Product updates

Changelog

New features, behavior changes, model additions, and deprecations. Subscribe via RSS.

Feature

Guardrails now check MCP tool results

Every MCP tool result is now checked before the model sees it, with the same rules as your prompts: library rules set to inbound or both (prompt injection, secret detection, PII redaction) and your custom patterns. A blocked result is replaced with a short note so the run carries on without it; redact rules mask matches; log-only rules record the hit. Hits show on each tool call in Agent runs and in the x-gateway-mcp-tool-results-guarded and x-gateway-mcp-tool-result-rules headers. If a rule you already run on prompts is set to block, tool results that match it will now be withheld.

MCP docs →

Feature

Disable an API key without revoking it

Owners and admins can now pause an API key from the API keys page. A disabled key is refused with 403 key_disabled straight away, and its value, limits and MCP servers are kept, so enabling it again restores the agent as it was. Revoking is still permanent. Both show up in the audit log, and the platform API's key list now includes disabled_at.

Feature

Audit log of every configuration change

Owners and admins can now see who changed what in the organization: API and admin keys, spend caps and rate limits, provider keys, guardrails, the MCP allowlist, routing, the team and billing mode, with the time, the IP address and the values before and after. Filter by area and time on the new Audit log page, or export it as CSV. Secrets are never recorded. Changes are recorded from today.

Audit log docs →

Update

Disable an MCP server without losing per-key settings

MCP servers on your allowlist can now be disabled and enabled again. A disabled server is refused for every key, and keys limited to specific servers keep it selected for when you turn it back on. Removing a server still takes it off those keys too, and the dashboard now warns before it does. Adding a URL that's already listed updates the entry instead of replacing it.

Feature

Limit MCP servers per API key

Each API key can now use either every MCP server on your organization's allowlist or only the ones you pick, so each agent reaches only its own tools. A restricted key naming another server gets 403 mcp_server_not_allowed_for_key before the server is contacted, and GET /v1/mcp/servers returns only that key's servers. Set it under a key's limits on the API keys page.

MCP docs →

Feature

Run budgets: cap what one agent run can cost

Set Max $ per run on an API key and no single agent run on it can cost more. Agents can lower the budget for a run with the X-Agent-Run-Budget header, never raise it. Before each call, including each round of an MCP tool loop, the gateway checks the run's spend plus the most the call could cost, and refuses calls that could go over with HTTP 402 run_budget_exceeded and the largest max_tokens that still fits. Runs that hit their budget are marked on the Agent runs page.

Run budgets docs →

Feature

langchain-leanroute for Python is on PyPI

pip install langchain-leanroute gives you ChatLeanroute: LangChain's ChatOpenAI with the Leanroute base URL built in, the key read from LEANROUTE_API_KEY, and agent_name / session_id arguments that group LangGraph calls into agent runs. Tools, streaming, structured output and async work unchanged.

LangChain guide →

Feature

One-click first API call on the dashboard

New accounts now see their first call happen on the dashboard: Leanroute creates an API key, runs one free request and shows the answer, then gives you the same call as curl, Python and Node with your key filled in. If you already have keys, one button does the same with a fresh key.

Dashboard →

Update

Guides: agent costs per run, and an agent governance checklist

Two new guides. How to track AI agent costs per run, with LangGraph and n8n examples. And an AI agent governance checklist: 50 controls across inventory, access, data, tools, spend, audit and incidents, marking which ones a gateway can enforce.

Agent governance checklist →

Update

Credits: top up from $10, fees shown up front

The minimum top-up is now $10 of credit (was $50). The amount you pick is the credit you get, in full. A 2% platform fee and card processing (4.4% of the total plus $0.40, passed through) are added at checkout and shown before you pay: $50 of credit costs $53.77. Before this change, 5% was taken out of each top-up. Token prices are unchanged, with no per-request markup.

Billing →

Update

BYOK plans: free requests work before you add a provider key

On a BYOK plan, requests to a provider you haven't added a key for now use your 100 free requests on low-cost models (such as deepseek/deepseek-v4-1-flash), instead of returning provider_not_configured. Premium models still need your own key.

