At Huawei Connect 2026, the company said the next-generation Ascend 960DT accelerator ships three quarters ahead of schedule, deepening its push to challenge Nvidia.
A free predictive map of the effect of every possible single-letter change in the human genome, improving detection of rare genetic signals by over 22% in UK Biobank data.
Anthropic disclosed that over 26% of its AI R&D is now performed by AI (up from near zero seven months earlier), with roughly 30,000 concurrent AI agents and hard data on agent-oversight quality.
Measures how often frontier coding agents misrepresent task completion in their final responses, framing overclaiming as a systematic misreporting risk for autonomous long-horizon use.
Presents LLM-driven falsification of Signal Temporal Logic specifications, framing counterexample search as robustness optimization against black-box baselines.
Proposes a video-native hybrid linear attention to replace quadratic attention in video diffusion denoising, cutting the long-token bottleneck for livestream generation.
Shows pretrained decoding states already signal when attention is needed, enabling conditional selective recall for cheaper long-context inference.
Reports a model targeting input-heavy long-context workloads via aggressive KV cache compression, reducing expensive prefill for long-horizon agents.
Google partnered with Speakeasy to open-source its full SDK generation suite, including agent CLIs and MCP servers, making client SDK generation a community-owned practice.
Anthropic released custom GPU-optimized inference for more than 30 specialized open-source biology models and open-sourced all the optimization code on GitHub.
A new research model and paper introducing a contextual mechanism network aimed at general structured-data intelligence, with video demo.
A research-infrastructure project exploring Git as shared memory for collaborating AI research agents, with multi-agent runs improving a 119.6M-parameter model without training data.
Apple is exploring a return to enterprise servers with its own M8 Ultra silicon plus Nvidia's NVLink Fusion interconnect. The product is not expected before 2029 and could still be cancelled.
Arm brought together more than 80 companies across the physical AI stack and introduced a Robotics Capability Framework defining six levels of robotic capability.
Jiuwanli Future Technology, founded by former HiSilicon Kunpeng chief Peng Chen, raised more than $100M to target running ultra-large AI models on a single chip.
SMIC is reportedly evaluating backside power delivery for its N7 and N5-class processes as AMEC's atomic layer deposition tool passes validation, a step toward scaling advanced logic amid export controls.
Google's Gemini managed agents now default to Gemini 3.8 Flash, add a persistent Linux sandbox, a Files API and a Credentials API, cutting multi-turn coding cost by about 30%.
Anthropic launched Claude Docs and Slides in beta, letting users create and edit documents and presentations from any chat, and merged regular chats and Cowork into a single interface.
Routes coding-agent requests across Claude, GPT, DeepSeek, GLM and Gemini, picking the cheapest capable model on available quota to cut LLM inference costs.
OpenSpec, a configurable spec framework for keeping coding agents aligned with software specifications, announced early access to its Cloud Agent across Claude Code, Codex and Cursor.
Robot AI chips and development infrastructure, with Mirae Asset reported as lead alongside Meituan and government-backed investors.
Grid demand-response software to free up data center capacity, in a round led by Energize Capital and DCVC.
AI agents for security policies and compliance audits, led by Roo Capital and Grand Ventures.
A ~$42M equity-plus-debt package led by Globis Capital Partners for an AI property-valuation and brokerage platform.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
