Skip to main content

​
October 2026

​
New Course: AI 104, Building with Gloo

The AI Learning Center has a new hands-on course for developers. Nine short lessons take you from your first streamed answer to an app that’s ready for real users, each with a small example you can run.
  • Build the core of an app — streaming, multi-turn conversations, structured JSON output, tool use and images.
  • Work with coding agents — how agents plan, run tools and loop until a task is done, and how to use Gloo Code or Claude Code with Gloo.
  • Point people to resources — recommend books, articles and videos from your content with the Recommendations APIs.
  • Run it well — keep costs down with prompt caching and the right model and reasoning effort, and handle retries, failover and model deprecations.
  • Docs — AI 104: Building with Gloo.

​
New Tutorials, Best Practices and Lessons

New step-by-step tutorials, best-practice guides and lessons for building on Gloo.

​
Guarded Follows Your Task and Format

Guarded Completions and Responses now follow the task and output format in your system message or instructions more reliably, while keeping Gloo’s safety and values.
  • Your format, as asked — JSON-only, single-label and other fixed formats come back as requested, without follow-up questions the format has no room for.
  • Your app’s task holds — text you pass along from transcripts, documents or web pages is much less likely to take over the task your system message sets.
  • Safety still comes first — your instructions can’t switch off safety screening or remove crisis resources.
  • Docs — Guarded Completions best practices.

​
Best Practices for Guarded Completions and Responses

The guarded Completions and Responses pages now show how to get dependable results from the guarded endpoints in your app.
  • Structured output — get JSON for app features such as suggested replies with a forced tool call, with a complete example in cURL, Python and TypeScript.
  • Instructions and content — put your task and output format in the system message or instructions, and pass transcripts, documents and web pages as clearly delimited data.
  • Handle what comes back — blocked requests, moderated replies, retryable errors and parameters a model didn’t use.
  • Docs — Guarded Completions best practices and Guarded Responses best practices.

​
See Every Parameter a Model Didn’t Use

When a model can’t use a parameter you send, your request still runs, and Gloo tells you which parameters weren’t applied, the same way for every model.
  • Every model — on Chat Completions and Responses, ignored_params, or the X-Ignored-Params header on a stream, now lists every parameter a model can’t use, including temperature on Claude 5 models.
  • Your temperature, as sent — temperature stays exactly as you send it, and if you leave it out each model uses its own default; if you relied on the previous default, set temperature: 0.7 in your requests to keep the same behaviour.
  • Effort levels that just work — a reasoning_effort level a model doesn’t list runs at the nearest level below it, or at the model’s lowest level when none is below, and is listed too.
  • Docs — Sampling parameters and Effort levels.

​
Choose Your Image Model in the Image Generation Tool

The image_generation tool on most OpenAI GPT-5 and GPT-6 models now lets you pick from the current GPT Image models, with gpt-image-2 as the default.
  • Current image models — set the tool’s model to gpt-image-2, gpt-image-2.5-flare or gpt-image-2.5-sunburst. gpt-image-1.5, chatgpt-image-latest and gpt-image-1-mini still work and retire on December 1, 2026.
  • gpt-image-1 keeps working — OpenAI retired it on October 23, 2026. Requests that name it run on gpt-image-2, with X-Gloo-Model-Redirected-From, Deprecation and Sunset headers so you can spot them and update.
  • Clear errors — a model that can’t run the tool, or an image model that doesn’t exist, now gets a 400 that says so, and a temporary outage gets a retryable 503.
  • Docs — Image generation tool and Image generation errors.

​
Reasoning Models Get Room to Think

Reasoning models now get room to think before they answer on the chat endpoints, and the Reasoning guide shows how to size your output limit for each task.
  • Room to think — on Completions, Grounded Completions and Direct Completions, a max_tokens below 8,192 is raised to 8,192 when the model is reasoning, so a small limit still leaves room for the answer. Larger limits are sent as you set them, up to the model’s maximum output.
  • Short answers stay short — reasoning_effort: "none" now keeps your max_tokens as you set it on every model that can turn reasoning off, from every provider. The model catalog’s can_disable says which models those are. See What Gloo sends.
  • max_completion_tokens works too — the OpenAI name for the chat output limit is accepted, and wins if you send both.
  • Size it with confidence — a sizing table by effort and answer length, how to spot a cut-off (finish_reason: "length" or status: "incomplete"), and each model family’s levels and limits.
  • Docs — Set your output limit for reasoning models.

