Fable Costs Twice as Much as Opus. It's Still Cheaper to Run.

Cognition, the company behind the coding agent Devin, has published data showing its most expensive model produces its least expensive agent runs. Fable 5 charges roughly twice what Opus 4.8 charges per token, yet a Fable-led team running Cognition’s Fusion architecture finished benchmark tasks for $1.86 per run, against $2.04 for an Opus-led team on the same tasks. The Fable team also scored higher, 60.7 versus 54.6 on FrontierCode 1.1, Cognition’s internal evaluation for agentic coding.

The finding matters because it undercuts a planning assumption most teams make by default: that a model’s per-token price predicts what an agent will cost to run. Cognition’s post, written by engineer Joon Hee Lee and published on the company’s engineering blog, argues the real cost driver is turn count and context volume. It is not the model’s list price.

This is a sequel, not a restatement. AI Insiders covered Cognition’s Fusion architecture as a dual-agent cost cutter that trimmed per-task spending by 35 percent, pairing a frontier lead with a cheaper sidekick that executes delegated work. That earlier story was about the architecture. This one is about which model belongs in the lead seat once the sidekick is fixed, and the answer inverts the obvious pricing logic.

The efficiency gap is visible in raw usage. Fable’s lead averaged 11.5 turns and 545,000 cumulative input tokens per run. Opus needed 26.5 turns and 1,679,000 tokens, more than three times the context. Fable’s lead also produced far fewer output tokens per run, 6,100 against Opus’s 19,000, near a third the volume.

The starkest gap is behavioral, not computational. Fable’s lead skipped hands-on coding entirely in 81 percent of its runs, delegating the full implementation to the sidekick model. Opus did the same in only 24 percent of runs, meaning its lead touched code directly four times as often as Fable’s did.

The two models manage differently, not just efficiently. Fable delegates early, writing specification-quality briefs that define constraints and outcomes upfront. Opus explores extensively before delegating, then frequently re-reads summarized files and corrects the sidekick’s output itself, at Opus’s premium token rate. Both models delegate roughly three times per run; the difference is when, and how well the handoff is specified.

One caveat belongs here. FrontierCode 1.1 is Cognition’s own evaluation, run on Cognition’s own product, and the post includes no results from an independent benchmark or a neutral third party repeating the comparison.

For any team pricing agentic work by the token, this data argues for measuring the completed run instead. A model priced twice as high per token can still finish a task cheaper if it delegates cleanly and skips corrective rework, so procurement decisions based on token price alone will misprice the job. Teams running multi-agent pipelines should re-run their own cost-per-task comparison before assuming the model with the lower sticker price is the model with the lower bill.

Cognition detailed these findings in a July 13, 2026 post by engineer Joon Hee Lee on its engineering blog, which the company amplified on X the following day.