Everything that shipped in AI this cycle, tested and boiled down for builders. Tap any to open the source.
An independent reproduction (HN front page Jul 18, press Jul 19) shows GPT-5.6 deriving a result that had stood open for three decades from one prompt. The takeaway for builders: frontier models can now do real research-grade math, so treat them as collaborators on hard analytical problems, not just code and prose.
A 2.8-trillion-param open mixture-of-experts with a 1M-context window, benchmarking near Opus 4.8 while priced like Sonnet 5. Weights land Jul 27. Already routable via Vercel AI Gateway, so A/B it against your current frontier model on cost-per-task before the weights even drop.
Days after Kimi K3, Alibaba's 2.4-trillion-param flagship goes live in preview via Token Plan, Qoder, and QoderWork at 10% of standard pricing, with open weights promised soon. It speaks both OpenAI and Anthropic API protocols, so you can A/B it without restructuring your stack. Active-param count and independent benchmarks are still undisclosed, so verify before trusting the second-only-to-Fable-5 claim.
Those two commands are now opt-in instead of firing automatically at the end of a turn. If you relied on the auto-behavior, wire /verify and /code-review back into your own hooks or run them manually.
Closes holes where Edit globs and long or quoted bash could dodge permission prompts, plus a new EndConversation tool that lets the agent cleanly close a session. Upgrade if you run Claude Code in any semi-autonomous mode.
The Sol, Terra, and Luna tiers were mislabeled; the real window is 272K tokens. Update so your context-budgeting and truncation logic uses the right ceiling.
New managed rules block active remote-code-execution and SQL-injection exploits in popular WordPress frameworks before you can patch upstream. If you front any WordPress on Cloudflare, confirm the managed ruleset is enabled.
Plus SDK 1.1 and a new neon diff for schema comparison. Postgres, blob storage, serverless functions, and model routing now live behind one Neon project. Try it if you want to collapse your agent backend onto a single provider.
A packaged set of Claude Code skills that autonomously hunt for vulnerabilities in a codebase. Drop it into your repo to run agent-driven security passes on in-flight changes.
The agent now proposes a plan before editing, spans multiple repos in one task, and picks work up across channels. Review the plan gate to keep large multi-repo edits under control.
A Jul 19 follow-up moves heavy peer deps to optional, roughly halving the footprint, and Kimi K3 is now selectable in the CLI. Reinstall for a lighter agent and a cheaper frontier option.
Deploy and manage Railway services from inside Grok Build, and push edge caching from the terminal with railway cdn. Handy if you drive infra from an agentic CLI.
Two new open frontier models one gateway call away, plus free outbound traffic from Vercel Sandbox and Slack agents in the Chat SDK. Swap models via config with no new SDK.
Share one identity across a workspace and pull data from Apollo and Wiz into builds. Reuse auth instead of re-connecting per project.
Subagents can now hold a multi-turn conversation, and a preview plugin manager lets you add and remove capabilities. Try it for delegated, longer-horizon agent tasks.
Each session gets its own git worktree so parallel work does not collide, and you can scope network access per session. Good for running several isolated agent tasks at once.
A ~6k-star harness that wraps coding agents with security checks. Part of the wider agentic code-security wave alongside VulnHunter and Strix; worth studying if you run agents against production code.
An open protocol that lets multiple coding agents coordinate on the same task. A reference to watch as multi-agent workflows go mainstream.
Runs a local review pass on agent output, then hands the findings back to the agent to fix. A concrete answer to the generation-versus-verification bottleneck.
A local MCP server that coordinates several agents editing the same files without clobbering each other. Useful scaffolding for parallel agent workflows.
Toggle feature flags from CLI, CI, or an agent without a redeploy. Wire it into automation to flip behavior without a full deploy cycle.
Point it at a URL and get structured, agent-ready JSON, exposed through an MCP server (~625 upvotes on Product Hunt). A quick way to feed live web data into agents.
See the rendered body of routed messages straight from the Activity log. Faster debugging for email-routing flows.
