Gemini 4 Neon is a point behind Claude Opus 5.5 in Devin at a third of the cost
Cognition added Gemini 4 Neon to Devin Desktop and Devin CLI, reporting 46.2% on its FrontierCode 2.0 benchmark at $4.38 per task, against $12.82 for Claude Opus 5.5.

Cognition has added Gemini 4 Neon to Devin Desktop and Devin CLI, and says the model scores 46.2% on its FrontierCode 2.0 benchmark, the best result from a Google model. Claude Opus 5.5 is about a point ahead, at roughly three times the cost per task.
The announcement, published on November 18, is a short post built around one chart. It gives the Neon score, the Opus 5.5 score and the cost of each, and little else.

A one-point gap that costs $8.44 less per task
According to Cognition, Neon is live in both Devin products. FrontierCode 2.0 is Cognition’s own benchmark, which judges models on real-world engineering tasks by quality and mergeability. The post puts Neon at 46.2% and Claude Opus 5.5 at 47.3%.
The price difference is larger than the score difference. Cognition lists $4.38 per task for Neon and $12.82 for Opus 5.5, which makes Opus about 2.9 times as expensive. The post defines cost as the mean spend per rollout at list API pricing, public discounts included. It also says solutions that fail blocking criteria score 0, so the score is a weighted aggregate of rubric items and not a simple pass rate.
Access works through the two products named in the post. Devin Desktop can be downloaded from Devin’s site, and Devin CLI installs from its documentation. The post says nothing about which Devin plans include the model, whether it costs extra, or whether it is the default.
Neon against its siblings Carbon and Hydrogen
Cognition has now posted three Gemini 4 models in nine days, and its earlier posts allow a comparison inside the family. The Carbon post of November 10 gave a FrontierCode 2.0 score of 40.0%, up from 26.3% for Gemini 4 Argon. The Hydrogen post of November 11 gave 38.2% at $2.40 per task, and put Carbon’s cost at $4.04.
Set side by side, Neon gains 6.2 points over Carbon for an extra $0.34 per task, about 8% more. Hydrogen is roughly 45% cheaper than Neon per task and eight points lower. Cognition also reported GPT-6.1 Sol at 39.0% in those earlier posts, so Neon is about seven points above it.
Neon: 46.2% at $4.38 per task. Claude Opus 5.5: 47.3% at $12.82.
These figures come from separate posts, and the cost settings behind each model’s number are not spelled out. They are a rough guide, not a controlled test. The step from Argon’s 26.3% to Neon’s 46.2% is nonetheless large for one model generation.
Token prices tell a different story than task costs
Cognition’s cost figure measures a whole task, and that differs from the price per token. In our report on the Gemini 4 Neon launch, the list price was $3 per million input tokens and $15 per million output tokens. Our pricing tracker lists Claude Opus 5.5 at $4 and $20, so Neon is only about a quarter cheaper per token. The task gap in Cognition’s chart is far wider, which suggests Neon needs fewer tokens or less effort per task. The post does not say which, and it does not say how many tokens either model used.
The Devin numbers also sit awkwardly beside Google’s own claim. Google said a Neon task costs 32% less than on Carbon at the default thinking level. In Cognition’s data, Neon at $4.38 costs more per task than Carbon at $4.04. The two companies use different benchmarks and possibly different reasoning settings, so they need not conflict. Readers choosing a model for Devin should treat the benchmark and its settings as part of the claim.
What the post does not establish
The benchmark belongs to Cognition, which also sells the product that runs it. The company does not claim independent verification, and it does not list the number of tasks, the reasoning setting used for Neon, or the variance between runs. The linked leaderboard may carry more detail, but the post itself does not.
The result still matters for people who run coding agents at volume. Per-task spend is the figure that scales with usage, and a model within a point of the leader at about a third of the cost changes that trade-off if the numbers hold up on real codebases. Cognition’s other recent launch, Devin Bots, which runs teams of Devins on a goal, points the same way: more agent work per dollar becomes more valuable as agents run unattended.
Cognition does not say whether it will adjust Devin’s default model or pricing in response, and it gives no timeline for more Gemini 4 models. For now, the evidence is one chart from the vendor that built the benchmark, and users who want to test the claim can pick Neon in Devin Desktop or Devin CLI and compare it with Opus 5.5 on their own work.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



