Two launches on November 4, 2026 point the same way: the cheap tier of agent models keeps getting more capable, and vendors are pricing for the bulk of everyday agent work rather than the hardest tasks.
Tencent Hy4: vision and computer use, same price
Hy4 is the full release after Hy4 Preview (August). Tencent says it is the first Hy model with visual understanding, and the first with computer use across command-line and graphical interfaces. It was trained on professional software such as Blender and CAD for long-horizon tasks.
Numbers Tencent published:
- Price: about $0.90 / $2.70 per million input/output tokens and $0.045 per million cached input tokens, unchanged from the Preview. TokenHub has it at 50% off through December 31, 2026.
- Context: OpenRouter lists 1,048,576 tokens of context and 65,536 max output tokens, with tool calling, structured outputs and reasoning.
- Hallucination: in Tencent’s internal tests the hallucination rate fell from 12.95% to 7.2%, and errors from unsupported assumptions fell from 23.1% to 5.6%.
- Game dev: +19.1% on GameDev and +9.8% on GameLogic versus the Preview, with Unreal, Unity and Godot integration.
Tencent’s earlier Preview listing describes a 770B-total / 49B-active MoE. Tencent has not restated the parameter count for the final Hy4, and the weights are not out yet; the company says weights and system cards will follow “in the coming days.” Most of the benchmark numbers are internal, so treat them as Tencent’s claims until independent evals land.
On safety, Tencent says offensive exploitation material was excluded from training, and Xuanwu Lab and Keen Lab found “no significant concerns” about offensive cyber capability in the areas they tested.
Cognition SWE-2 Flash: a small model for routine agent work
Cognition’s pitch is direct: smaller models now handle a large share of what coding agents do, such as automations, code review, feature work and docs. SWE-2 Flash is built for that share.
- Base: Thinking Machines’ Inkling-Small, 276B total / 12B active.
- Recipe: for the first time, Cognition added a mid-training stage of a few trillion tokens (scientific articles, blogs, code) before large-scale RL.
- Gains over the base model: Terminal-Bench 4.0 0.9% → 20.0%, FrontierCode 2.0 3.2% → 24.0%, DeepSWE 1.1 17.4% → 60.1%, SWE-Bench Pro v2 Hard 56.5% → 89.9%.
- Versus its big sibling: on FrontierCode 2.0 it scores 24.0 against SWE-2’s 23.8, but SWE-2 still leads on Terminal-Bench 4.0 (27.3) and DeepSWE (73.0).
- Speed and price: about 200 tok/s versus SWE-2’s 60, at one-fifth the price.
Cognition also pushed its ExploitBench score from 11.3% to 34.8% to help defenders. Because the same skills help attackers, the weights are not public. Cognition says a vetted sharing program for partners and labs is coming “in a few weeks.”
What it means for your routing
Both models fit the planner/executor split many teams already use: a frontier model plans and reviews, and a cheaper, faster model does the high-volume edits, searches and test runs. Hy4 is the one you can call today through OpenRouter at a known price. SWE-2 Flash is only reachable inside Devin for now. If you route executor work by cost, Hy4’s $0.045 cached input price is the number to test against your current executor.