Updates, improvements, and fixes for the SpiderGate unified AI API gateway.
We've migrated the Changelog to a fully componentized architecture, just like the press newsroom!
Components: The changelog is now powered by sg-changelog-list and sg-changelog-hero.
Data Bindings: Uses the live changelog data source on Edge.
Enjoy the new architecture!
A prompt you got right is worth keeping. Save the composer as a named bundle — the system prompt, the model, its settings and any reference images — then reload it in one click from the Prompts gallery, or type /prompts to open it from the keyboard. Saved prompts live inside a Studio project, so each project keeps its own set.
Your agents can call them by name. Anything holding your API key — an agent, a script, our CLI or MCP tools — can put prompt.<name> or prompt:<id> in the prompt field of a chat or media-generation request, and SpiderGate expands the bundle server-side. Anything the request sets itself wins, so a saved prompt is a sensible default you can override per call rather than a straitjacket. A reference that does not resolve is now an error — a 404 when nothing matches, a 409 naming both candidates when two prompts share a handle — instead of being quietly treated as ordinary prompt text.
When the model you have selected accepts an image to work from, an Add to prompt button appears in the composer, or type /reference. It opens a visual grid of this project's media, your wider media library, and an upload tile for something new. The image you pick is shown inline in the conversation rather than reduced to a filename, and after a generation runs, Use as next reference carries it into the following prompt. On a model that declares no image input the button is still there but disabled, and it tells you why.
The Images, Video and Audio canvases now open on a grid of everything you have already made in that project, newest first. Reuse loads a generation's whole recipe — the exact prompt, the model, the aspect ratio and the panel settings — back into the composer, so you can change one word and run it again. Use as reference attaches the generation itself to your next run.
Find all three in your dashboard under SpiderGate → Studio. The full guide is in the Studio documentation.
The Studio now keeps the work instead of just the last message. A project holds your conversations and everything you have generated inside it, you can put one prompt to several models at once, and every run leaves a trace you can open afterwards.
Fork and Merge. Send one prompt to several models at the same time. Fork streams the answers side by side so you read every lane yourself. Merge passes those answers to one more model that fuses them into a single reply, with the raw lanes kept behind a toggle. Up to five lanes, on the Quality, Budget and Fast presets or on models you pick.
Projects hold the whole session. A project is now a real container: its conversations, its saved prompts and its generations all live together, and switching project switches the lot. Chat, Images, Video and Audio share it.
A trace on every run. Each request records which model actually served it, how long it took and what it cost, including one row per lane when you fork. Open any of them from the Traces tab.
Full walkthrough in the Studio guide.
Provider keys used to live wherever somebody left them. A teammate held the Groq account, a client owned the OpenRouter billing, and every one of those secrets had to travel through a chat window to reach me. The Vault replaces that. A key goes in encrypted, and I never see it.
Contribute a key without exposing it. Send an invite and the other person pastes their key on their own machine. No login, no account with us. The key is encrypted before it is stored, and only a masked preview comes back. 31 keys have arrived this way.
Three key kinds, sorted by how they bill. Free and subscription keys join the shared pool. Pay as you go keys never do, and that rule is held by a database trigger, a constraint, and a rejected write rather than by a code review.
Round robin selection with failover. Every request picks the healthiest, least used key. Three consecutive errors drop a key from selection, and the next request simply completes on the next healthy one. Across the last 30 days, 68 different keys served 169,581 requests.
Spend and quota visible per key. Every request is logged against the key that served it, so cost reads per key instead of as one monthly total. Subscription plans meter against their own rolling window rather than a dollar amount.
As of July 2026 the Vault holds 75 active keys across 15 providers, 50 of them pooled and rotating.
Read how it works in the docs at https://gate.spideriq.ai/docs/dashboard/vault
SpiderGate's model catalog now tells you what each model is actually good at. Nine capability categories, a score in each, and the benchmarks behind every number. A daily agent keeps the facts current.
Capability scores on 221 models. Each scored model carries a number in nine categories: Coding, Design/Frontend, Reasoning, Math, Vision, Research, Tool Use, Chat and Translation. The score is a within-category percentile, not a raw benchmark average.
Rank the catalog by what you need. Filter by a capability chip and sort by that category's score, in the dashboard or through the gateway. Your agents read the same ranking your team does, so a model choice made in code matches the one you would make by hand.
The benchmarks behind every score. Open any score to see the results feeding it, drawn from 4,821 benchmark rows across LLM-Stats and DesignArena, with vendor-reported numbers flagged as vendor-reported so you can weigh them yourself.
A daily agent keeps the catalog current. It pulls specs and pricing from OpenRouter, benchmarks from LLM-Stats and lineage from Wikidata, stamps every field with its source and fetch time, and never overwrites a row a human has curated.
Coverage is stated plainly: 221 of the 1,216 models in the catalog are scored, and every other model is marked not scored yet rather than ranked last.
Full release notes and the method behind the scores: https://gate.spideriq.ai/blog/rank-models-by-capability
SpiderGate routes chat across providers on one key. This release extends the same token, the same call shape and the same trace view to media: image, video, speech, transcription and embeddings.
Media generation now runs on your gateway token. POST /media/generations takes a provider/model id and a params object, returns a stored URL, and records the spend beside your chat traffic. 33 of the 81 catalogued media models are active today.
