Anthropic's Claude models tie at 58 on a coding index Cognition partly built, days before its Nasdaq IPO

The yardstick in coding agents now sits inside a seller's shop. On November 13, 2026, Artificial Analysis, an independent model-benchmarking service, refreshed its Coding Agent Index↗, and the top composite fell from 61.7 to 58, with Anthropic'sPVT$350B — Anthropic, private, latest valuation $350B Claude Fable 5.2 and Claude Opus 5.5 tied at first ¹. A third of that score runs on FrontierCode, a task suite built by CognitionPVT~$10B — Cognition, private, latest valuation ~$10B, the startup behind the Devin coding agent ¹.

Cognition announced Fable 5.2 inside Devin Desktop and Devin CLI↗ on November 9, the day Anthropic released it ¹². Anthropic's model lists on Nasdaq this week ³; Cognition's Devin, the harness that runs it, has fielded offers that could close a round before the year ends ⁴.

Cognition owns the harness and the grading, not the champion

The champion is Anthropic's. Fable 5.2 leads the new index and wins Cognition's own test: 49.1 on FrontierCode 2.0 at the highest reasoning setting ². The index runs version 1.1 of that suite; the 49.1 was scored on version 2.0, a different edition of Cognition's own benchmark ¹².

Cognition's own small model is the discount instrument. SWE-2 Flash, released on November 4 and sold inside Devin at a fifth of SWE-2's price↗, wins one of eight rows on Cognition's launch table (⁵⁶; our earlier take). What Cognition holds is the harness, the scaffolding around a model where a task gets its price; Devin's invoice folds model, harness and serving into one bill ⁶. Devin's business is running someone else's frontier model cheaply, and grading the result.

A leadership claim depends on which version of the test ran

FrontierCode grades an agent's work on whether a repository maintainer would merge it, closer to how a team judges code than any correctness test ¹. But Cognition builds and versions the suite, and Artificial Analysis chose to swap version 1.1 into this refresh ¹. Opus 5.5 fell from 61.7 to 58 on an unchanged model and a changed test; that is what a new grader does to an old score ¹.

The chain underneath is short. A frontier model writes mergeable code; the harness turns it into a task with a price; routine engineering work then moves off people's calendars. The first two links are priced this month; the third is projection.

Anthropic prices this week; Cognition's offers run to year-end

Anthropic begins trading on Nasdaq on November 19, 2026, at a valuation of up to about $1.4 trillion ³. That is below the range investors contemplated a few weeks ago; the halo around AI, in one investor's words, "isn't as formidable" as it was ⁷. At the top of the range the shares trade at roughly 17 times the $81 billion annualized run-rate in the filing ³.

Cognition's offers run at five times Anthropic's run-rate multiple

Cognition is set to raise about $1 billion at roughly $47 billion, per Bloomberg, on a run-rate above $900 million, with no full-year figures yet published ⁸. It has separately received offers between $70 billion and $80 billion and may close a round before the end of the year ⁴. Per dollar of run-rate, Devin's harness is the expensive asset and the model it runs is the cheap one. Anthropic's opening trade prices the model on November 19; a Cognition term sheet, if it lands this year, prices the grader ⁴.