OpenAI Releases GPT-6.2 Astra and Sol with Better Coding and Lower Cost
OpenAI Releases GPT-6.2 Astra and Sol with Better Coding and Lower Cost
OpenAI introduced GPT-6.2 Astra and Sol, claiming better professional outputs, coding, computer use, and efficiency at lower cost across ChatGPT Work, Codex, and the API. No specific benchmark scores or pricing figures were released. For developers, the new models can be used directly for code generation and automation tasks, and the lower cost means more requests per budget.
OpenAI Releases GPT-6.2 Astra and GPT-6.2 Sol
OpenAI introduced two GPT-6.2 models: Astra, a frontier model for the hardest projects, and Sol, an efficient, versatile model for everyday work. Both are available in the API as gpt-6.2-astra and gpt-6.2-sol. Simon Willison shipped LLM 0.38 with support for the new models. Developers can switch between the flagship and lightweight tiers based on task difficulty to balance cost and latency.
GitHub and Microsoft Turn Thousands of AI Bug Findings into 99 Fix-Ready Reports
In the Codename MDASH pilot, GitHub and Microsoft turned thousands of raw AI findings into 99 checked, private reports with suggested fixes for six widely used open source projects, and maintainers have already fixed more than half. The effort addresses maintainers being flooded with low-quality AI bug reports by shifting from reporting bugs to supplying fixes, reducing triage burden.
GitHub Open-Sources Agent Host Protocol for Cross-Device Sessions
GitHub released the open Agent Host Protocol, aimed at keeping agent sessions accessible across editors, apps, phones and the cloud. According to the GitHub Blog, the protocol lets the same agent session continue across devices instead of restarting on each terminal. For users working across multiple development tools, it means session state can migrate between devices.
Perplexity Hits $1B Revenue Run Rate, Up 4x in 2026
Per Latent Space’s newsletter, Perplexity has reached a $1B revenue run rate, up 4x in 2026. The same issue notes Computer launching Teammates, Perplexity shipping a Monitor API, and a new research lab launching. The growth puts Perplexity among the fastest-scaling AI search products, with a $1B run rate signaling that its commercialization has reached scale.
Video-as-Experience: Audio-Visual Understanding Improves Computer-Use Agents
A new arXiv paper introduces Video-as-Experience, teaching multimodal LLMs computer use from audio-visual demonstrations. Existing video-based learning approaches typically rely on visuals alone and discard audio cues. This work trains computer-use agents that perceive GUIs, reason over instructions and execute actions using combined audio-visual understanding. It points to a new route for training operation agents from everyday screen recordings with narration.
Datasette Lite Rewritten to Run Full Server App in Browser via Service Workers
Simon Willison released a complete rewrite of Datasette Lite, using service workers instead of the previous approach to run the Datasette Python app in the browser via WebAssembly and Pyodide. The rewrite is now almost indistinguishable from the real server-side application, letting users do in-browser what previously required a server. It means deploying a data query app no longer needs a backend.
Simon Willison Explores External Sandboxes for Coding Agents
Simon Willison detailed his exploration of external sandboxes for code executed by coding agents, focused on making it safer to try local models without risking damaging actions from mistakes or prompt injection attacks. He worked on a prototype that isolates agent-executed code in a controlled environment. For developers running agents locally, such sandboxing lowers the risk of trial and error but adds environment setup overhead.
OpenAI Launches ChatGPT for Sales with 30+ Partner Plugins
OpenAI launched ChatGPT for Sales, featuring a new Sales plugin plus 30+ new and expanded partner plugins. The plugins help research prospects, prepare pitches, update the CRM and move deals forward. Sales teams can consolidate prospect research, materials prep and CRM entry inside ChatGPT, cutting the need to switch between tools.
SpecRelay Cuts Round-Trip Overhead in Edge-Cloud Speculative Decoding
A new arXiv paper introduces SpecRelay, a scheme targeting the round-trip communication bottleneck in edge-cloud speculative decoding. The architecture runs autoregressive drafting on an edge device while a larger cloud model verifies draft tokens in parallel, overlapping the two under a synchronous draft-verify schedule. Prior methods were constrained by communication overhead under low-latency networks; SpecRelay amortizes or avoids the round trip to improve overall efficiency. For developers deploying LLMs on edge devices, this means less waiting between edge and cloud while keeping cloud-level verification quality.
