eesel.ai Review, Deep Read: Smarter, but Verbose, Slow, and Chat Is Not Its Strength

eesel.ai published one of the first systematic reviews within three days of launch, and its headline carried two key judgments: "smarter" and "closed for now." The review was written before the weights release — its "closed" judgment has since expired, but its judgments about the model's abilities still stand. The headline: the jump from 13 to 31 is real, especially in coding agents, vision and document work; but the price is real too — per-task cost and time both multiplied — and chat is not its strength.

Where it improved: coding, vision and documents

The review works through the official benchmark table and finds progress concentrated in three directions: coding-agent work, vision, and document/knowledge work. SWE-Bench Pro v2 (Hard) at 86.3%, PerceptionBench at 64.7% and EnterpriseBench-CoreCraft at 71.8% are the standouts; AA-Briefcase rising from 912 to 1527 Elo suggests document-agent work improved most. The 73.3% AA-Omniscience non-hallucination rate is the highest in the table — the review reads this as the factuality gap narrowing, with the model inventing answers less often.

Against rivals: vision and finance lead, coding close behind DeepSeek

The review's verdict is "mostly vision and finance, with coding close behind DeepSeek." But on the independent AA index, Inkling 2 Small's 30.6 sits below DeepSeek V4.1 Flash at 32.6, Claude Haiku 5.5 at 36.5, and even slightly below Gemini 3.8 Flash at 30.8 — the vendor table and the independent board do not rank it the same way, so it matters who is doing the measuring.

The cost: pricier, slower, more verbose

Artificial Analysis's independent measurements are the most notable part of the review: $1.18 and 15.7 minutes per index task, against $0.20 and 7.0 minutes for Inkling Small. Nearly 6× the cost and twice the time, driven by output volume: about 150K output tokens per task, 4× the predecessor's ~37K. AA's words were "notably slow per task and very verbose." The review also notes that the $0.30 input / $1.00 output unit prices look cheap, but AA considers the input price "somewhat expensive" for its class — the class median is $0.24.

Two reservations: chat, and the then-closed weights

The review quotes the lab's own line: chat is not its strength. The model card further warns of hallucination, imperfect instruction following, weaker long multi-turn chats and uneven quality across languages — the lab's own listed weaknesses, not the reviewer's harshness. The other reservation is timing: written around November 8, when the weights were still unreleased, "closed" was a fair criticism; after the November 12 release it has expired. Read the review with its date in mind: the capability judgments still hold, the access judgments have changed.