Google reports that Gemini 4 Argon scored 77.9% on Deep SWE, ahead of the OpenAI and Anthropic models in its comparison. It is a strong reported result, but it does not settle which model is best at coding.
A Lead on Deep SWE, With Questions About the Comparison
Google tested Argon with mini SWE software, then took the rival scores from leaderboards and provider reports. No one has independently rerun that comparison. The scores in Google’s table therefore come from different sources.
The benchmark has questions of its own. Epoch found 23 flawed Deep SWE tasks, and all 113 tasks and solutions were public when it launched. Those details matter when using a single score to rank coding models.
Google’s own table also puts Argon behind rivals on Frontier SWE and Terminal Bench. Argon leads in Google’s Deep SWE comparison, while the other results point in a different direction. The coding ranking remains unsettled.
Longer Output and Code Changes Have Limits
Google raised Argon’s output limit from 64,000 to 1 million tokens. That gives it room to produce much longer work. The limit alone does not show whether that work stays coherent throughout.
Google says Argon agents helped deploy code changes that freed over 300 teby of memory. It also says Argon is helping rewrite C and C++ libraries in Rust, though those rewrites are still being tested and reviewed. They are not in production.
For now, access is limited to selected Fairwind partners. Paid API customers and AI Ultra subscribers are first in the wider release. Those limits are part of what we can say about Argon today, alongside its reported benchmark score and Google’s account of its code work.
Research Agents Report Gains Across Different Tasks
Darwin 27B RSI’s card reports a GPQA Diamond score rising from 72.85 to 78.09. With 16 votes, the reported scores move from 79.80 to 83.59. The training procedure has not been released, so the loop behind those gains cannot be checked or reproduced from the published material.
An earlier Darwin 27B Opus card reports 86.9%. That test used retries and verification after an initial pass, making it different from the RSI card’s starting score. The two figures are not a clean comparison.
Other reported gains concern narrower jobs. Google DeepMind says AlphaEvolve sped up a Gemini matrix multiplication kernel by 23%, cutting Gemini training time by 1%. Alibaba says Qwen 3.8 Max completed 33 automated improvement cycles in a month while its Artificial Analysis score rose from 40 to 45.
OpenAI says it met its target for an automated research intern working on defined tasks under human direction. Anthropic says Claude authored more than 80% of the code it merges as of May 2026, while people still choose the problems and judge the results. In a Princeton-led research exercise, human reviewers rejected papers produced by Claude Opus 4.8 agents.
These reports cover different kinds of work. They do not give us one score for autonomous research ability, just as Argon’s Deep SWE result does not settle the broader coding ranking.
Chip Memory and Space Radiators Leave Measurements Missing
Alibaba says its Zhen Ru 900 delivers three times the performance of its previous M890. It published 216 GB of memory per chip and 1.2 TB per second between chips. But it has not published compute speed, memory bandwidth or power draw, and there are no independent ML performance results yet.
The memory figure allows a narrower calculation. At FP8, a weights-only estimate puts Qwen 3.8 Max at a minimum of 12 V900S chips, versus 18 Nvidia H200 chips. A running model needs more memory than its weights, and Alibaba has not said what precision it uses for Qwen 3.8 Max. Those counts are minimums, not a deployment plan.
Space compute has a different missing measurement: sustained operation while getting rid of heat. A server on Earth can move heat into air or water; a data center in space needs radiators that emit it. At roughly 27° C, a simplified estimate puts a radiator at about 413 watts per square meter from each side. Sunlight and heat from Earth can make that harder.
Google’s MVP satellite carries four TPUs and has about 1 kilowatt of solar power. It is described as running in bursts of about 15 minutes, with a launch scheduled for October 1. Google has not published its radiator size, and the reason for the bursts is not established. Its research says the heat problem still needs work in orbit.
SpaceX’s proposed AI1 lists 150 kW of peak compute and a 110 m deployable liquid radiator. AI1 has not flown. Its listed hardware and Google’s planned satellite do not yet show how much compute can run continuously in space.
For the headline coding claim, the next useful result is an independent Deep SWE run of Argon’s reported lead.