Tencent's Hy4 joins a benchmark race where cost and proof matter more

Tencent opens its latest model

The release matters less for its features than for comparison shopping. Buyers now weigh Google's index-topping Gemini 4 Carbon against an open-weight model Tencent says will be downloadable. They also weigh Gartner's estimate that open models handle most agentic work at a fraction of frontier cost.

The pitch is broad. Hy4 adds visual understanding and computer-use capabilities for the first time. It can work through command-line and graphical interfaces. Tencent says it trained the model on professional software including Blender and CAD. The aim is long, multi-step tasks. On Tencent's WorkBuddy assistant, the company reports an average performance score of 83.18 and a pass rate of 96.15 percent. Hy4 ranked first in nine of twelve scenarios in mainland China. A blind evaluation run with Tencent Financial Technology used 213 financial professionals. Of those, 153 came from outside institutions. The run covered 579 tasks.

Every one of those figures comes from Tencent's own testing. No outside benchmarker produced them. Open weights change the terms of that comparison. That holds if Tencent publishes what it promised. Buyers can then run the model themselves and check the claims.

Google takes the benchmark crown, narrowly

Tom's Hardware reported that Gemini 4 Carbon scored 59 on Artificial Analysis Intelligence Index V5. That put it one point ahead of Claude Opus 5.5. It was two points ahead of OpenAI's GPT-6 Astra.

Gemini 4 Carbon took the top score on four of the nine evaluations in the suite. It did so at a lower cost per task than other high-ranked models. It also sits nine points clear of Meta's Muse Wonder Full. It is not best at everything. Tom's Hardware noted it trails Claude Code with Claude Opus 5.5 on programming. It also falls behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro on knowledge work.

The margin is thin. Thin is normal. Rankings on that index have moved fast and often through 2026. A single point is not a durable lead. What Google can point to is price. Gemini 4 Carbon cost $7.76 per task on the Intelligence Index test. That is cheaper than Claude Opus 5.5 and Sonnet 5.5. It is roughly half the $12.99 of Claude Fable 5.1. It is still dearer than OpenAI's GPT-6 Astra. It costs about six times as much as GPT 6.1 Sol.

On AA-Omniscience v1.1, it scored 54. That benchmark tracks knowledge recall and hallucination. The score sits on a scale running from -100 to +100. Even the best frontier model of late 2026 is a long way from reliable recall.

The cost argument

Gartner reckons open models can handle 90 percent of agentic work. That is at a twentieth of the cost of a frontier LLM, The Register reported.

Put that next to the price list and the trade-off gets concrete. Gemini 4 Carbon's $7.76 per task is about 13 times what open-weight DeepSeek V4.1 Flash 1017 costs. That model scores 32 on the Intelligence Index. Buyers are choosing how much accuracy they need per unit of work. They are also asking whether a benchmark gap of a few points justifies a multiple on price.

Google's own framing leans on security work. It is giving cyber defenders access first through its Fairwind program. Wider availability comes later. Most buyers cannot test the top-ranked model yet.

What to watch: weights, access and proof

Anthropic published a blog post on Tuesday describing FlashProve. FlashProve is an open-source system for verifying which model produced a given output. It works by having labs publish a fingerprint of each model. Labs can match that fingerprint to outputs and check it cryptographically. The check does not reveal the weights.

The researchers released a technical specification and working code on GitHub. They said they have no plans to deploy it at Anthropic or any other lab. That makes it a proposal, not a product. Still, the problem it addresses is real for buyers. A company could claim an appropriate model approved its business plan. Or it could claim a frontier security model scanned a library. A customer would have no way to confirm it, the researchers write.

For now, the practical questions are simpler. Tencent says Hy4's weights and system cards arrive in the coming days. Google says Gemini 4 Carbon will reach wider availability after the Fairwind program. Until then, the benchmark claims stand largely as their makers published them.