​
Use Claude Code and the Anthropic SDKs With Gloo

Already building with Claude? The new Anthropic Messages endpoint speaks Anthropic’s own Messages API, so the official Anthropic SDKs, the Claude Agent SDK and Claude Code work with your Gloo API key. It serves Claude models only; to call other providers’ models, use Direct Responses or Direct Completions.
  • Two-line setup — point the Anthropic SDK at https://platform.ai.gloo.com/ai/v2/direct/anthropic and pass your Gloo API key as auth_token. Your requests, streaming and error handling stay as they are.
  • Claude Code and the Claude Agent SDK — set ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, and choose a model such as claude-opus-5-5.
  • The Claude features you use — streaming, tool use, web search, prompt caching, and Claude’s thinking with Anthropic’s own thinking and output_config.effort settings. Tokens bill at the Claude model’s published rate, and web searches are billed per search.
  • Docs — Anthropic Messages, including Claude Code setup and Thinking.

​
Reasoning on Every Stream

Set a reasoning effort and watch the model reason as the answer streams, on every chat and Responses endpoint.
  • Streamed reasoning everywhere — chat streams carry delta.reasoning_content before the answer on Completions, Grounded Completions and Direct Completions. Responses streams carry a reasoning item with OpenAI’s events when you set reasoning.summary.
  • The same setting on chat and Responses — the nested reasoning: {"effort": ...} form from the Responses API now works on the chat endpoints too, and the models list shows each model’s effort levels and whether it reasons by default.
  • Usage on cancelled Claude streams — if you cancel a Claude 5 stream, or it times out, while the model is still thinking, your usage includes the thinking it had already done, up to your output limit, as it would on a finished stream. If another provider takes over the request, you’re billed once.
  • Docs — Reasoning, with each model family’s defaults.

​
Clearer Errors You Can Show Your Users

Errors from the direct endpoints now include a short message you can show your end users, with the details you need in separate fields. Every endpoint tells you plainly when to retry.
  • A message for your users, details for you — every error on Direct Chat Completions and Direct Responses has a short message, a developer_message with the next step, and an X-Gloo-Error-Code header to branch on. The message text has changed, so don’t match on it.
  • Real status codes on direct streams — a stream that fails before any output now returns an HTTP error instead of 200, and a Direct Responses stream that fails later ends with response.failed.
  • Clear retry signals — 503 with 3003 / SERVICE_UNAVAILABLE is now retryable: wait for Retry-After, then retry. On the direct endpoints, Gloo’s own timeout is now 504 (was 408).
  • Docs — Direct Endpoint Errors and Error Code Reference.

​
Gemini 4 Argon and Six More Models

Google’s flagship, Gemini 4 Argon, is now available, along with six new models for coding, vision and long documents.
  • Gemini 4 Argon — gloo-google-gemini-4-argon is built for complex coding, math and science, and multi-step agents. Set reasoning_effort to low, medium or high; the model always reasons.
  • Coding and vision — Meta Muse Spark Open (gloo-meta-muse-spark-open), NVIDIA Nemotron 3.5 Super VL (gloo-nvidia-nemotron-3.5-super-vl) for images and video, and Moondream 3 Preview (gloo-moondream-3-preview).
  • Million-token context — DeepSeek V4.1 Flash 1017 (gloo-deepseek-v4.1-flash-1017), Mercury 3 (gloo-inception-mercury-3) and Pareto 26.10 (gloo-unbiased-pareto-26.10).
  • Docs — Supported Models and Reasoning by model family.

​
Automatic Failover on the Direct Endpoints

Direct Chat Completions and Direct Responses now keep your requests going when a provider has trouble, and tell you clearly when to retry.
  • Automatic failover — when a provider can’t take your request, Gloo sends it to another provider that serves the same model, including when this happens as a stream starts.
  • One clear, retryable error — if no provider can serve the model, you get 503 with 2002 / PROVIDER_UNAVAILABLE and retryable: true. Retry with exponential backoff.
  • Clear content-policy errors — when a provider’s moderation blocks a prompt, you get 400 with provider_content_policy_violation, so you know to change the request rather than resend it.
  • Docs — Every provider unavailable and Content Policy Refusals.