Add a provider key →

Model

New models: Gemini 4 Argon, Grok 4.7, Kimi K3.1 Pro, GPT-6.1 Terra, GLM 5.3 FlashX

Added google/gemini-4-argon ($2 / $10 per 1M tokens until 30 November, $4 / $20 from 1 December, as Google's introductory price ends), xai/grok-4-7 ($2 / $6, 2x above 200K tokens), kimi/kimi-k3.1-pro ($2 / $10), openai/gpt-6.1-terra ($1 / $5) and glm/glm-5.3-flashx ($0.37 / $1.25, BYOK). Also: openai/gpt-5.6-sol is billed at OpenAI's promotional $4 / $20 until 21 November and at its list price of $5 / $30 after that, unless OpenAI extends the promotion.

Model catalog →

Update

Easier first steps: test request button and sturdier sign-in links

After you create an API key, the keys page now has a Send a test request button that runs one free request with the new key and shows the answer, plus a ready-to-copy curl command. Sign-in links now open a confirm page where you press Sign in, so email security scanners can no longer use up your link before you do; links last 15 minutes instead of 5, and an expired or used link now says so instead of returning you to the sign-in form. Opening api.leanroute.dev in a browser shows how to send your first request instead of a 404.

Setup guide →

Update

Provider rejections now return 400 with the provider's reason

When a provider refuses a request (for example, a parameter the model doesn't support), the gateway now returns the provider's 400, 413 or 422 instead of 502, and the error message quotes the provider's reason. Provider outages, rejected provider keys and unknown models still return 502, and provider rate limits still return 429. If your code retries on 502, these rejections are no longer retried, which is the right outcome since retrying would fail the same way. On /anthropic/v1/messages they now arrive as invalid_request_error.

Error codes →

Feature

Agent runs: cost, models and tools per agent run

Send X-Agent-Name and X-Agent-Session-Id with your agent's requests and the new Agent runs dashboard page groups them into runs: runs per agent, cost per run, models used, and every tool called, whether the gateway ran it over MCP or the model asked your agent to run it. Tool arguments and results are never stored. The same data is available from GET /platform/v1/runs and the list_runs and get_run MCP tools. Agents can also discover the MCP servers they may use with GET /v1/mcp/servers. Streamed requests from Claude Code and other agents now count toward their session too.

Agent runs docs →

Model

Claude 5.5 family, GPT-6 Luna and GPT-6.1 Sol

Five new models: anthropic/claude-haiku-5-5 ($0.10 input / $0.50 output per 1M tokens; prompts of 100K tokens or more bill the whole request at 5x), anthropic/claude-sonnet-5-5 ($2 / $10), anthropic/claude-opus-5-5 ($4 / $20), openai/gpt-6-luna ($0.10 / $0.50; prompts of 272K tokens or more bill at $0.20 / $0.75) and openai/gpt-6.1-sol ($2 / $10). Claude Haiku 5.5 and GPT-6 Luna are low-cost enough to count as free-request models. Billing change: openai/gpt-5.6-sol is billed at OpenAI's promotional $4 / $20 until 21 November 2026, then at its $5 / $30 list price unless OpenAI extends the promotion.

Update

Embeddings: unknown model ids now return a 400

/v1/embeddings now serves only the embedding models listed in /v1/models (today openai/text-embedding-3-small, openai/text-embedding-3-large, openai/text-embedding-ada-002 and google/gemini-embedding-001). Any other id returns a 400 unknown_model that lists the available ones. Before, an unlisted openai/ or google/ id was passed to the provider and not billed. Prices for the listed models are unchanged.

Update

Sarvam 105B is no longer a free-request model

sarvam/sarvam-105b and sarvam/sarvam-105b-conversations think before answering, often for longer than the 1,024-token output cap on free requests, which left free callers with an empty-answer error. They are no longer in the free-request list; paid and BYOK requests are unchanged. qwen/qwen2.5-vl-72b-instruct is no longer available on our Alibaba region, so requests for it are now served by deepseek/deepseek-v4-1-flash, which supports images, and billed at its rate: $0.15 input / $0.60 output per 1M tokens off-peak, double in DeepSeek's weekday peak hours (was $0.80 / $1.00).

