What we've shipped
Reverse-chronological list of what's actually shipped on the InferAll AI gateway and the marketing site. Each entry links to the public PR where the change landed; gateway-side changes that live in the private infra repo are marked as such.
2026-10-16
nvidia/nemotron-3.5-super-vl-120b-a12b is available on NVIDIA NIM at $0 input and $0 output, on both /v1/chat/completions and /v1/messages, and it accepts images. It reasons before it answers, so give longer answers room: at max_tokens 800 a 150-word reply can be used up by reasoning.
2026-10-15
stepfun-ai/step-5-preview is available on NVIDIA NIM at $0 input and $0 output, on both /v1/chat/completions and /v1/messages. It is a preview model, so NVIDIA may change or retire it without much notice. The no-card trial can call it like any other $0 model.
2026-10-13
gemini/gemini-4-argon is available on paid balances: a 1M-token context, 32K output and built-in thinking. It costs $2.60 per million input tokens and $13.00 per million output tokens (thinking included) until November 30. Google doubles its own price on December 1, and ours follows on that date: $5.20 and $26.00.
2026-10-12
nvidia/llama-3.1-nemotron-safety-guard-8b-v3 has not answered since 10-11 (every call timed out at 45 seconds, and NVIDIA itself gave no answer within 120 seconds), so we no longer list or route it. Requests for it now get an immediate refusal that names nvidia/nemotron-3.5-content-safety, a live safety model. We will bring it back when it answers again.
2026-10-09
anthropic/claude-haiku-5-5 is available on /v1/messages and /v1/chat/completions, and in the playground. Anthropic prices it by prompt length: up to 100,000 prompt tokens it is $0.10 in and $0.50 out per million, above that $0.50 and $2.50, where the prompt counts cached tokens too. temperature and top_p are dropped for this model because Anthropic rejects them.
Cache hits on claude-sonnet-5-5 now bill at Anthropic's rate of 5% of the input price ($0.10 per million before our multiple) instead of 10%.
2026-10-08
nvidia/llama-3.1-nemotron-safety-guard-8b-v3 is served again. It was removed on 10-06 after repeated timeouts; 13 test calls today all answered (median 6 seconds). If it starts timing out again, /v1/chat/completions stops waiting after three timeouts and answers 503 with Retry-After for ten minutes instead of making each caller wait.
Each model in the OpenAI-style list at /v1/models has a pricing object with the prompt and completion price you are billed per token, in USD, as decimal strings (for example claude-opus-4-8: prompt "0.00000425"). Free models show "0". The existing id, object, created and owned_by fields are unchanged.
2026-10-07
grok-4.7 is available on /v1/chat/completions as xai/grok-4.7, and in the playground. It is billed at xAI's reported cost for each call plus 30%. /models now lists every Grok model at its normal rate (it used to show the long-context rate); prompts over 200K tokens cost twice as much at xAI.
claude-sonnet-5-5 is available on /v1/chat/completions and /v1/messages as anthropic/claude-sonnet-5-5, and in the playground, at Anthropic's rate minus 15%. Like the rest of the Claude 5 family it does not accept temperature or top_p; the gateway drops both for you on /v1/chat/completions. top_p is now also dropped for the other Claude 5 models there, where it used to be forwarded and rejected. Cache reads on Claude Fable 5.1 and Opus 5.5 are now billed at Anthropic's lower cache rates for those models.
gpt-6-astra, gpt-6.1-sol, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna are available on /v1/chat/completions with the openai/ prefix (for example openai/gpt-6-sol), and in the playground. Like the gpt-5 line they reject max_tokens and a temperature other than 1 upstream; the gateway sends max_completion_tokens and drops temperature for you. Prompts over 272K tokens bill at OpenAI's long-context rate. Prices are on /models.
A text part marked with cache_control: {type: "ephemeral"} in a system or user message now reaches Claude, so a repeated long prefix is read from the cache at 10% of the input rate. The marker used to be dropped on this endpoint. Cache writes and reads are now billed at Anthropic's cache rates on both /v1/chat/completions and /v1/messages; before, cache tokens were not counted, and a non-streamed /v1/messages call to Claude was billed a flat 2.5 cents whatever its size. See Prompt caching in the docs.