​
Claude Reasoning Summaries

Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 now show how they reached an answer. Set a reasoning effort and the response includes a summary of the model’s reasoning alongside the answer.
  • Where to find it — reasoning_content on Chat Completions, and a reasoning item on Responses when you set reasoning.summary. Direct Completions also streams it.
  • Earlier Claude 5 models too — Opus 5, Sonnet 5, Fable 5 and Fable 5.1 return summaries the same way.
  • Docs — Reasoning summaries on Direct Completions, Direct Responses, Responses and Completions.

​
Reasoning Details in the Model Catalog

Pick the right reasoning model from the catalog. GET /platform/v2/models now describes how each model reasons.
  • New reasoning object — the effort levels a model takes, whether it returns a summary, and whether it reasons by default.
  • Accurate supports_reasoning — now true for every model that reasons, including the Claude models above.
  • Docs — Supported Models and the models API reference.

​
More Than 160 Models, One API

Pick the right model for the job without changing how you build. The catalog now lists more than 160 models from 14 provider families, including embedding models, all behind one API key.
  • OpenAI — the GPT-6 family (Astra, Sol and Luna, with Pro variants), GPT-6.1 Sol and Sol Pro, and Pro variants of GPT-5.6.
  • Anthropic — Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5.
  • More providers — Grok 4.7, GLM 5.3, MiMo 2.6, four NVIDIA Nemotron models, Mistral Large 4, DeepSeek V4.1 Flash, the Qwen 3.8 family, Command A Plus and Gemini Nano Banana 2.1.
  • Docs — Supported Models, the Models page and GET /platform/v2/models.

​
Retired Models Keep Working

When a provider retires a model, your requests keep working. Calls to its Gloo model ID go to the named replacement, so you can move on your own schedule.
  • See it coming — GET /platform/v2/models shows is_deprecated, a deprecation_note, and the replacement_model.
  • Docs — Model Lifecycle.

​
Gloo Code 1.22: Find, Pick Your Branch First, Make It Yours

Find what you need in any conversation and set up a session the way you like.
  • Find — ⌘F finds words in a session or an open file, with next and previous.
  • Branch before you start — choose the branch before your first message, or stay on the current one. Settings ▸ General sets the default.
  • Appearance — Reduce motion, conversation text size, avatar colours, and a light/dark switch in the profile menu.
  • Docs — What’s new in Gloo Code.

​
Gloo Code 1.21: Work in Tabs, Pick Your Branch

Run several sessions side by side and publish from the branch you choose.
  • Tabs — the desktop app opens each session or Home in its own tab. Press ⌘T for a new tab and Ctrl+Tab to move between them, drag tabs to reorder, or choose Open in new tab from a session.
  • Branch picker — on the default branch of a shared checkout, click the branch name to rename the branch or pick an existing local branch. In a worktree session you can rename it. Publishing uses the branch you chose.
  • Quitting — Gloo Code now always exits cleanly when you quit the app.
  • Docs — Tabs and the branch picker.

​
Gloo Code 1.20: Notion and Linear Built In, With Safety You Control

Bring your planning tools into the same session as your code, with you in charge of what the agent can do.
  • Notion and Linear are built in — connect from Settings, or run gloocode notion login or gloocode linear login, and choose which tools the agent gets.
  • Writes ask by default — every Notion or Linear write shows what it will send and waits for your OK. Settings ▸ Safety lets you set a tool to Allow.
  • Link an issue or page to a session — a strip above the prompt shows the branch, its pull request, and the linked Linear issue or Notion page.
  • Reserved names — an MCP server named notion or linear, or with a name starting notion_ or linear_, is ignored with a warning, so rename it. See MCP servers.
  • Docs — Notion and Linear and Safety settings.