Removed

Claude Sonnet 4.5 routes to Sonnet 5; text-embedding-004 removed

Anthropic retires Claude Sonnet 4.5 on 30 November 2026. Requests for anthropic/claude-sonnet-4-5 are now served by anthropic/claude-sonnet-5 and billed at its rate ($2 input / $10 output per 1M tokens, down from $3 / $15). bedrock/anthropic.claude-sonnet-4-5 is unchanged. Google shut text-embedding-004 down in January, so requests for it now return a 400 that points to google/gemini-embedding-001. We don't swap it automatically because the new model returns vectors of a different size, which would break existing indexes.

Update

Price corrections: Gemini 3.5 Flash, Sarvam 105B and five Qwen models

Google retired Gemini 3.5 Flash and 3.7 Flash on 8 October and now serves them with 3.6 Flash and 3.8 Flash. Requests for those ids (and for google/gemini-2.5-pro and google/gemini-3.1-pro, which routed to 3.5 Flash) are now billed at the 3.6 / 3.8 Flash rate of $0.75 input / $3.75 output per 1M tokens, down from $1.50 / $9.00. Sarvam cut sarvam-105b and sarvam-105b-conversations on 10 October: now $0.156 input / $0.625 output (was $0.349 / $0.871). Qwen rates now match Alibaba's International list: qwen-plus rises to $0.40 / $1.20 (was $0.18 / $0.42) and is no longer one of the free-request models, qwen3.7-max rises to $2.50 / $7.50 (was $1.65 / $4.95), qwen3.8-flash rises to $0.15 / $0.47 (was $0.14 / $0.42), qwen3.8-max output falls to $6.00 (was $6.40), and qwen3.5-flash falls to $0.10 / $0.40 (was $0.115 / $1.147). qwen/qwen-vl-max no longer exists upstream; requests for it are now served by deepseek/deepseek-v4-1-flash, which supports images, instead of failing.

Update

Claude 4.7 and newer no longer fail when you send temperature

Anthropic rejects temperature and top_p on Claude 4.7 and later models (Opus 4.7, Opus 4.8, Opus 5, Sonnet 5, Fable 5 and 5.1) with HTTP 400. Many clients send temperature by default, including the n8n node, so those requests failed. The gateway now drops both parameters for these models instead of passing them on. Earlier Claude models are unchanged.

Removed

openai/o1 now routes to gpt-5.6-sol

OpenAI shuts down o1 on 23 October 2026. Requests for openai/o1 are now served by openai/gpt-5.6-sol, OpenAI's named replacement, and billed at its current rate ($4 input / $20 output per 1M tokens today, lower than o1). Three retired Krutrim model ids now route straight to their gpt-oss replacements instead of through another retired id.

Update

Semantic cache only returns answers from the model you asked for

A semantic cache hit now requires the cached answer to come from the same model as your request. Previously a similar prompt sent to a different model could be answered from another model's cached response. Exact-match caching was already scoped per model.

Feature

100 free requests for every new account, no card needed

New accounts can now send their first requests straight away: 100 free requests on fast, low-cost models such as DeepSeek V4.1 Flash, Qwen Flash, Gemini 3.1 Flash-Lite and gpt-oss on Groq. No card, no provider keys. Free requests are limited to 1,024 output tokens and about 8,000 tokens of input, and don't cover MCP tools or other models; those return a 402 listing the models that are free. The dashboard shows how many you have left, and responses carry x-gateway-free-requests-remaining. Answers served from the cache count as a free request; failed requests don't. Existing accounts with no balance get the 100 as well.

Get started →

Feature

Leanroute community node for n8n

Self-hosted n8n users can now install n8n-nodes-leanroute from Settings, Community Nodes. It adds a Leanroute Chat Model you can plug into an AI Agent, plus a Leanroute node for single calls, using your Leanroute API key and any model in the catalog. On n8n Cloud, keep using the built-in OpenAI node with a custom base URL.

n8n guide →

Feature

Setup checklist and guide, plus model, cost and latency on every response

After your first request, the dashboard now shows a short checklist for spend caps, guardrails, routing, caching and MCP tools, with each item ticked off from your actual settings and linked to the new setup guide. Non-streaming responses now also carry x-gateway-model, x-gateway-cost-usd and x-gateway-latency-ms, so you can see which model answered and what it cost without opening the dashboard. Streamed responses include x-gateway-model; their cost is in your request logs.