Every model publishes the parameters it accepts. GET /media/models returns each active model's declared inputs, so an agent reads the contract instead of guessing. An undeclared parameter is dropped, and a declared one carrying an illegal value returns 422 invalid_param before a key is spent.
Embeddings answer to task aliases. agent/embed-small, agent/embed-large and agent/embed resolve server side to the matching embedding model, so retrieval code stops naming a vendor. Raw model names still pass through unchanged, and every call is metered and traced like chat.
Generation runs under a wall clock. Image and speech cap at 600 seconds, video at 900, and an overrun returns 504 generation_timeout instead of hanging. Every dispatch failure now writes a typed row, so a degraded model shows up as a number.
Media is bring your own key: all 33 active models are paid tier, and none of them draw on the shared pool.
Full endpoint reference in the docs at https://gate.spideriq.ai/docs/reference/api.
A provider key does not announce that it has died. It starts returning 401s, and on a shared pool the damage is quiet: requests fall through to a fallback model and quality drops before anyone thinks to open the vault.
A failing key leaves the pool on its own. Three consecutive failures mark a credential unhealthy and routing skips it. A credential that keeps failing authentication, three times inside 24 hours, is deactivated outright, so a revoked key cannot keep taking a share of your traffic.
The person who can fix it gets told, and only them. The contributor who added the key gets a re-authentication link, up to three times, 48 hours apart, and never inside their quiet hours. Brand admins get a single daily digest at 09:00 local instead of one message per key.
Re-authentication happens in place. The link stays valid for 120 hours, outlasting the whole reminder cycle. Paste a fresh key, or re-login for OAuth, and the same vault entry updates: same key id, same sharing and limits, history intact. The engine reloads without a restart.
A key having a bad minute comes back by itself. A credential that has been unhealthy for 20 minutes with no authentication errors returns to the pool automatically. Rate limits and timeouts heal on their own, so only genuinely dead credentials wait on a human.
The result: a revoked key leaves the rotation on its own, and the only person who has to act is the one who can actually replace it.
Read the full guide at https://gate.spideriq.ai/docs/dashboard/key-health.
SpiderGate has been logging every request your agents send through the gateway since March. This release is where that log becomes something you can open: one trace per request, filtered by agent, showing the model that actually served it and what it cost.
Every request has a trace you can open. The Traces tab lists one row per request with the agent, the requested model, the model that actually served it, latency, tokens and cost. Click any row for the full span waterfall: auth, routing, the model call, tracking.
Spend attributed to the agent that caused it. Filter the list to a single agent token, or jump straight there from any agent card in Keys. On a shared key pool your provider bill is per key, so this is the only place the answer is per agent.
The trace list survives its own backend. Request metadata goes to SpiderGate's own store on every call. Full message bodies and span timings live in LangFuse, a separate service. If LangFuse is unreachable, the list and the summary stats still load, without the bodies.
Fallbacks stop being invisible. When a request falls through to the second model in an alias chain, the trace records which model answered, so a quiet drop in quality has somewhere to be looked up.
Across the gateway that is 327,759 traces recorded since 29 March 2026, as of 10 August.
Read the full guide at https://gate.spideriq.ai/docs/observability/traces.
Invite Contributor lets a brand admin collect a provider credential from someone who has no SpiderGate account. You send a signed link, they submit one key, and it lands in your vault. Live since 10 April 2026.
A contributor needs no seat. The invite link opens a single-purpose page that sits outside the dashboard login. The person holding the key never creates an account, never picks a password, and never reaches anything else in your workspace.
Keys are validated before they are stored. Every submitted key is tested against the provider first. A key that fails is reported on the page and never written to your vault, so a dead credential cannot sit there looking healthy.
One invite can request up to 25 providers. A single link carries the whole list with a progress bar and closes only once every requested provider is in. Re-submitting a provider updates the existing entry instead of creating a duplicate.
Every link expires on its own. An invite is good for 72 hours. You can cancel a pending invite or resend it, and each key is labelled in the vault with who contributed it, so months later you know where a credential came from.
You get the capacity you need for a client's work without holding their password or paying for a seat.
Full walkthrough in the docs at https://gate.spideriq.ai/docs/dashboard/invite-contributor.
SpiderGate is one OpenAI-compatible endpoint sitting in front of every model I can reach. In v2.0.0 I stopped addressing models by name and started addressing them by job, so the model behind a call can change without the caller ever knowing.
You ask for a task, not a model id. Send spideriq/coding or spideriq/extraction where you used to send a vendor model string. There are 32 of these aliases today, and each one resolves to a real model at request time rather than at deploy time.
Every alias has somewhere to fall back to. Those 32 aliases carry 91 ordered slots between them, so a chain runs 2.8 deep on average. When the first model rate limits or goes dark, the next slot answers and your call still returns.
One catalog, one credential. 1,216 models across 16 providers are catalogued, and 61 are key backed right now. You hold one SpiderGate token; I hold the provider keys, the pooling and the routing behind it.
Routing runs on litellm.Router. BerriAI's router does the least busy selection, the 3 retries and the 60 second cooldown. What SpiderGate adds on top is the task alias, per tenant auth and a usage record for every call.
A model id you pinned in March is a model id that breaks in June. An alias is not.
Read the routing reference in the docs at https://gate.spideriq.ai/docs/using-the-gateway/task-aliases