​
September 2026

​
Gloo Code 1.19: See Your Allowance, Keep Going

Know where you stand before you run out, and get back to work fast when you do.
  • Your allowance, always on screen — a ring beside Send (and allowance 42% in the terminal footer) changes colour as you approach the limit, with a warning at 80% and the limit at 100%. Click it for the reset date and View usage.
  • Top up and keep going — at 100%, an admin gets a Top up button in the error card and can resend the same message once the top-up is done. Members see the admins’ names and can check again.
  • Steadier long runs — repeated identical calls or repeated errors get a nudge first, then the turn ends. At 80% of its step budget, the agent is told to write a handoff note.
  • Docs — Allowance and top-up and Loop protection.

​
Gloo Code 1.18: GitHub, and Commit and PR Without Leaving the App

Go from a finished change to an open pull request in one place.
  • Review, commit, and open a pull request — in the desktop review panel, pick files, edit the drafted branch name, commit message, and PR title and description, then commit, push, or open a PR in one dialog.
  • GitHub without installing gh — paste a PR link or #42 and the agent reads the issue or pull request, including private ones. By default, every comment asks first. Connect in Settings ▸ GitHub or with gloocode github login.
  • Retry and Continue — a failed or stopped turn keeps its output and can be retried or continued.
  • Unleashed mode — answers permission prompts across open projects while deny rules still apply. Also available as --unleashed for the terminal and gloocode run.
  • Docs — Commit and open a PR and GitHub in sessions.

​
Embeddings Endpoint

POST /ai/v2/direct/embeddings turns text into embedding vectors with any of the 30 embedding models in the Gloo catalog, from OpenAI, Google, Voyage AI, Mistral, Qwen, BAAI and others. It uses the OpenAI Embeddings request and response shape, so the OpenAI SDKs work by pointing the base URL at https://platform.ai.gloo.com/ai/v2/direct. Requests are billed per input token at each model’s published rate. See the Embeddings guide.

​
Canonical AI Endpoint Paths

The v2 inference endpoints now have explicit paths that identify whether requests are guarded or grounded. Existing integrations can continue using the previous paths because they remain aliased to the canonical endpoints; new integrations should use the canonical paths. The /ai/v1/grounded/responses endpoint is unchanged. /ai/v1/responses has no guardrails, so its behavior-preserving replacement is the direct endpoint. See Endpoint types.

​
Direct Endpoints

New direct endpoints send your request straight to the model you name, with no guardrails, output moderation, values-alignment, or routing added: POST /ai/v2/direct/responses and POST /ai/v2/direct/chat/completions. Your application owns its own safety and prompting.
  • Still applies — authentication, metering, the Gloo model catalog, and platform rate limits.
  • Published model rates — direct endpoints bill the published per-model rate.
  • Same request body — direct Responses takes the same body as the guarded endpoint, so moving between them is a URL change.
  • Docs — Endpoint types compares guarded, direct, and grounded endpoints. Direct Responses and Direct Completions cover the endpoints.

​
Bug Fixes and Improvements

  • Output limit reported — an answer cut off at the output limit ends with finish_reason: length (Chat Completions) or an incomplete status (Responses) instead of stop.
  • Long reasoning — a reasoning request that uses its whole output budget returns a 200 response marked incomplete, with a suggestion to raise the limit, instead of a 503.
  • Stop sequences — Chat Completions now honor stop. Up to four sequences are sent to the model; any beyond that are listed in ignored_params.
  • API keys — fixed an issue where a valid API key could be rejected.

​
August 2026

​
Prepaid Credits and Guarded Endpoint Pricing

Pay-as-you-go API access is now funded with prepaid credits: you buy credit before you use it, and each request draws down the balance at the model’s published rate.
  • Prepaid credits — buy $1 to $100 at a time from the Billing page, with optional auto-refill. Credit is available within seconds. See Prepaid Credits.
  • Guarded endpoint pricing — guarded endpoints (Responses and Completions) are priced at the model’s rate plus 6.5%, the platform price for guardrails, moderation and values alignment. Direct endpoints and Embeddings bill the published per-model rate.
  • Separate pools — a Gloo Code allowance and API credits are separate pools, and neither can be spent as the other.
  • Docs — Billing & Plans describes the current lineup and how legacy plans keep working.