Gemini models are billed from Google's published rates plus 30%. Before this, most Gemini ids were billed at a placeholder $0.15/$0.60 per million tokens, well under Google's price for models like gemini-3.5-flash ($1.50/$9.00) and gemini-3.1-pro-preview ($2.00/$12.00); gemini-2.5-flash is now $0.30/$2.50 before the margin. Pro prompts over 200k tokens bill at Google's long-context rate, cached input at 10% of the input rate, and image models at their image-output rate. Gemini ids with no published price (the -latest aliases, some preview ids, embedding and transcribe ids) are no longer listed or served; use a listed id such as gemini-3.8-flash.
/models now lists the rate you pay for every model: the provider's rate plus 30%, or 15% below it on Anthropic models, with the provider's own rate beside it. Media models list the per-image or per-second price you pay, including the $0.01 per image and $0.005 per second minimums. Before this, /models and the credits section of /billing said premium models bill at the provider's rate with no markup, which was not accurate. GET /ai/v1/models adds billedInputPerM and billedOutputPerM to token models and billedPerImage or billedPerSecond to media models; inputPerM and outputPerM still carry the provider's rate.
Token models bill at exactly the provider's rate times the margin above. A $1 per million token minimum used to lift the cheapest paid models (gpt-4.1-nano, gpt-4o-mini, Claude Haiku 4.5) above that, up to twice the provider's rate.
2026-10-06
The open NVIDIA NIM models are $0 input and $0 output; a 0.5 cent per-request minimum that applied to accounts with a card is gone. On /ai/v1/generate (the endpoint the Python and TypeScript SDKs use), a request that fails is no longer recorded as spend in your usage totals.
2026-09-27
Current OpenAI SDKs and the OpenAI Agents SDK default to the Responses API; it now works with every model, including tools, tool results, structured output (text.format), images and token-by-token streaming. previous_response_id and OpenAI's hosted tools (web search, file search) are not supported yet and return a clear error. A guide for the Agents SDK is on the blog.
Streamed requests to Claude and Gemini used to deliver the whole answer as one chunk at the end. They now stream as the model writes: on a 250-word answer, Claude's first text arrived after about 1.1s instead of 4.8s.
Gemini 2.5 and 3 models think before answering, and that thinking counted against max_tokens: a 600-token limit could return a 20-word answer. Thinking is now off for Flash models and minimal for Pro models unless you ask for it with reasoning_effort (low, medium or high) or Anthropic's thinking parameter. A reply that is cut off now reports finish_reason length instead of stop.
NVIDIA's Nemotron 3 models reason before answering. The gateway now leaves room for that reasoning so structured output is not cut off, and reasoning_effort none or minimal turns it off.
When NVIDIA reports an error at the start of a stream, the request now moves to another free model instead of failing. A request for a free model only ever falls back to other free models.
Current Claude Code failed on its first request to Claude models (a missing anthropic-beta header) and to Gemini models (a tool-schema field Gemini rejects). Both are fixed and tested with the real CLI.
OpenAI's open-weight gpt-oss-20b is served through NVIDIA at $0. It had been wrongly reported as retired. Free models now also have their own pages under /models.
2026-09-26
A streamed response could stop partway and never finish when you did not set stream_options.include_usage (the default). It hit OpenAI models most often: in our test, 4 of 5 streamed gpt-4o-mini tool calls hung. The cause was in our stream filter, not the provider. Fixed; the same request now completes every time in under a second.
Images sent to Gemini on either endpoint, to OpenAI models on /v1/messages, or to Claude on /v1/chat/completions were dropped, and the model answered without them. Both OpenAI image_url and Anthropic image blocks, base64 or URL, are now converted for every provider.
Gemini: nested models ($ref/$defs), optional fields (anyOf with null) and properties named title or default now reach the model intact; before, Pydantic-style tools could come back with malformed arguments. /v1/messages now honours tool_choice (forced, named and none) for OpenAI, NVIDIA and xAI models. A named tool call from an OpenAI model now reports stop_reason tool_use, so Anthropic-style agent loops run it. On NVIDIA models, tool_choice required now returns a structured tool call instead of text.
The Anthropic token-counting endpoint now exists. Claude models return Anthropic's exact count; other models return an estimate, and the X-Inferall-Token-Count response header says which.
Some NVIDIA models reject fields they do not know. The gateway now removes safety_identifier, verbosity and any field whose name starts with an underscore before sending to non-OpenAI providers.
