New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Alibaba just launched Qwen3.8-Flash, an open-weight multimodal MoE model with a 125B backbone that activates just 6B parameters per token. It handles 262K native context, previews the upcoming Qwen4 architecture, and cuts API input costs to $0.16 per million tokens while matching heavyweights in coding benchmarks. Weights are live on Hugging Face and ModelScope.
Google just launched Gemini 3.5 Transcribe, a speech-to-text model hitting a 2.6% word error rate for pre-recorded audio and 4.0% for streaming. It strips filler words, handles real-time self-corrections across 85+ languages, and can trigger background Gemini models via function calling. Developers can already access it via Google AI Studio.
Yutori just released Navigator n2, a 27B computer-use model that seamlessly switches between desktop GUIs, CLIs, and code across Windows, macOS, and Linux. Scoring 85.3% on OSWorld-Verified, it matches frontier-class agents at an API cost of just $0.50 per million input tokens.
Sam Altman just revealed that an AI model meeting his bar for AGI will exist internally at OpenAI by year-end. Chief scientist Jakub Pachocki noted that its Astra model can already handle a week of researcher work solo. Why it matters: AI creating genuine new knowledge takes model capability into uncharted territory, even as OpenAI slows frontier rollouts to refocus on safety.
Z.ai just launched GLM-5.3-Flash, a 320B-parameter multimodal model that activates just 18B parameters and costs $0.045 per task—about a tenth of comparable alternatives. It crushes GLM-5.2 on coding and agentic benchmarks to near Claude Opus 4.8 levels, all while running entirely on Chinese AI silicon.
OpenAI just finished pretraining "Bel," a massive foundation model packing over 10 trillion parameters, according to insider reports. The model is slated to serve as the base for Astra and GPT-6 after reinforcement learning, with sources claiming it could reach an AGI threshold—while Anthropic reportedly braces for a quiet stretch amid severe compute constraints.
Thomson Reuters just rolled out Thomson, its first proprietary large language model fine-tuned for legal work. Built on an open-weight base with Westlaw data and RLHF guided by hundreds of attorneys, the $40M project debuts as the default engine for CoCounsel's Tabular Analysis. It proves domain experts can rival frontier labs by training on high-value internal data.
A stealth reasoning model dubbed Ox Alpha just dropped for free on OpenRouter, sending AI circles into a frenzy over who actually built it. Designed for heavy coding and sustained agentic workloads, the model earned high praise from Stripe CEO Patrick Collison. Theories on its origin currently point to Chinese lab Z.ai or Microsoft's unreleased MAI.
OpenAI just dropped API and credit pricing for GPT-5.6 Sol by over 20% for the next three months. The company credited ongoing efficiency improvements alongside model capability upgrades for the cut. Why it matters: developers running production workloads catch a significant cost break through the next quarter.
DeepSeek just rolled out DeepSeek-V4-Flash-Vision-Exp on its API platform. The experimental multimodal model retains V4-Flash's text and reasoning chops while closing in on Opus-4.8 on multimodal agent benchmarks. Images cap out at 384 tokens each, and a new Files API lets developers upload once and reuse across calls.
Anthropic just brought Claude Mythos 5 into Claude Security, giving enterprise customers frontier-level scans to spot vulnerabilities and suggest code patches. The company also launched a $35M Defender Advantage Fund to help secure open-source software and is expanding its Cyber Verification Program. To prevent offensive dual-use risks, defenders get targeted security outputs rather than direct model access.
OpenCode just launched Ox Alpha, a multimodal stealth model that is completely free to use for the next week. It packs a massive 1M token context window, zero data retention, and generous rate limits backed by 100T tokens per day of capacity to stress-test the system.
ElevenLabs just made its Eleven v3 Conversational speech model generally available via ElevenAgents and its API. The real-time model covers 70+ languages and 11,000+ voices, adding audio tags that let developers cue emotions like whispering, laughing, or sarcasm. Starting at $0.05 per 1,000 characters, it brings humanlike nuance to voice agents.
Ornith.ai just launched Ornith-1.5, a family of open models built around a self-improvement loop where the AI proposes its own training tasks, builds scaffolds, and learns via reinforcement learning. The flagship 397B version matches Claude Opus 4.8 on key coding benchmarks, while the compact 9B variant can run on smartphones.
xAI just made its latest flagship model, Grok 4.6, generally available on Amazon Bedrock. Built for long-running agents and complex visual workflows, it features a massive 500K context window and configurable reasoning levels from low to extra high. Developers in supported AWS regions can now integrate it directly into their enterprise pipelines.
OpenAI just previewed Private Safety Processing, a system that detects misuse across multiple API interactions while keeping customer prompts unread and unstored. It pairs automated cross-session pattern detection with customer-controlled encryption keys, so human reviewers never see the underlying content. Why it matters: Enterprise users no longer have to pick between cutting-edge safety guardrails and strict data privacy.
Z.ai just launched its GLM-5.3 API, custom-built for coding, defensive cybersecurity, and long-horizon agentic tasks. Best of all, it keeps the exact same pricing as GLM-5.2, meaning developers can upgrade their automated workflows with zero extra cost. It is available right now via Z.ai's official API and partner model gateways.
OpenAI just paused frontier reinforcement learning runs for its upcoming model Astra after early signs showed it might breach Critical cybersecurity risk thresholds. The lab is enforcing strict sandboxing and adding multistage monitoring that eats an estimated 20% compute overhead. Why it matters: when models get dangerously smart at breaking networks, even their creators have to pull the plug to build thicker walls.
Z.ai just released GLM-5.3, a coding-focused model that achieves open-weights SOTA on Terminal Bench 3.0 and Agents' Last Exam purely through post-training scaling. Unexpectedly, the model more than doubled its cyber exploitation scores, uncovering 2,436 real-world vulnerabilities dating back up to 40 years. Model weights will drop in two weeks after a safety review.
Microsoft just introduced MAI-Cyber-1-Flash, a compact security-focused model integrated into its MDASH multi-agent harness. By routing 90% of routine security tasks to this smaller in-house model and saving GPT-5.4 for the hardest 10%, the system cuts operational costs in half while hitting a 96% score on the CyberGym vulnerability benchmark.