Parameter Merging Restores VLM Language Skills in VLA Models
A new arXiv paper examines a degradation problem in Vision-Language-Action models (VLAs): VLAs are fine-tuned from pretrained Vision-Language Models (VLMs) on robot demonstration data, and that fine-tuning erodes language ability. The work uses parameter merging to restore language capabilities while keeping action prediction intact. Better language grounding helps robots follow instructions while executing actions.
GRAFT Reuses LoRA Across Backbones for Continual Learning
A new paper introduces GRAFT, a method addressing the difficulty of reusing pretrained LoRA adapters across models in continual learning. Existing approaches typically require adapters and models to share the same backbone, limiting reuse. GRAFT enables cross-backbone LoRA reuse, letting language and multimodal models borrow already-trained adapters from other models during continual learning. For teams that frequently update model capabilities, this can reduce the cost of retraining adapters from scratch.
Read, Route, Patch: Robust Test-Time Evolution for LLM Agents
A new paper proposes the Read, Route, Patch framework for test-time evolution of LLM agents, improving long-horizon interactive decision-making. Existing test-time evolution methods face three information bottlenecks: memory retrieval overemphasizes textual similarity, and candidate screening is insufficient, among others. The framework addresses these bottlenecks across reading, routing, and patching stages. For developers building long-horizon task agents, this offers a path to continuously improve agent performance at inference time without retraining.
Cloudflare Brands Itself a Neolab, Builds System One Models for Millisecond Decisions
Cloudflare, known for speeding up the web, is now building System One models aimed at making decisions and pulling data in milliseconds. The company has labeled itself a “Neolab” and joined the so-called “Jevolution” trend. The move extends Cloudflare from infrastructure into the model layer, where its edge network could offer a latency advantage for inference, though specific model capabilities and release timing were not disclosed.
Baidu and Xiaomi CEOs Warn of Growing Gap Between AI Hype and Returns
According to SCMP, the CEOs of Baidu and Xiaomi both publicly warned that the gap between AI hype and actual commercial returns is widening. The excerpt does not include their specific remarks or cited figures. Coming from leaders of two major Chinese tech companies, the statements reflect growing industry concern over AI return on investment and offer a management-level view on whether the sector is overheating.
Claude Code v2.1.327 Adds Confirmation Before GitHub App Installs
Claude Code released v2.1.327 with a confirmation step before /install-github-app changes a repository, a new claude mcp add-json command that reads server JSON from stdin so tokens stay off the command line (not available inside a Claude Code session), and a notice when a resumed session’s earlier messages can’t be loaded. The changes target safety and usability, reducing misoperations and credential exposure.
Vercel Blob Adds TTL Auto-Deletion from 1 to 730 Days
Vercel Blob can now delete blobs automatically. Pass ttlDays when writing a blob and it is deleted that many days after upload, from 1 to 730, counted from the upload date. It suits data you would otherwise delete yourself, such as session transcripts, exports and temporary files. Developers can drop custom cleanup jobs and let expired data disappear on schedule.
Vercel Adds Pull-Request-Only Preset for Preview Builds
Vercel rolled out a new Ignored Build Step preset for preview builds that lets teams skip preview builds for branches without an open pull request. Production deployments and previews for branches with open PRs continue to build as usual. On GitHub and GitLab projects, skipped preview deployments are canceled before consuming any build CPU minutes. For teams, this cuts build overhead on idle branches and concentrates build resources on PRs that actually need previews.
Vercel Deprecates Native Lint and Typecheck Deployment Checks
Vercel is deprecating its native lint and typecheck Deployment Checks. These checks ran a repository’s lint and typecheck tasks on Vercel in parallel with deployment builds; new checks can no longer be added and existing ones will stop. Vercel advises using Vercel CI or your own CI workflows to keep running these tasks. Teams relying on the feature need to migrate checks into their CI pipeline.
HIMX Enables Mixed-Format Matrix Multiplication with Custom HIFLOAT4
A new paper presents HIMX, a mixed-format matrix multiplication scheme built on customized HIFLOAT4 representations, targeting 4-bit weight-activation quantization (W4A4) for efficient LLM inference. Recent hierarchical formats like HiFloat4 improve the representational range of 4-bit quantization through multi-level scaling; HIMX builds mixed-format matmul computation on top. For inference deployment, such W4A4 approaches aim to compress models while preserving accuracy, cutting memory and compute costs.