​
API Key Authentication

The Gloo AI API now uses API keys for authentication — the same pattern as OpenAI, Anthropic, and Stripe. Instead of exchanging a Client ID and Client Secret for a short-lived bearer token, you create a single API key (sk_...) in Gloo AI Studio and send it directly as Authorization: Bearer sk_... in every request. No token exchange, no refresh logic.
  • Simpler integration — one credential, one header, no token management. Set GLOO_API_KEY and you’re done.
  • Industry-standard format — sk_... keys are familiar to anyone who has used OpenAI or Anthropic.
  • Key lifecycle — create keys in Studio, revoke when compromised. Revocation takes effect within 60 seconds.
  • Key security — keys are shown once at creation, then displayed as an obfuscated value (sk_...3456). WorkOS manages key material — no raw secrets stored in our database.
Existing OAuth2 client credentials are being deprecated and will be sunset by December 2026. Client ID / Client Secret credentials continue to work during the transition period. To migrate, create a new API key in Studio and update your applications to send it directly as a bearer token.

​
Model Lifecycle and Deprecation Fields

GET /platform/v2/models is now the one place to detect new models and deprecations. Deprecated models are flagged with is_deprecated, and carry a deprecation_note and a replacement_model where one has been chosen, so you can find models that are about to change without watching for an announcement.
  • Docs — the Model Lifecycle guide explains how models are added, what happens when one is retired, and how to detect both from the API.

​
July 2026

​
Responses API v2 — Now with Gloo’s Full Guarded Pipeline

POST /ai/v2/responses is the new Responses API endpoint. It keeps the same OpenAI-compatible request/response shape as v1 and adds the full pipeline that powers /ai/v2/chat/completions — guardrails, blocked-content handling, values-aligned tradition responses, and output moderation.
  • Same shape, more features — input, instructions, typed output[], and named SSE streaming, now with guardrail parity with Completions V2.
  • tradition on Responses — the values-aligned tradition parameter, previously Completions-only, now works on the Responses shape.
  • Multimodal, guarded — image input (vision) and image generation work on the same endpoint, with the guard layer applied to the conversation.
  • Predictable blocked handling — guardrail static responses arrive as normal Responses payloads; hard blocks return the same 403 as Completions V2.
  • v1 deprecated — POST /ai/v1/responses is deprecated in favor of v2. It keeps working. v1 behaves like the direct endpoint (no guardrails or output moderation), so the behavior-preserving move is a URL change to /ai/v2/direct/responses. Move to /ai/v2/guarded/responses when you want guardrails, moderation and values alignment, at the guarded platform price. See Endpoint types.
  • Docs — the Responses guide now covers v2.

​
Grounded Responses — RAG on the Responses API

Grounding comes to the Responses API. The new POST /ai/v1/grounded/responses endpoint accepts the standard Responses API request extended with grounding controls, retrieves relevant passages from your uploaded publisher content, and returns the familiar typed output[] — plus a sources_returned flag (or X-Sources-Returned header when streaming) confirming the answer was grounded.
  • Same shape, grounded — no need to switch to the chat-completions format to get publisher-grounded answers.
  • Grounding controls — rag_publisher scopes retrieval to your content, sources_limit (1–10) tunes how many sources are injected, and certainty_threshold sets a minimum similarity bar for sources.
  • Explicit grounding signal — every response reports whether sources were used, including the zero-sources case.
  • Docs — A new Grounded Responses guide covers the request format, streaming, and how it compares to Grounded Completions.

​
June 2026

The new Responses API (POST /ai/v1/responses) is the recommended way to build on Gloo AI. It’s OpenAI-compatible and exposes the full model catalog through a single endpoint for text, vision, reasoning, tool use, and native image generation — with one typed request/response shape across every provider.
  • Multimodal in one endpoint — image input (vision) and image generation alongside text, no separate endpoints or bespoke payloads.
  • OpenAI-compatible — use the official OpenAI SDKs (responses.create); just point the base URL at https://platform.ai.gloo.com/ai/v1.
  • Typed, structured output — responses return an output[] array of typed items (messages, tool calls, generated images, reasoning).
  • Completions V2 stays supported — fully backwards-compatible, and still the home of auto-routing, model_family, tradition, and grounded completions.
  • Docs — A new Responses API guide covers usage, pricing, tool use, and migrating from Completions.

