November 21, 2026BenchmarkCodingResearch

Models Engineering Index: Neon Wins, the Cost Column Decides

A week of launch-day charts finally got an outside referee, sort of. The Models Engineering Index, a one-person project that hit the Hacker News front page late Thursday with 360 points, ranks Google's Gemini 4 Neon first on engineering tasks with a score of 68.4. GPT-6.2 Astra is second at 67.1, Claude Fable 5.2 third at 66.7. Site: https://git-aarvin.github.io/mei/

Know what it is before you quote it. MEI does not run anything. It is a composite of 16 coding benchmarks other people published, adjusted for benchmark difficulty and missing results so the average model scores 50. That makes it a useful way to see the field at once and a poor way to settle a 1.3 point gap. The top three are within noise of each other.

The interesting columns are the other two. Cost and token use are normalized the same way, relative to a typical model. Neon and GPT-6.2 Astra both sit at 2.1x cost, but Astra uses 0.37x the tokens to Neon's 0.8x. Fable 5.2 buys its third place at 5.8x cost and 1.4x tokens. GPT-6.2 Sol scores 60.6 at 0.54x cost, close to Opus 5.5's 61.5 at 4.2x. Qwen4-Max, launched last week on a cost-per-task pitch, lands ninth at 51.9 and 9.5x cost on this index. DeepSeek V4.1 Flash scores 45.9 at 0.15x. Read down the cost column and the ranking reshuffles completely.

That is exactly the argument Fireworks co-founder Chenyu Zhao made onstage at Slush on Thursday, in a TechCrunch interview: open models are "at most six months behind, probably closer to three", and once a product passes 10,000 to 100,000 users or $10,000 to $100,000 a month in tokens, it is time to move off frontier models. Zhao predicts "a two-model world", frontier for planning and cheap models for everything else, with routing and cost per task replacing cost per token. Worth noting the same interview covers Fireworks' two security incidents this fall. Interview: https://techcrunch.com/2026/11/19/anthropic-rival-says-companies-can-mostly-skip-the-pricey-models/

My read: the leaderboard top is a three-way tie, and the real decision is which model you put in which slot of your agent. A planner can be expensive. The worker that runs a thousand tool calls a day should be whatever sits highest on score per dollar, and MEI's cost column is the quickest place to start looking.
← Previous
Two Million Canaries: Your Cheap Inference Route Is Reading Your Prompts
Next β†’
Anthropic Closes $93.5B IPO After Full Greenshoe, $119.5B With the Chip Placement
← Back to all articles

Comments

Loading...
>_