Fifty-two percent. That is the best score on a coding benchmark that does not ask whether the tests are green. It asks whether a real open-source maintainer would hit Merge.
The grade that used to be a person
On October 12, 2026, Artificial Analysis opened a public leaderboard for FrontierCode 1.1, Cognition's private agentic coding eval. The metric is blunt: complete a change in a real repository in a way that would actually get merged by maintainers of those repos.
Top of the board, independently run by AA on the 100-task Main set:
- Claude Opus 5.5 (max, default fallback): 52.2%
- GPT-6.1 Sol (max): 49.1%
- Claude Fable 5.1 (max, default fallback): 49.0%
Twenty-two models. Three attempts per task. A fail on any blocking check, or a trajectory that looks up the solution, scores zero. Full board: artificialanalysis.ai/evaluations/frontiercode.
Read that again. The frontier of "would a human ship this" is still roughly a coin flip - for the models that win press cycles every other week.
Name it: the Maintainer Gate
Cognition's original write-up, from June, already said the quiet part: today's coding benches proved models can write correct code. Correctness is table stakes. The live question is whether they write good code - scope, style, test quality, repo standards, the taste layer that only shows up when someone who owns the tree reviews the PR.
Call that filter the Maintainer Gate.
AA's verifier stack makes the gate concrete: classical tests, command checks, reverse-classical tests, adaptive classical tests, scope checks, and an LLM prompt review. Blocking items are hard stops. Non-blocking items still drag the score. The harness is Stirrup, capped at 250 turns. GPT-6 Luna (medium) sits as the LLM judge. Methodology lives on the same AA page.
This is not SWE-bench with better marketing. Cognition claims an 81% lower false positive rate versus SWE-Bench Pro, with tasks built by 20+ maintainers who spent 40+ hours each defining what "mergeable" means in their tree. The suite stays private to limit contamination.
Two orders of magnitude for the same gate
Score is only half the chart. Cost per task on FrontierCode spans about $0.09 for GPT-6 Luna (max) to $12.00 for Claude Fable 5.1 (max with fallback) - two orders of magnitude for attempts at the same merge standard (AA thread).
Four models sit on the Pareto line for score versus cost, with Luna at 42% and nine cents, MiMo-V2.6-Pro at 45% among the cheap-ish pack. Time-per-task Pareto puts Luna, Astra, Sol, and Opus 5.5 on the frontier - GPT-6 family wins on speed via token behavior, not magic.
So the buyer question is no longer "which model codes." It is "which model clears the Maintainer Gate often enough at a cost your PR volume can survive."
Same day, different storyboards
The same October 12 window shipped louder narratives.
Meta put out Muse Wonder and Muse Wonder Full - first frontier-class Muse models, pitched at personal and professional agents with Active Memory and Active Focus. AA's Intelligence Index put them at 52 and 56 at max effort, up from Muse Spark 1.3 at 48 in six weeks (AA on Muse). Japanese tech accounts amplified claims of beating Fable 5.1 / Astra on agent tasks. Weights stay closed. Pricing chatter lands around $4 / $20 per million tokens depending on tier.
Microsoft unveiled Maia 300, aiming to make Maia the primary inference silicon for its own products by multi-gigawatt rollouts from 2027: 40% more tokens per megawatt than "leading GPU systems" (vendor claim, comparator unnamed), up to 7x rack-scale throughput versus Maia 200, 96 accelerators per rack, HBM4, TSMC N2P chiplets - designed explicitly for long-context reasoning and multi-step agentic work (The Verge, PC Watch).
Google Cloud clarified Gemini Agent is enterprise-first: included at no extra charge for Gemini Enterprise subscribers, multi-model under the agent (Gemini family plus Claude), Japan GA still TBD (ITmedia AI+).
None of those stories are fake. They are just different layers. Silicon and orchestration race ahead. The merge rate on real trees still tops out near half.
What the Gate is really measuring
If you run coding agents at work, you already feel this. Green CI is cheap now. A polite "LGTM" from the person who owns the module is not. FrontierCode formalizes that gap as a leaderboard instead of a Slack argument.
Industrial PR factories still matter - Block's public story of Fable orchestrating design while dozens of Opus/Sonnet workers grind thousand-scale pull requests is the volume side of the same coin. Volume without the Gate is just a larger reject pile.
So here is the uncomfortable cut: agent coding demos optimize for "it ran." FrontierCode optimizes for "a maintainer would keep it." Until those two numbers meet, every "AI wrote our sprint" claim should be read with a 52% asterisk - and a cost axis that can stretch from a dime to twelve dollars for the privilege of still missing half the Gate.