​
Prompt Caching Is Widely Available

A major cost-saving feature for all customers. When you send repeated prompt prefixes — system instructions, few-shot examples, large context blocks — the platform caches them and serves subsequent requests at a fraction of the cost. Cache read tokens are priced well below standard input tokens.
  • Studio — The Usage page now breaks spend down by cache segment (fresh, cache read, cache write), so you can see exactly what caching is saving you. Model cards show cache pricing alongside standard rates.
  • Programmatic access — Cache token counts and pricing are now reported programmatically, so savings can be calculated automatically.
  • Docs — A new Prompt Caching guide covers how caching works across Anthropic, OpenAI, DeepSeek, and Gemini, with code examples and a supported-models matrix.

​
GlooCode v1.6 — Web UI & More

Four GlooCode releases shipped this week (v1.5.3 → v1.6.2), a significant milestone for the product.
  • Web UI — GlooCode now includes a built-in web interface with 30+ themes, file diffing, session management, and support for 18 languages.
  • Faster, more reliable updates — gloocode update now performs direct binary swaps, and install.sh always performs a fresh install.
  • New built-in agent optimizations — Better and more intelligent default agent models.
  • More reliable releases — Hardened testing and guaranteed-fresh installers on every release.

​
Platform Security & Trust

  • Org-level access controls — Administrative tooling to immediately revoke platform access at the organization level.
  • GDPR right to erasure — Enterprise customers can request full deletion of user data across the platform, along with data export for subject-access requests.

​
Data Engine

  • Cache cost reporting — Usage reporting and CSV exports now break spend down by cache segment.
  • Performance under load — Improved reliability and throughput at scale.

​
Billing

  • Free Plan retired — New users must add a payment method during onboarding to activate their Pay-as-you-go plan and unlock the Studio Playground, API access, and other features. Existing paid plans are unaffected. See Billing & Plans for the updated lineup.
  • Enterprise billing flexibility — Enterprise admins can now switch to pay-as-you-go billing.
  • Automatic subscription hygiene — Downgrades automatically clean up associated subscriptions.

​
Week of May 26, 2026

​
11 New AI Models

  • Claude Opus 4.8 — Anthropic’s latest, available across all providers with fast-mode support
  • GPT-5.3 Codex — OpenAI’s reasoning-focused coding model with 400K context
  • Claude Opus 4.7 Fast & 4.6 Fast — Fast-output Opus variants for speed-sensitive workloads
  • Qwen 3.5 Plus, Qwen 3.5 27B, Qwen 3 235B A22B variants — Expanded Qwen family, including thinking/reasoning variants
  • GLM-4.7 Flash — Z-AI’s fast inference model
  • MiMo V2 Flash & V2.5 — Xiaomi’s new MoE models optimized for reasoning, coding, and agentic tasks
  • Gemini 3.1 Flash Lite — promoted from preview to GA

​
Studio

  • Spend visibility for all orgs — Every organization can now see its spend breakdown on the Usage page.
  • Enterprise → pay-as-you-go — Enterprise admins can switch to flexible billing directly from the Billing or Plans page.
  • Smarter usage charts — Date labels respect your timezone, and charts clearly indicate hourly, daily, weekly, or monthly granularity.
  • Improved GlooCode usage filters — Faster, more reliable filter dropdowns.

​
Completions

  • Prompt caching (beta) — Now works consistently across all completion types, with cache hit rates and token counts in usage reporting.
  • More reliable streaming — Improved stream stability for reasoning models, including during long thinking pauses.
  • Smarter guardrails — Tuned content filtering reduces unnecessary blocks, with graduated warnings replacing hard blocks in more cases.
  • Clearer errors — Provider issues now surface as clear, actionable errors.

​
GlooCode

  • v1.5.0 → v1.5.2 — GlooCode now runs on an improved background service for more reliable sessions; existing installations upgrade via gloocode update.
  • More reliable releases — Hardened release process and installer delivery.

​
Data Engine

  • Enterprise self-serve billing — Enterprises can transition billing directly in Studio.
  • Timezone-aware metrics — Usage reporting now supports timezone-aware grouping with correct daylight-saving handling.
  • Ingestion refinements — Improved YouTube ingestion consistency for publisher metadata and status tracking.

​
Documentation

  • Restructured and refreshed console documentation and developer reference for better discoverability.

