A no-name model tied GPT-6 Astra, and Claude now leads 26% of Anthropic's AI R&D

The week's frontier story started with a model nobody would put a name on. OpenRouter listed a stealth model called Union Alpha on September 16, 2026, and inside a day users had identified it as Pareto 26.9 from Unbiased. 1

Unbiased says Pareto 26.9 is not one set of weights. Several models work the same task behind a harness that checks the work and calls in a stronger model when a task needs one, and the company says there was no training run. 2 OpenRouter lists it as a multimodal composite model with a 262K context window. 3

On Unbiased's own model card, Pareto 26.9 scores 74 on DeepSWE, level with GPT-6 Astra and DeepSeek 4.1 Flash, while Claude Fable 5.1 sits at 67. The same card shows where it trails Astra: 51 against 58 on Terminal-Bench 4.0, 78 against 87 on MMMU-Pro, 49 against 54 on HLE without tools, and 88 against 91 on ArXivMath. The card adds that measured task costs and a composite score were not published. 4

The price for all of that is $2.50 per million input tokens, $0.25 cached and $7.50 output. Unbiased calls it less than a quarter of Astra's list price. 24

The launch did not hold up under its own demand. The free trial was planned to run a week and ended after one day, when traffic reached a billion tokens a minute and response times became unusable. Unbiased says AWS helped it triple capacity overnight and that it still was not enough, so the free period was cut short and the model went paid. 25

In the same week Anthropic published how much of its own research its model now does. As of August 2026, Anthropic reports that Claude "leads" 26% of its AI research and development — a level where the model completes most of a task end to end from a high-level prompt under human supervision — up from under 1% in February. More than 90% of that work is at or above the level Anthropic calls "AI collaborates". About 30,000 agents were doing research and engineering work on its most-used internal platform at any one time, and about 6% of the compute going to AI R&D that week was allocated to safety. Of more than a billion research-agent decisions analysed over August, the company says its online monitor blocked 0.002%, about 1 in 47,000. 6

Anthropic is asking other developers to publish comparable numbers on a public methodology, so the figures can be compared across labs and over time. 6