Bench
For builders and anyone following AI. We report what shipped, what it costs and who it affects, with the primary source and a second publication behind every story.
27 verdicts · Updated Oct 24, 2026
Type
Sort
27 verdicts shown
The leaked-secret monitoring preview gains alert management and integration tools. GitHub's announcement and documentation conflict on eligibility for some enterprise customers.
By Jason4 min read
News
Each totals_by_* list expands from five entries to 25 per record, while earlier report days retain generic labels and existing counting rules remain unchanged.
By Jason4 min read
News
Moonshot lists Pro cheaper than the earlier K3 and says Highspeed runs 2-3x faster than Pro at twice its price; Vercel offers both through AI Gateway.
By Jason3 min read
News
The benchmark lists $2-per-million input pricing and 77.1 output tokens per second, while sources differ on the model’s proprietary and open-weight status.
By Jason3 min read
News
The demonstration includes 49,401 email-cited facts. Human choices determined which organizations and internal units received pages.
By Jason3 min read
News
The browser demo predicts screen changes from keystrokes. Its creator reports faster updates and improved simulated file recall, but filesystem consistency is unreliable.
By Jason4 min read
News
The TorchTPU engines add a PyTorch serving path alongside existing JAX runtimes. Project documentation limits what the launch establishes about model coverage and deployment.
By Jason4 min read
News
Vercel reports faster tasks and shorter outputs. External measurements cover the Max variant, while Gateway-specific rates and configuration remain unconfirmed.
By Jason4 min read
News
The repository lists 8B and 14B checkpoints. Performance and efficiency figures remain author-reported, without independent corroboration.
By Jason3 min read
News
The agent can group up to 25 alerts into a draft pull request. Enabling it replaces the free alert-level fix action with a workflow that consumes AI credits and Actions minutes.
By Jason4 min read
News
The internal review system's reported average cost fell to $0.77, but the comparison uses its most expensive weeks and does not establish unchanged defect detection.
By Jason4 min read
News
The promotion applies only when MiniMax serves the request. Vercel says requests pinned to MiniMax will move to standard token pricing after the offer ends.
By Jason4 min read
News
Three publishers' guides agree that one benchmark score hides which habit broke, but split on whether tests should assert an agent's steps or only its results.
By Jason5 min read
Analysis
Subscribers get Colab compute units and faster accelerators. Ultra adds background execution and Premium GPU access, and Colab subscriptions are unchanged.
By Jason3 min read
News
The 740M-parameter model shares one 768-dimension space across five modalities, with an 8K context and Matryoshka truncation to 128 dimensions.
By Jason4 min read
News
Jobs on macos-14, macos-14-large and macos-14-xlarge fail on purpose in scheduled windows; three of eight brownouts have passed as of October 19.
By Jason3 min read
News
Clef (27B) and Clef-flash (9B) return typed answers with probabilities. Cloudflare's benchmarks put Clef ahead of Typesafe AI's Jev on most tests; none are independent.
By Jason3 min read
News
Meta's personal AI agent runs on its own cloud computer. Meta lists no prices, and the tier figures come from Zapier.
By Jason4 min read
News
GitHub says each trusted publishing configuration can now move tags such as latest and next with short-lived OIDC credentials, off by default.
By Jason4 min read
News
The Nvidia-backed lab says the text-only model matches GLM-5.2 on reasoning with 3-4x less inference compute; the benchmarks are the company's own.
By Jason4 min read
News
The Oct. 6 change leaves Enterprise organizations opted out until an admin enables it; the agents keep running locally, so the computer must stay on and online.
By Jason4 min read
News
Lovable and Supabase tokens are partner patterns reported to the provider, while the two Pydantic tokens raise alerts in the repository. None has push protection.
By Jason4 min read
News
Sonnet 5.5 input and output prices match Sonnet 5. The Oct. 7 cache-read cut is the only rate change, and Anthropic says it trims most agentic task costs by about 20%.
By Jason4 min read
News
Microsoft's Sept. 30 release also puts the Copilot harness in an agent host process and adds HydraFusion to the model picker as a research preview.
By Jason4 min read
News
A Harvard-Jellyfish working paper and an NBER study both find commit-level gains from coding agents shrink by the time work is reviewed, merged and released.
By Jason5 min read
Analysis
The smallest of four AI library installs was @anthropic-ai/sdk at 13316 KB; langchain was 52100 KB. Check your bundle before deciding.
By Jason6 min read
Verdict: DependsCurated
The agents run on GPT-6 Astra, with one included for ChatGPT Pro and Business Premium subscribers outside the EEA, Switzerland and the UK.
By Jason3 min read
News