The developments most likely to change what AI builders do next.
Models, APIs & Pricing
DeepSeek launches DeepSeek-V4-Pro GA with 1M context and speculative decoding
Summary The 1.7T parameter MoE model brings three adjustable reasoning levels, native Responses API support, and improved agentic coding benchmarks.
The details
Features a 1.7-trillion parameter MoE architecture, a 1,048,576 token context window, 384,000 maximum output tokens, and integrated DSpark speculative decoding.
Achieves 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, 83.3 on Cybergym, and 60.0 on Humanity's Last Exam with tools.
Adds three configurable API reasoning effort levels (low, high, max), native OpenAI Responses API compatibility, and direct Codex integration.
Introduces peak ($1.32 input / $3.96 output per million tokens) and 50 percent discounted off-peak pricing starting August 16, 2026.
SpaceXAI releases Grok 4.6 with expanded agentic RL capabilities
Summary Grok 4.6 improves on Grok 4.5 in long-horizon coding and agent benchmarks while maintaining identical base API pricing.
The details
Grok 4.6 matches Grok 4.5 pricing at $2 per million input and $6 per million output tokens with a 500,000-token context window, offering an optional fast tier at double the token price.
On benchmark evaluations, Grok 4.6 High scored 65.9% on DeepSWE v1.1 (vs 54% for 4.5), 69.9% on CursorBench v3.2 (vs 66.7%), and 57.5% on APEX-Agents (vs 47.1%).
Training combined supplemental pre-training, filtered SFT trajectories generated by Grok 4.5, and agentic RL across kernel optimization, CAD, and web development environments.
LangChain Launches Managed Deep Agents Beta on LangSmith
Summary Deploy directory-based, code-first agents with durable execution and memory via one CLI command.
The details
Managed Deep Agents is in public beta on LangSmith Cloud in the US region via the managed-deepagents CLI.
Local code-first project directories contain agent logic, system prompts, custom tools, memory mounts, skills, sandboxes, and evals.
Every durable thread receives an isolated default sandbox for secure code execution and filesystem operations.
LangSmith Context Hub provides persistent memory across agent threads.
Built-in channel integrations allow agents to process external events, including Slack app mentions and direct messages, without custom backend servers.
LlamaIndex releases ExtractBench to test enterprise document extraction
Summary The benchmark reveals frontier VLMs silently truncate long lists, while LlamaExtract Agentic Plus preserves completeness.
The details
ExtractBench evaluates 14 systems across 370 enterprise documents and 4,869 pages, scoring value F1, long-record completeness, grounding, and cost.
Commercial VLMs on 50+ page documents scored 8.9% to 35.8% F1 due to silent list truncation where spot checks pass high precision.
LlamaExtract Agentic Plus leads overall with 95.6% F1 (96.1% on long-list tasks) at 8.1¢ per page, outperforming Codex (93.6% F1 / 27.8¢) and Claude Code (87.1% F1 / 16.2¢).
SQLite Patches 16-Year-Old WAL-Reset Data Race Bug
Summary Tailscale and SQLite maintainers resolved a rare concurrency race condition that silently dropped committed database pages during checkpointing.
The details
The flaw affected every SQLite version from 3.7.0 in July 2010 through 3.51.2 in January 2026.
The issue manifests as a tight-timing race condition during WAL-reset checkpoints, causing committed frames to be silently overwritten under concurrent writes.
SQLite patched the flaw in version 3.51.3 and issued backports for versions 3.44.6 and 3.50.7.
Tailscale logged 19 production corruption events over six months across its database shards before isolating the bug.