Model launches are announced with coding scores. Almost none of them are announced with agentic coding scores, and after working through LiveBench’s 2026-06-25 release, the reason looks uncomfortably simple: the second number is much worse.
LiveBench splits the two. Its Coding category asks models to generate and complete code — the familiar shape of a benchmark question. Its Agentic Coding category, added in May 2025, runs a harness against issues from real GitHub repositories in JavaScript, TypeScript and Python, and checks whether the issue is actually resolved.
Across the 28 models in the current release, the best Coding score is 83.6. The best Agentic Coding score is 56.1. The average model gives up 30.9 points when the task moves from a prompt into a repository.
The scores are not the interesting part. The prices are.
LiveBench also publishes what each evaluation cost, and declares each model’s token prices in its own repository. Put the two together and the ranking stops looking like a ladder.
GLM 5.2 scores 51.9 on agentic coding — 93% of the leader’s 56.1 — at roughly a fifth of the cost per successful task. grok-build-0.1 lands at 45.8 for about two cents a task, against roughly a dollar at the top of the board.
One honest wrinkle in that comparison: LiveBench’s score table and its cost table label effort levels differently, so the top model’s cost is taken from its nearest published variant, which runs at a higher effort setting. The direction of the finding is safe — the leader is several times dearer per successful task — but read the exact multiple as indicative rather than precise.
Eight of the 28 models sit on the cost/score frontier, meaning nothing on the board is both cheaper and better. That is a crowded frontier for a market usually described as a two-horse race, and it means “which model” is a budget question as much as a capability one.
This is the argument behind model routing generally, and behind Cognition’s Fusion framing specifically: if a large share of agent work is mechanical, paying frontier prices for all of it is a choice rather than a requirement. The LiveBench numbers put third-party figures under that claim instead of leaving it as vendor assertion.
Three caveats that matter more than the ranking
Devin is not on this board. LiveBench scores models; Cognition does not submit Devin to it, and no Cognition entry appears. What the data describes is the behaviour of the underlying models an agent routes to — useful for exactly that, and misleading if read as a ranking of agent products. Cognition evaluates Devin on SWE-bench instead.
It is a monthly snapshot, not a live feed. LiveBench ships new questions on roughly a monthly cadence and replaces about a sixth of them each release, deliberately, to limit contamination. That design is the point — but it also means scores are not directly comparable across releases. A model that moves may be facing new questions rather than having changed.
Agentic scores are harness-dependent. LiveBench’s agentic category runs through a SWE-agent-style harness with a step limit, and the changelog records that both the harness and the limit have changed over time. A score measures a model inside a particular scaffold, which is precisely why a product with a better scaffold can outperform its raw model number.
Figures in this piece are derived from LiveBench's 2026-06-25 release, distributed under the Apache License 2.0. Our full sortable view, method notes, and the cost/score frontier live on the agentic coding scoreboard. Cite the benchmark as White et al., LiveBench: A Challenging, Contamination-Free LLM Benchmark, ICLR 2025.
What to take from it
If you are choosing a model for agent work, the headline coding score is close to useless on its own — the field is bunched at the top there and spread out on repository tasks. Look at the agentic column, then look at what a successful task costs, and notice how often those two rankings disagree.
And treat the 30-point gap as the honest state of the art. Agents are genuinely useful today at wide, well-scoped, verifiable work. The benchmark says what practitioners already report: the further a task gets from a clean question, the more the scaffolding around the model has to earn its keep.
References
- LiveBench leaderboard — livebench.ai
- LiveBench source and changelog — github.com
- LiveBench: A Challenging, Contamination-Free LLM Benchmark — arXiv:2406.19314 (ICLR 2025)
- Declared token prices — github.com
- Agentic coding scoreboard — Devin Central
From Devin Central — subscribe for Devin news, features & tutorials. Devin is a trademark of Cognition; other product names belong to their respective owners.