Claude Opus 5.5, Opus 5, Sonnet 5 and Fable 5.1 are available with the anthropic/ prefix. Claude Opus 4.5 to 4.8 were priced at the older Opus 4.1 rate ($15 / $75 per million tokens) instead of Anthropic's $5 / $25. That is corrected; current rates are on /pricing.
2026-09-02
Entries below, and 17 blog posts, the billing and keys pages, the comparison pages, the shipped code examples and an outreach template, said premium provider tokens (OpenAI, Anthropic, Google) are billed at the provider's published rate with zero markup. That was not accurate: InferAll bills premium tokens at its own published per-token rates, which sit above the underlying provider cost; the applied margin measured 1.30x across those providers. Every live surface has been corrected and the blog posts carry dated correction notes; current rates are on /pricing. ⚠️ Amended 2026-09-03: "every live surface" was not true: the shipped @inferall/cli still said "$0 markup" in two places and went uncorrected on npm for another week. A completeness claim in a correction is itself a claim, and this one was not checked. Now fixed. Older entries in this changelog are left as written because a changelog is a record of what was announced at the time, and this entry is the correction to them.
site (private) · infra (private)
2026-07-23
npx @inferall/cli init detects your project (Node or Python), writes INFERALL_API_KEY to .env (and .env.example, and makes sure .env is gitignored), and points your existing OpenAI or Anthropic SDK client constructors at the gateway — the OpenAI SDK gets baseURL https://api.inferall.ai/v1, the Anthropic SDK gets the host root since it appends /v1/messages itself. Simple constructors are rewritten deterministically; ambiguous ones get an AI-suggested patch you review first, generated by InferAll's own $0 NIM models (the wizard dogfoods the gateway). Google Gemini SDK usage is detected too — that SDK speaks its own protocol, so instead of a risky rewrite you get the exact OpenAI-compat snippet to migrate at your own pace. The final step is a live call that reports your real account state (active, or the one-time $5 activation with the link) instead of failing opaquely. Also ships: doctor (CI-friendly setup health check, non-zero exit on any problem), models (the live catalog with real prices — 180+ models, 80+ at $0), and status (tier, activation, balance, trial, 7-day savings). Zero runtime dependencies, Node 18+. Published as @inferall/cli 0.2.0 on npm.
2026-07-13
Every 402 payment-required response now includes X-Inferall-Gate-Reason: paid | cap | no_card | card_required. Lets SDK wrappers branch on the reason without JSON-parsing the body — e.g. auto-swap the request to a $0 NIM model when reason=paid, or halt retries entirely on reason=cap|no_card|card_required. Companion X-Inferall-Upgrade-URL header (and the upgrade_url JSON field) also now carry a ?src=402_<reason> query tag, so /billing landings from terminal-rendered 402 errors capture the reason in PostHog and let us measure the missing middle step: hit-gate → click-URL → open-checkout → pay. All four 402-emitting sites tagged: /ai/v1/generate (index.ts), /v1/messages (compat.ts), /v1/chat/completions (openai-compat.ts), and the free-model card-required gate (limiter.ts checkCardRequired). Full documentation on the response-headers table at /docs/rate-limits.
infra (private)
The pre-payment 200-call trial was gate-able on has_card (any Stripe customer relationship), and has_card was set at /api/billing/start when we create the Stripe Customer — before any card is attached. Bot clusters were minting throwaway Stripe customers and burning NIM quota per key without ever paying. New gate is has_paid_successfully — a settled real-money charge. In exchange, the 200-call cap is gone entirely: once your $5 activation settles, NIM open-source models stay $0 in/out with no per-user request cap beyond the standard free-plan daily limits (50 chat / 100 text / etc.). The $5 also becomes spendable balance for premium providers (OpenAI, Anthropic, Google) at zero markup [corrected 2026-09-02]. Reframed the marketing on /pricing, /billing, /auth, /docs, /docs/rate-limits, and /keys to match. Trade-off is honest: no more free pre-payment trial — but if you want to evaluate before committing, reply to any /billing email and we'll grant an evaluation credit manually.
