Skip to main content
What’s new on chat.ironlabs.ai. Reverse-chronological. Newest first.

​
June 2026

  • Model lineup cleanup. Deprecated 31 prior-generation models from the chat selector: OpenAI gpt-4.1/gpt-5.0/gpt-5.1/gpt-5.2/gpt-5.3-codex, Anthropic claude-sonnet-4 and claude-opus-4.5, Google gemini-2.5-pro/gemini-2.5-flash, xAI grok-4 (plus legacy grok-3/grok-3-mini), Together Qwen3-30B-A3B and Kimi-K2-Thinking, plus older 2024 models. Newer generations stay: Claude Haiku 4.5 / Sonnet 4.5 / Opus 4.6+, Gemini 3.x, Grok 4.3+.
  • Brand: now Foundry. The chat product is officially Foundry by IronLabs. Page title now reads “IronLabs — Foundry Results, not chat.”
  • Fixed: billing display. Cost and throughput numbers display correctly again; closed a $0 leak in chat / cache billing paths.
  • Fixed: stream crash. Prevented unhandled-rejection crashes on no-output streams + a freeModelCount app-shell crash.
  • Fixed: agent-mode UI sync. Agent Mode now correctly syncs UI on fallback, skips doomed primaries, and alerts Sentry on critical failures.

​
May 2026

  • 3D Website Mode wired into the composer. New composer UI exposes 3D Website Mode as a first-class option alongside Chat / Image / Video. Backed by the OpenRouter pipeline introduced earlier as the first Workflow preset.
  • SLM passthrough completions. New /api/v1/slm/completions endpoint with PBKDF2 API-key hashing and prompt support.
  • Vulnerability sweep. Updated all vulnerable dependencies to resolve every critical and high CVE.

​
April 2026

  • Promo banner added.
  • UX polish. Scrollbar at the rightmost edge, full-width header divider; image-mode edit-user-message fix; search + attach allowed for signed-in users.
  • Pricing tweak. Web Search now costs 2× and Deep Research costs 3× a normal chat reply.

​
December 2025 → January 2026

  • Docs restructure. docs.ironlabs.ai reorganized into two tabs — Refinery (developer platform) and Foundry (chat product). 35 new Foundry pages: Get Started, Chat, Modes, Reasoning, Sources, Connectors, Skills, Tasks, Workflows, Personalization, Tutorials.
  • Brand consolidation. iron.cx shortlink kept as the chat-app short URL; forge.ironlabs.ai legacy domain folded into foundry.ironlabs.ai.

​
November 2025

  • Image Mode model refresh. Added NanoBanana Pro, GPT Image 1.5, and Qwen-Image-2512 to Image Mode. NanoBanana (the original) also landed earlier in the month.
  • Reasoning Effort Control. Pick Low / Medium / High in Settings to trade off cost and depth. Effort control →
  • Live Reasoning View. Watch the model think in real time before it answers. Live view →
  • Memory Support. Foundry now remembers what you told it last week — your projects, your tone, your favorite formats. (Roadmap: full dynamic memory; today: tuned via System Prompt.)
  • Zapier integration. Drop Foundry into any Zap and run prompts inside your existing automations. Zapier →
  • MCP Support. Foundry can interop with Model Context Protocol tools and contexts.
  • Voice input. Click the mic in the composer and speak your prompt. 99+ languages auto-detected. Voice →
  • Dark Mode. Everywhere — chat, dashboard, settings.
  • New keyboard shortcuts. ⌘N new chat, ⌘⇧A add attachment, ⌘K search.
  • Browser notifications when a long response finishes — switch tabs without missing the result. Notifications →
  • Custom Router (Foundry side). Build your own router with your chosen models, logic, and preferences. Self-serve, no human-in-the-loop.
  • Capabilities-based Model Selector filter. Find the right model by what it does, not its name.
  • Feature Drop notifications. In-app toasts when something new ships.
  • Latest models added to routing: Qwen 3, Grok 4-Fast, Gemini 3 Pro / Flash, Claude Haiku 4.5, Claude 4.5 Sonnet & Opus, Kimi K2-Thinking.
  • Fixed: multiple file upload bugs. More reliable attachments.

​
October 2025

  • Image Generation Mode. Generate high-quality images right in chat. Prototyping, design ideation, creative tasks. Image Mode →
  • New reasoning models in chat + SDK. GPT-5 and newer Grok with better reasoning capabilities.
  • Fixed: Ironlabs Studio UI bugs. Cleaner profile section. More reliable subscription validation and chat-subscription flows.

​
September 2025

  • Reasoning tokens now first-class. Live reasoning + reasoning summaries. Billing accounts for reasoning tokens accurately.
  • Regenerate. Don’t like a reply? Regenerate from the same or a different model — even mid-conversation.
  • No more jumpy scrolls. Long replies no longer auto-scroll. A Scroll to Bottom button appears so you can read at your own pace.
  • Keyboard shortcuts. A full set, designed to streamline navigation.
  • Cold-start latency. Cut significantly via a Keep-Alive server (Vercel).
  • Dodo payments for both custom top-ups and auto top-ups.
  • Model Selector modal redesign — more intuitive, less clutter.
  • Fixed: various UI inconsistencies.

​
August 2025

  • Anonymous / Guest users. Try Foundry without signing up. Great for first-time exploration.
  • Ironlabs Studio. Powerful playground for developers, researchers, AI tinkerers — customize prompts, test responses, view token cost + latency in real time.
  • New models added: GPT-5 series, GLM-4.5 series, Kimi K2.
  • Throughput tracking. Better visibility into system performance and request handling.
  • Mobile responsiveness pass. ChatsBar, Chat, Dashboard, ModelSelector — all tuned for small screens. Mobile overlay backdrop, refined sidebar toggle, mobile header with menu dropdown.
  • Rebranded domains. ironlabs.ai is the landing page. iron.cx is the chat-app shortcut. Both live and healthy.
  • Background jobs. New cron-job infrastructure for more reliable automation.
  • Fixed: anonymous feedback flow, emoji rendering inconsistencies.
  • Fixed: general latency and bug-fix pass. Faster everywhere.

​
July 2025 — version 1.0

The first release. The start of an intelligent platform.
  • Unified chat interface for top AI and LLM models — all in one place, no juggling.
  • Automatic LLM routing. Foundry picks the best model per query (performance, cost, or latency).
  • Hyper-personalization via feedback. Thumbs-up / thumbs-down individual responses; the router learns and tailors future routing.
  • Routing optimized on 1M+ training data points. Real signal, not hand-tuned rules.
  • Arcade Mode. Run multiple models side-by-side and pick the response you like best.