​
Week of May 18, 2026

​
9 New Models in the Catalog

  • Gemini 3.5 Flash — 1M context, major coding gains at high speed
  • Gemma 4 31B — Google’s most capable open-weight model, with vision
  • GPT OSS 20B — OpenAI’s open-weight MoE model for consumer hardware
  • Ministral 3B / 8B / 14B — Mistral’s edge-to-server lineup
  • Qwen 3.7 Max — 1M context for agentic workloads
  • Qwen 3.5 397B and 122B — Large MoE models
  • Gemini 3.1 Flash Lite — promoted to GA

​
GlooCode v1.5.0 — Sub-Agents & Persistent Sessions

GlooCode is Gloo’s autonomous AI coding agent — it writes, tests, and ships code while intelligently optimizing for cost and quality.
  • Improved background service — More reliable sessions and sign-in.
  • Explore sub-agent — Fast codebase context-gathering for more relevant results.
  • Session persistence — Sessions persist through long, context-heavy work.
  • Org-level usage analytics — Workspace, user, project, and branch tracking, so organizations can see exactly who’s using what.
  • Studio integration — New GlooCode product and get-started pages with a cost comparison calculator and agent overview, plus a usage analytics dashboard with user/branch filters and cache hit rate metrics.

​
Studio

  • New Data Engine pages — Dedicated product pages for ingestion, search, enrichment, and a hub overview, each with polished motion design and market context.
  • Community section — A new Summer Challenge page with toolkit cards for YouVersion and Gloo AI APIs, judging criteria, and a responsive timeline.

​
Enterprise Billing Overhaul

  • Dedicated enterprise billing view with Pro upgrade path
  • Billing history scoped to org admins
  • Per-cache-segment cost breakdowns and per-model output limits across all providers

​
Reliability & Quality

Continued investment in platform resilience: more stable streaming for reasoning models, more reliable spend-limit handling, hardened content enrichment, clearer provider-error reporting, structured-output validation for applicable models, and faster recovery during provider capacity events. Studio also received fixes for mobile layout, usage filters, and billing views.

​
April 21 – May 15, 2026

​
New Features

  • GlooCode v1.0 — Gloo’s autonomous AI coding agent launched as a single native binary for macOS and Linux. Its specialized agent roles (plan, build, explore, review, QA) are each optimized for cost and quality.
  • Citations — Grounded completions now return source citations including title, URL, author, publisher, date, and content snippets.
  • 20+ new models — Including GPT-5.5, GPT-5.5 Pro, GPT-5.1, Claude Opus 4.1, DeepSeek V4 Flash, DeepSeek V4 Pro, DeepSeek R1-0528, Devstral 2, Qwen 3.6 Max Preview (1M context), Kimi K2.6 (vision), and others.
  • Model Details pages — Every model in the catalog now has a dedicated Studio page with specs, pricing comparison, modalities, and capability flags.
  • Enterprise provisioning — Enterprise accounts are provisioned and activated automatically, including plan changes and cancellations.
  • Usage page model breakdown — New “By Capability” and “By Model” views with human-readable names, cross-provider merging, stacked input/output bars, and per-credential filtering.
  • Multilingual safety messaging — French and Dutch support for safety responses and corrections; refusal language now respects the user’s language.
  • Prompt caching — Automatic across supported providers, with full cache-hit visibility.

​
Bug Fixes

  • Completions streaming responses are more resilient, with retries for empty responses
  • Tool-call-only streams are no longer misclassified
  • Spend-limit responses return clear status and retry guidance
  • Structured output is now forwarded consistently for applicable models
  • Model refusals now surface explicitly so callers can handle them programmatically
  • Rate-limit handling improved for long-standing customer accounts

​
Improvements

  • Significant streaming latency reductions across the completions pipeline
  • All error codes enriched with name, category, and description
  • Studio organization administration surfaces revamped with detail pages, stat cards, search, sorting, and inline editing
  • Polished page transitions across Studio