infra (private)
2026-06-25
If you made a $5 starter-pack purchase and then created additional API keys, those new keys could end up with the right tier but the wrong has_paid_successfully flag — leading to surprise 402s when you tried a premium model from a fresh key. Two fixes shipped: (1) the trial gate now bypasses entirely when has_paid_successfully=true on the calling key, regardless of tier — your access is governed by balance and rate limits, not by the 200-call trial cap; (2) new keys created after a successful payment now inherit has_paid_successfully from the user's existing key set at create time, so the inconsistency doesn't recur. Defense-in-depth: the bypass covers any pre-existing mis-tagged keys still in the DB; the inheritance prevents new ones from being created in the bad state. Caught via a new real-time observability signal (X-ALERT:claude_cli_blocked) that fires when a Claude Code / SDK user accumulates 3+ trial-block 402s in 5 minutes — surfacing the issue within an hour of the alert going live, on the first qualifying user.
infra (private)
Returns a one-shot snapshot of where your account stands: tier, trial { used, remaining, cap=200 }, balance in cents and USD, has_card, has_paid_successfully, and a savings_7d block with vs_gpt4o_mini_usd / vs_claude_sonnet_usd (computed from your free-NIM tokens routed in the last 7 days). Auth is via the same API key you use for /v1/chat/completions (Bearer or x-api-key) — no separate dashboard JWT. Lets SDKs and CLIs render upgrade nudges proactively ("150 of 200 free calls remaining, $0.64 saved vs Claude Sonnet so far") without scraping headers. Pairs with two response headers also added: X-InferAll-Trial-Usage: N/200 and X-InferAll-Trial-Status: active | halfway | near_cap, so per-call awareness comes for free without an extra round-trip.
infra (private)
The trial-block error messages were running 200–300 characters with the action URL embedded mid-sentence. In a terminal where the SDK renders the error inline (Claude Code, in particular), the URL often got truncated or buried — users would read the prose, not see a clickable URL, and bounce. Rewrote the four 402 surfaces (trial_blocked_no_card, checkCardRequired, trial_blocked_paid, trial_blocked_cap) to lead with the URL, then a short one-line explanation, then stop. Sample: "https://inferall.ai/billing — visit to activate 200 free NIM-model calls. No payment required, just initiating checkout unlocks the trial." Same gate, same flow, sharper terminal rendering.
infra (private)
2026-06-24
Every successful /v1/chat/completions response from a free-trial (tier=pending) key now includes the header `X-InferAll-Trial-Usage: N/200`, where N is the post-success count over the 200-call lifetime trial cap. SDK clients can surface "you've used 20 of 200 free calls" inline without an extra usage endpoint. Paid keys don't see the header (no relevant cap). Threaded through all four response paths — NVIDIA NIM and OpenAI passthrough, plus the Gemini- and Anthropic-translated paths — so the counter is consistent regardless of which provider answered the request. Closes a real friction case: heavy free users who pick NIM models that work well for them never bump into the upgrade gate and never realise there's a cap waiting. Same JSON body shape; OpenAI SDK clients keep working unchanged.
infra (private)
Pricing lookup was strict-match: if you sent model="nemotron-3-super-120b-a12b" without the "nvidia/" provider prefix, the gateway couldn't tell it was a free NIM model and the trial gate fired with a billing_required 402 — wrong outcome for a model that costs $0. Now the lookup falls back through common provider prefixes (nvidia, meta, deepseek-ai, google, anthropic, openai) when the bare name doesn't match directly. Affects every code path that depends on per-model pricing: trial allowance, free-model detection, cost reservation, and the inline-NIM-alternative suggestion in 402 copy. Bare gpt-oss-20b and gpt-oss-120b also resolve correctly now.
infra (private)
2026-06-23
Users hitting the trial gate on a paid model (e.g. gpt-4, claude-sonnet, gemini-2.5-flash) used to see a 402 that said "add $5 to unlock premium providers, your trial covers NIM." True but not actionable — they had to figure out which NIM model to swap to. Now the 402 message also says: "Or try a $0 NIM model right now — e.g. swap model=meta/llama-3.1-8b-instruct or meta/llama-3.1-70b-instruct (or any of the 118 NIM models) and your trial covers them today." Gives stuck users a one-line script edit to keep going without paying. The $5 pack CTA stays in place; X-Inferall-Upgrade-URL header stays. Within an hour of the deploy a Go-SDK user that had been retrying gemini-2.5-flash for 30+ attempts swapped to meta/llama-3.1-70b-instruct and made 22 successful calls.