Session-wide WebSearch caps, per-session subagent spawn limits, auto-backgrounding of long MCP calls, /fork into background sessions, and a screen-reader mode. 2.1.212 fixed a plan-mode bug that ran file-modifying Bash without permission. Upgrade to 2.1.213 and use the new caps to bound runaway agent runs.
Frontier (Sol), balanced (Terra), and high-volume (Luna) tiers, plus Programmatic Tool Calling, explicit prompt-cache controls, persisted reasoning, and beta multi-agent orchestration in the Responses API. Route by tier and adopt Programmatic Tool Calling to cut agent-loop cost.
xAI's strongest coding/agentic model at $2/1M in and $6/1M out, live in Cursor on all plans day one. Benchmark it for agentic coding before defaulting to it.
A full agentic coding harness you can fork and study. But a wire-level analysis found the CLI uploading entire tracked git repos to a GCS bucket, forcing a public deletion promise. Great reference harness, and a cautionary tale on agent data exfiltration.
An open-weight ~1T model that natively ingests text, image, and audio, with day-0 support in transformers, SGLang, and llama.cpp, plus routing on Vercel AI Gateway. An NVFP4 build (~600GB VRAM) or llama.cpp quants make it self-hostable; try HF Inference Providers first.
Agents run async server-side (background: true) and call remote MCP servers. Offload long-horizon Gemini work to the background and reach private data via remote MCP instead of local shims.
Inline scan of in-flight changes for injection, XSS, path traversal, and weak crypto, with severity and confidence scores. Available on Free and Pro too. Run it on a diff before you open the PR.
Better interruption handling and alphanumeric recognition for production voice agents. Swap to gpt-realtime-2.1 (or the cheaper mini) for voice input, overlapping speech, or spelling of codes and emails.
Blocks more destructive shell forms (expanded rm variants) before execution in autonomous modes. If you run Codex at higher autonomy, upgrade, though these builds are pre-release.
Point at any OpenAI-compatible endpoint with your own key, customize Claude agent providers, and run local sandboxing (preview). Bring a self-hosted model into JetBrains Copilot and keep it local.
Startup-time fingerprinting of MCP server config against a trusted baseline (a supply-chain guardrail), plus GA of automated MSVC upgrades. Verify MCP servers haven't been tampered with.
Public and private DNS unified with shared policy. Consolidate split-horizon DNS for agent and backend flows behind one set of internal resolvers.
Agents handle mid-flow form and URL user input over MCP. Build agents that pause to collect structured input and then resume.
The newer Kimi frontier model is one gateway call away, with no separate Moonshot integration. A/B it via gateway config, no new SDK.
Inject a message that interrupts an in-progress turn, plus full LLM, TTS, and end-to-end latency breakdowns emitted automatically. Build barge-in voice UX and diagnose latency without extra instrumentation.
The Agent creates and edits diagrams without credential setup, and non-owner editors can answer the Agent's routine questions. Delegate Agent Q&A to teammates and wire domains in-app.
The CLI hits any Platform API route, and MCP branch create accepts expiresAt. Spin up auto-cleanup DB branches from agents or CI.
Deploy and manage infra from chat, plus feature flags and CLI spend caps. Trigger infra ops from chat and cap agent or CLI spend.
Open-weights Arabic ASR that beat Whisper in 96% of human comparisons. Evaluate it as a Whisper replacement for Arabic audio.
One log pane across Edge Functions, DB, and Realtime. Trace cross-service issues in a single view.
A 1M-context multimodal agentic model on a new OpenAI-compatible Meta Model API (US preview), reported near $1.25/$4.25 per 1M with $20 free credits. Point your OpenAI SDK base URL at it to A/B test.
Run multiple agent sessions side by side with full-session cost visibility, plus Marketplace model-provider discovery and a unified context and reasoning picker.
grok-voice-think-fast-1.0 realtime plus grok-tts and grok-stt are callable through AI Gateway on AI SDK 7, with no separate xAI key and built-in spend controls.
The /side command opens a durable side chat for tangents without polluting the main thread, Cmd+K searches agent transcripts, and new beforeSubmitPrompt, afterAgentResponse, and afterAgentThought hooks let teams instrument agents.