​
February 11 – April 20, 2026

​
New Features

  • Multi-provider routing — Integrations with Anthropic, OpenAI, Google Gemini, DeepSeek, and AWS Bedrock, with automatic health-based failover that shifts traffic when a provider degrades and seamlessly continues responses on a backup — without the caller ever seeing the failure.
  • 13 new models — Claude Opus 4.6, Claude Sonnet 4.6, Claude Opus 4.7 (available same day as Anthropic’s release), GPT-5.2 Pro, GPT-5.4, GPT-4.1, GPT-4.1-mini, GPT-5-Nano, Llama 4 Maverick, DeepSeek V3.2, DeepSeek R1, and Llama Guard 4 12B.
  • Public model catalog — Pricing, speed ratings, and capability flags (tools, streaming, reasoning, vision) for all available models are now publicly accessible.
  • Vision support — Image inputs across all providers.
  • Structured error responses — Every error includes an error code, fault attribution (client/provider/platform), and a retryable flag.
  • Trace IDs — Every response includes a trace ID for request correlation.
  • Self-serve Pro tier — Plan selection, upgrades, promo codes, and payment management entirely in-product. A developer can discover, evaluate, pay, and build without a sales conversation.
  • Spending limits — Per-organization budget caps with a contact-sales path for over-cap requests.
  • In-product support — A Support button in the Studio sidebar opens a chat with Fin, Gloo’s AI support agent.

​
Bug Fixes

  • OpenAI tool conversion issue resolved
  • Content updates appear in search and grounded completions much faster
  • Organization auto-activation extended to all new organizations
  • Prayer guardrail detection improved

​
Improvements

  • Major latency reductions across the request pipeline
  • Streaming reliability upgrades — faster first responses and more robust mid-stream handling
  • YouTube ingestion overhauled with automated reliability monitoring
  • Content enrichment hardened with automatic retries
  • Studio Content Library rebuilt — instant loads, with live updates for in-progress items

​
January 25 – February 10, 2026

​
New Features

  • Spending limits — Weekly budget caps with retry guidance when limits are reached
  • New models — Claude Sonnet 4, Llama 4 Maverick, and DeepSeek R1 added to the catalog
  • Smarter guardrails — Edge cases now receive contextual guidance toward appropriate alternatives instead of hard blocks
  • Batch file upload — Upload multiple files to Data Engine in a single request
  • Performance tracking — Response-time measurement to drive latency improvements
  • Payment method gating — Payment method required before platform access is granted
  • Per-organization billing — Spending controls and billing attribution scoped to individual organizations

​
Bug Fixes

  • YouTube transcription failures resolved across multiple scenarios including duplicate detection, live streams, playlists, and timeouts
  • Batch ingestion now reports per-item errors with full transparency

​
Improvements

  • Better handling of large document batches
  • Smarter model routing for cost and performance

​
January 18 – 24, 2026

​
New Features

  • Enterprise organization controls — Plan-based search access and admin controls for publisher settings
  • Payment method onboarding — Self-serve customers add payment methods during signup, with the flow adapting to organization type
  • JSON file uploads — JSON files can now be uploaded and processed in Data Engine
  • Sitemap fallback — Sites with broken sitemaps now fall back to direct page crawling instead of failing

​
Bug Fixes

  • Hardened input validation in Data Engine ingestion
  • Search failures resolved for publishers without enrichment configured
  • Content upload failures for new publishers resolved
  • Content Library progress indicators fixed
  • Access denied errors for customers with valid credentials resolved
  • Prayer response guidelines for Faith Assistant now trigger reliably

​
Improvements

  • AI model upgrade in Faith Assistant delivers better responses at lower cost
  • Full light mode support across Studio
  • User role badge added to the profile dropdown
  • Long API keys display correctly in the Studio API Key modal

​
January 11 – 17, 2026

​
New Features

  • Streamlined billing onboarding — Connecting a payment account now provisions metering automatically, with optional pay-as-you-go enrollment, so organizations activate without manual setup
  • Automatic organization activation — Organizations activate automatically when onboarding completes
  • Dynamic search routing — Completions automatically route to the correct data collection based on publisher configuration
  • Context window error guidance — Requests exceeding a model’s context limit now return the limit, the affected model, and suggested higher-capacity alternatives

​
Bug Fixes

  • Content enrichment status now reflects actual completion state
  • 5 broken models removed from the catalog

​
Improvements

  • Guardrail accuracy improved for theological content