infra (private)
2026-06-22
Telemetry expansion: every 402 from the gateway's trial-block / card-required / credits-required gates now carries a structured error_code in ai_usage_log plus an X-Inferall-Upgrade-URL header for SDK/IDE retry loggers that don't parse JSON bodies. The trial-block path previously emitted 402s with no log entry at all — that entire cohort is now visible. Closed two billing bugs: (1) finalize_spend was debiting tier=pending trial users' balance on every free-NIM call because the MINIMUM_FREE_MODEL_COST_CENTS floor flowed into estimated_cost_cents; gate finalize_spend on real provider cost instead. (2) Stripe webhook endpoint was only subscribed to 4 of 9 handled events — patched to all 10 so charge.succeeded / charge.refunded / checkout.session.expired actually fire in production. checkout-expired events now also persist to billing_errors so dormant-checkout outreach can be queried programmatically. Added a per-user gate-storm tracker that emits [ALERT:gate_storm] when one user hits 15 trial-block 402s in 60s — surfaces SDK auto-retry loops in real time instead of via hourly sweeps.
infra (private) · site (private)
2026-06-15
Shipped a 21-commit run touching billing race conditions (the mode=payment fix that unblocked our first paid customer, plus the mirror fix for mode=subscription), Stripe checkout improvements (abandoned-cart recovery, allow_promotion_codes, product description on the Stripe page, Stripe Customer idempotency keys to defeat double-click races), a charge.refunded handler so refunds actually subtract from balance, the compat /v1/messages fallback engine now short-circuits on non-recoverable upstream errors instead of cascading the whole chain, and a full timeout sweep — every outbound fetch in the gateway codebase (Supabase REST, Stripe API, and all five provider adapters: minimax/replicate/runway/stability/elevenlabs) now uses fetchWithTimeout with explicit budgets so a hung upstream socket can no longer pin a worker forever.
infra (private)
Earlier entries in this changelog (and across the marketing site) described the Free tier as "100,000 tokens/month" — that was never the actual gateway behavior. The real model: open NVIDIA NIM models bill at $0 input/$0 output regardless of tier, premium providers (OpenAI, Anthropic, Google) bill at the provider's published per-token rate with zero markup [corrected 2026-09-02], and every tier has a daily request count limit per operation (Free: 50 chat/100 text per day; Pro: 5,000 chat; Team and Enterprise scale up). Quotas reset daily UTC. Marketing copy across the homepage, /billing, /keys, /solutions/*, /use-cases/*, /extension, and the major blog posts was rewritten to describe this honestly. Old entries below are left as the historical record but the framing they describe is superseded by this one.
site (private)
2026-06-13
Added a $5 one-time activation pack as the entry pricing tier alongside Pro ($29/mo). The $5 becomes spendable balance: open NVIDIA NIM models bill at $0 input/$0 output against your balance, premium providers (OpenAI, Anthropic, Google) bill at the provider's published rate with zero markup [corrected 2026-09-02]. Real card + real money cleared by Stripe is the bot defense — bots don't burn $5 on a card they can't extract value from, and real evaluators get to start using the gateway without committing to $29/mo. Uses inline Stripe priceData (same pattern as the existing $100 credit pack) so no preconfigured Stripe Price ID was needed.
infra (private)
2026-06-12
Removed "no credit card required" / "no card needed to start" claims across the marketing site (hero, /pricing, /keys, /billing, /solutions/*, /use-cases/*, /integrations). The free tier still costs $0 within the 100,000-token monthly allowance on 80+ NVIDIA NIM open-source models, but activation requires a verified card on file — that's the gate that's holding back the bot signup wave that started in early June, and reversing it would re-open ~$7k/mo in upstream Anthropic outflow. The site now says so up-front. Pending-tier (no card yet) keys get a small trial allowance to evaluate before the activation prompt.
infra (private)
2026-06-03
ElevenLabs text-to-speech is now live as a gateway provider. POST /ai/v1/generate with provider: "elevenlabs", operation: "tts" — returns binary audio/mpeg. Supports eleven_multilingual_v2 and named voice aliases (confident_narrator, warm_narrator, friendly_narrator, deep_narrator) or raw ElevenLabs voice UUIDs. Same ifu_ key as all other providers.
