V2 (Full)

95% interval shown · Scale Labs

Show all 12

MCP-Atlas

95% interval shown · Scale Labs

Show all 12

GDPval-AA

95% interval shown · Artificial Analysis

Show all 12

No margin published · dots only

Terminal-Bench 4.0

no margin published · Artificial Analysis

Show all 12

GDPx

no margin published · GDPx by SemiAnalysis

Public votes (Arena) · kept separate from the tests

Agent Arena

95% interval shown · Arena

Show all 12

WebDev Arena

95% interval shown · Arena

Show all 12

Each colour is the model's lab · full colour = leader or too close to call · faded = behind · shaded band = the leader's 95% interval · hover for details

Summary and insights

Summary

This is the first week of the tracker (collected 2026-11-13), so there are no week-to-week changes yet. No single model leads every test. Among independent tests, Claude Fable 5.2 ranks first on SWE Atlas (too close to call) and GDPx. Separately, in public-vote rankings it ranks first on Agent Arena and WebDev Arena (too close to call there). Other models lead elsewhere: Kimi K3.1 Pro on Terminal-Bench 4.0, Gemini 4 Argon on MCP-Atlas, Claude Opus 5.5 on GDPval-AA and Claude Opus 5 on SWE-Bench Pro V2. Several of these top spots are statistical ties, so the honest headline is "a small group of leaders" rather than one winner.

Insight 1

Claude Fable 5.2 leads most often, but not everywhere.

On SWE Atlas (Scale's test of professional coding work: answering questions about a codebase, writing tests and refactoring), it leads with 80.5%, but Claude Opus 5.5, Claude Fable 5.1 and Gemini 4 Carbon are too close to call (source). On Terminal-Bench 4.0 (coding tasks done in a computer terminal), it scores 59.6%, behind Kimi K3.1 Pro at 67.2%, with no margin published (source). Separately, in public votes on Agent Arena (people rating AI agents on their own real tasks), it leads with a score of 0.303, clearly ahead of the rest (source).

Insight 2

The two work benchmarks disagree on who is first.

On GDPval-AA (realistic office-work tasks, rated with a chess-style Elo score), Claude Opus 5.5 leads with 1866, but Kimi K3.1 Pro (1861), Muse Wonder Full, Claude Sonnet 5.5 and Claude Fable 5.2 are all too close to call (source). On GDPx (very hard expert tasks in fields such as chip design, finance and law), Claude Fable 5.2 leads with 16.4% ahead of Claude Opus 5.5 at 12.0%, though GDPx publishes no margin (source, GDPx by SemiAnalysis). The low GDPx top score shows that even the best models still fail most of these expert tasks.

Insight 3

Higher scores often cost much more per task.

On GDPx, Claude Fable 5.2 scores 16.4% at $88.15 per task, while Gemini 4 Carbon scores 10.9% at $16.29 per task (source, GDPx by SemiAnalysis). In Agent Arena's public votes, Gemini 4 Carbon ranks third at a median cost of $1.46 per task, while the leader, Claude Fable 5.2, costs $7.43 (source). For many everyday uses, a cheaper model near the top may be the better choice.

Insight 4

One coding test is nearly maxed out, so it can no longer pick a winner.

On SWE-Bench Pro V2 (fixing real issues in software projects), the top score is 99.4% (Claude Opus 5), and three other models are too close to call (source). SWE Atlas still spreads models out, from 80.5% at the top to 62.1% at the bottom of the top 12, which makes SWE Atlas the more useful coding signal for now (source).