Broadcast now sends and receives ArrayBuffer across clients, REST, and the DB (auto in supabase-js 2.91.0 and up), cutting encoding overhead for telemetry and live-frame streaming. Watch for silent drops to older clients.
A stronger reasoning LLM inside Voice Agent API with no separate Anthropic key or orchestration. The claude-sonnet-4-20250514 snapshot is deprecated and now errors, so audit any pinned configs.
Native agy CLI agent support plus user-defined model routers turn the terminal into a multi-model routing hub. Route cheap tasks to small models and hard ones to frontier models automatically.
Point Zed's agent at a local llama.cpp model for fully offline, private agentic coding. Note that format-on-save is now off by default.
Remote plugins are on by default for a bigger catalog, and new macOS and Windows system-proxy support finally lets corporate users behind proxies run the Codex CLI cleanly.
A safer middle ground that approves edits but not arbitrary commands, plus MCP auth without opt-in. Useful for teams that want agents to edit files but not execute.
Design Guidance generates three design directions before building to fight generic AI aesthetics, and a new GitHub API connector wires repos, issues, and PRs into issue-aware builds.
Flux conversational STT now returns per-word start and end times for precise captions, highlighting, and barge-in timing. Bump the Python SDK to 5.3.4 or later if on 5.x.
Orgs can mandate OTel telemetry destinations and push managed Copilot settings (permissions, model, plugins) via Intune, Jamf, Group Policy, or Ansible for org-wide control.
toMarkdown and AI Search now ingest .gif and .bmp via object detection into a vision model, indexed automatically on next sync with no config. Stop pre-filtering these formats.
Agent Customization hooks, in-IDE MCP server management, and /create-skill or prompt or agent generators turn JetBrains into a configurable multi-agent IDE rather than a single-agent chat.
Devin Local shows autonomous edits as reviewable diffs with in-client ACU usage, making autonomous edits auditable and cost-visible. Review diffs before merge and watch the ACU meter.
A 560B-param open-weight reasoning model optimized for speed charted on Product Hunt. One to evaluate for a self-hosted reasoning stack, but verify weights and license before production.
All three tiers are live in the API, ChatGPT, and Codex; Terra is roughly GPT-5.5-class at about half the cost ($2.50/$15 per 1M).
A frontier coding/agent model at $2/$6 per 1M, available day one in Cursor on every plan and in the xAI console. Not in the EU yet.
Blends Codex and ChatGPT to act across apps and files and stay on a single task for hours at a time.
New listen-and-speak-at-once models power ChatGPT Voice; API access is "coming soon," so it is ChatGPT-only for now.
Selectable today in desktop, web, iOS, CLI, and SDK, for multistep tasks at $2/$6 per 1M.
The macOS and Windows client gets inline code editing, annotations, and GitHub PR support.
One line of Wrangler config puts a tiered regional cache in front of any Worker; cache hits bill no CPU.
A sub-second edge VLM (query, caption, point, detect) callable via env.AI right at the edge.
Joint transcription plus diarization (30.17% cpWER vs Nova-3's 37.92%) across 18 languages.
Vibe-coded Lovable projects get git-synced production hosting on Vercel.
Ask the Agent to "add payments" and it creates and connects Whop; plus a new native desktop app.
World-model policies, a reward-model API, six sim benchmarks, and FSDP cloud training.
Encrypt PII, health, and financial columns via CipherStash and still query them.
"Give me a high-level overview" produces purpose, stack, and a contributing guide (or a README) for fast codebase onboarding.
A 6x speedup (2.75s to 0.44s) via FP4 quant plus kernel fusion; new Base/Fast/Instant tiers for image-heavy apps.
One-click OAuth app setup and governed MCP access across servers in Claude, without per-server login.
Design-to-app builds get a stronger model for more accurate first passes.
A single RGB camera plus a plain-language instruction moves a robot through a space. Access is gated ("talk to our team").
Versioning, ownership, lineage, and traceability, a system of record for teams shipping controlled deployments.
The CLI routes model choice by task and lets you set per-session AI-credit limits for safer, cheaper scripted agent runs.
Stay on top of what shipped in AI. Weekly update.