infra (private)
Added the full GPT-4.1 family (gpt-4.1 at $2/$8/M, gpt-4.1-mini at $0.40/$1.60/M, gpt-4.1-nano at $0.10/$0.40/M), OpenAI reasoning models o3 ($10/$40/M), o4-mini ($1.10/$4.40/M), and o1-pro ($150/$600/M). Also added five Anthropic Claude 4 variants that were live on Anthropic's API but absent from our catalog: claude-opus-4-1, claude-opus-4-5, claude-opus-4-6, claude-opus-4-7, and claude-sonnet-4-5. Fixed a routing bug where o4-mini requests didn't route to OpenAI. Catalog is now 180+ models, 0 dead.
infra (private)
The /keys page now shows a persistent copy-paste curl snippet for pending-tier users who haven't made an API call yet — so the first-call path is visible even after you close and reopen the tab. Previously the snippet only appeared immediately after creating a new key; returning users saw their key prefix but no guidance. Also added a copy button and telemetry to track the never-activated cohort.
Fixed a bug where the trial progress bar (N/200 requests) was invisible to pending-tier users if the /ai/v1/usage fetch returned a non-401 error. The bar now defaults to showing 0/200 rather than hiding entirely, so users near the cap always see a warning.
PR #85 (private)
The daily GitHub Actions cron that reports usage to Stripe for metered billing was silently failing due to three missing secrets. Set and verified; first successful run confirmed. Usage-based subscriptions are now tracked correctly.
infra (private)
2026-05-30
You no longer need a card on file to try InferAll. Create a key and start calling the 110+ free NVIDIA NIM open-source models (Llama 3.1 70B/8B, Mixtral, Nemotron, CodeLlama, and more) right away — $0. A card is needed only when you want a paid provider (OpenAI, Anthropic, Google) or to continue past the free trial. Same ifu_ key, same OpenAI- and Anthropic-compatible endpoints — the card wall that used to sit in front of the free tier is gone.
infra (private)
2026-05-29
The gateway now monitors its whole catalog. A daily check (and a public GET /ai/v1/models/health endpoint) compares every model against its provider's live model list and flags any the provider has stopped serving — so when a provider deprecates, renames, or rotates a model ID (the drift that retired DALL·E, sunset Gemini 1.5, and regularly rotates NVIDIA NIM IDs) it gets caught and fixed instead of 404'ing your calls. Detection is registry-based and $0, so monitoring never burns the free tier it's watching.
infra (private)
If your app is still pinned to a deprecated Claude snapshot the provider has retired, the gateway now transparently routes it to the current equivalent and labels the response with the model that actually served it — so you keep working while you migrate instead of failing during an incident. We also removed catalog IDs providers no longer serve (the catalog now lists only models that actually run), repointed a stale fallback hop, and moved OpenAI image generation to the current gpt-image-1 (DALL·E was retired upstream) at the same price.
infra (private)
Raised the upstream timeout for the primary provider attempt so large, long-running generations complete instead of being aborted at 30 seconds — while keeping fallback hops short so a real provider outage still fails over quickly. Fixes a class of large requests that were erroring after the primary and every fallback hit the old 30s ceiling.
infra (private)
api.inferall.ai now runs on two machines across two regions (sjc + iad) with HTTP health checks, so a single machine or a regional edge incident no longer takes the gateway down — the proxy routes around an unhealthy instance. Per-machine memory was raised 256MB → 512MB to remove GC stalls under streaming load. Resolves the intermittent connection resets observed during a Fly SJC edge incident on 2026-05-28.
infra fly.toml (private)
Added GPT-5.5, GPT-5.5 Pro, GPT-5.4 (+ mini/nano), GPT-5, and Claude Opus 4.8 / Sonnet 4.6 to the model catalog (GET /ai/v1/models). Premium tokens bill at the provider's published price with zero markup [corrected 2026-09-02], on your existing ifu_ key.
gateway change (private)
After you create a key, the quickstart now offers a copy-paste curl that makes a $0 call on a free NVIDIA NIM open model via the OpenAI-compatible endpoint — alongside the Claude Code snippet — so you can confirm your key works in one paste, with no premium spend.
Fixed two broken model IDs in the OpenRouter comparison and Claude Code use-case pages (they named models not in the catalog, so copy-paste examples 404'd), and added a CI check that fails if any blog / comparison / use-case / docs page references a model ID not in the live catalog. Prevents the bug class going forward.
2026-05-26
Both 429 responses from the gateway's limiter — spending-limit-exceeded and daily-rate-limit-exceeded — now lead with https://inferall.ai/billing as the primary upgrade CTA. Previously the spending-limit message only pointed at the Stripe portal (meant for existing subs adjusting card cap), and the daily-rate-limit message had no upgrade pointer at all.
gateway change (private)
Replaced the generic "Select" buttons on /billing with tier-specific labels ("Upgrade to Pro", "Upgrade to Team", "Get free API key" → /keys, "Contact sales"). The current tier gets a left-border accent and a disabled "Current plan" label. Pro carries a "· Most popular" badge. Post-checkout, ?success=true now triggers a "Thanks — your subscription is active" banner.
Single shell script anyone with the Stripe CLI authed can run to refresh founder metrics: customer count, signup velocity, active subscriptions + plan breakdown, paid invoices, churn, void invoices, funnel breakdown. Outputs human-readable markdown + a JSON snapshot. Aggregate-only — no customer IDs, emails, or payment details written to disk.
PostHog wired to the marketing site (project 359601) — pageviews on soft navigation, plus explicit funnel events (inferall_auth_success with identify, inferall_key_created with tier, inferall_first_call_snippet_copied). Removed a fabricated aggregateRating from JSON-LD (no real ratings exist; honesty + Google structured-data policy). /live now acknowledges that the gateway passes any Replicate or NIM model id through directly, so the routable catalog is much larger than the 207 enumerated.
SECURITY.md and all nine Code-of-Conduct translations (en + ja, ar-sa, ko, es, zh-tw, pt-BR, zh-cn) now route security reports and abuse reports to the InferAll team instead of the upstream Cline fork's contact addresses.
2026-05-25
Python (inferall-ai) and TypeScript (@inferall/sdk) hello-world scripts plus a three-path README (Claude Code env vars, Python SDK, TypeScript SDK). Linked from the docs Quick-start.
Server-rendered with 1h ISR. Shows total / free / paid counts, the table of every vendor (google, nvidia, openai, meta, mistralai, anthropic, runway, microsoft, qwen, …), and live vs static source mix. Counts match what `curl /ai/v1/models | jq 'keys|length'` returns.
When you create a new key, the page now embeds a two-env-var snippet with the actual key inline and its own Copy button. Time from key → working `claude` command is one paste.
Homepage hero now shows a one-line live readout: total models, free models, gateway healthy/degraded — fetched server-side from the public endpoints. /status replaces its previous placeholder with the same data while the full multi-region status page is built out.
Dashboard-created user keys now mint with the InferAll-branded `ifu_` prefix instead of the legacy `kr_user_`. Docs, SDK examples, /compare snippets, and the security page were updated to match. Existing keys remain valid forever.
gateway change (private) · PR #28
Site now says "200+ models" / "150+ free open-source models" / "120+ NVIDIA-hosted models" — each rounded down from what /ai/v1/models actually enumerates today (207 / 158 / 123). Survives catalog fluctuation without overclaiming.
Removed every "no credit card" claim. The Free tier still costs $0 within the 100,000-token monthly allowance on NVIDIA NIM open-source models, but activation requires a card on file — we say so upfront across the marketing site and on /keys.
Open Graph + Twitter image, generated favicon, web manifest, sitemap expansion, /docs metadata. Link previews on Hacker News, X, Reddit, Slack now render correctly.
2026-05-24
The native TypeScript SDK is live on npm. `npm install @inferall/sdk`. Reads `INFERALL_API_KEY` from the environment; same `Inferall` class shape as the Python SDK.
Concrete SLA at /sla: 99.9% gateway control-plane uptime per calendar month, tiered service credits (10% / 25% / 50% of monthly fee), exclusions, failover behavior, and incident response.
2026-05-23
Pricing tiles on /billing now reflect the real Team tier (10M tokens/month) and add an Enterprise contact-sales tile linking to contact@kindly.fyi.
/docs SDK examples now show the actual published Python (`inferall-ai`) and TypeScript (`@inferall/sdk`) shapes, with `ifa_…`-prefixed keys.
Homepage hero leads with the published Python SDK install + a minimal usage snippet.
Earlier
Server-side retry against the next capable provider in the chain when an upstream rate-limits or errors. Context-size-aware, cheapest-first selection. A single Anthropic outage or Google quota reset no longer surfaces to your application.
gateway change (private)