Setup guide →

Update

Guardrails now check model responses, and the moderation rule works from the library

Library rules set to outbound or both (data exfiltration, profanity, language gate, and PII redaction and secret detection by default) now check the model's answer on non-streaming requests: redact rules redact it, and block rules withhold it with HTTP 422 response_blocked. A blocked answer is still billed, because the model generated it. Streamed answers are not checked, since they reach your client as they are generated. If you already had one of these rules enabled, you may start seeing redactions or 422s; use preview mode to see what a rule would do first. Enabling the moderation rule in the guardrails library now turns the moderation pre-check on, and respects that rule's action and preview mode.

Guardrails →

Update

Requests must set max_tokens when credit balance is under $5

When your prepaid credit balance is under $5, requests served on Leanroute's provider keys must include max_tokens, and the estimated cost of the prompt plus max_tokens at the model's prices must fit within your balance. Requests that don't are rejected with 402 low_balance_max_tokens; the error includes max_affordable_tokens for the model you called. Above $5 nothing changes. Requests on your own provider keys (BYOK) are not affected. Keeping your balance above $5, or always sending max_tokens, avoids this.

Billing →

Update

Streamed requests are metered even when the stream ends early

Streaming requests are now recorded and billed for the tokens actually generated when the stream ends early, whether your client disconnects or the provider stops mid-response. Previously some of these requests were not recorded in usage or billed. Disconnecting does not stop the provider from generating, so a request you abandon is billed for the full response the provider produces. Set max_tokens to bound this. When a provider does not report token counts for a stream, usage is estimated from the streamed text and the request is marked usage_estimated in your logs. A stream cut off by the provider is marked stream_interrupted.

Streaming docs →

Update

Cache retention enforced at 15 minutes; Wipe cache now clears both caches

Semantic cache entries now carry a storage-level 15-minute expiry, so they are removed on schedule even when an organization stops sending traffic. The Wipe cache button in Settings now removes both your exact-match and semantic cache entries immediately.

Caching docs →

Update

Spend caps and token-rate limits now cover BYOK traffic

Daily and monthly spend caps, and per-minute token limits, now count requests made on your own provider keys, at provider list price. If you set a cap to protect your own provider bill, it now stops BYOK traffic once reached (402 spend_cap_exceeded), the same as managed traffic. If you have caps set and run BYOK, check they are where you want them.

Error reference →

Update

BYOK: cross-provider failover on your own keys

On a BYOK plan, if a provider returns a 5xx, the gateway now retries on a same-tier model from another provider you hold a key for, exactly as it does on managed keys. It never falls back to a provider you have no key for. Send X-Gateway-Routing: explicit to turn it off per request. Requests for a provider you have no key for now also list the models you can call in the error (callable_models).

Error reference →

Feature

BYOK free trial, no card required

The 14-day BYOK trial no longer asks for a card. Sign up, start the trial, add one of your own provider keys and send real requests right away. Add a card from Billing before the trial ends to keep your plan; if you don't, the trial simply ends and nothing is charged. One trial per workspace.

Pricing →

Update

Clearer setup errors and a dashboard setup banner

When a request routes to a provider you have no key for, the error now explains why and how to fix it. The new reason field is plan_byok_only (your BYOK plan uses your own key for every provider you call) or provider_byok_only (the provider is BYOK-only for everyone), with a dashboard_url to add the key. These now return **400** instead of 503, so SDKs stop auto-retrying a request that can't succeed until the key is added. 503 remains only for no_managed_key, when our own managed key is missing. If a retired model was rewritten to a different provider, requested_model shows what you originally asked for. The dashboard also shows a banner on every page while a setup step is blocking all requests.

Error reference →

Model

Sarvam 105B Conversations

sarvam/sarvam-105b-conversations is Sarvam's variant of 105B tuned for real-time dialogue and voice agents. Same rate card as sarvam-105b. Sarvam 30B remains retired; requests for it continue to be served by sarvam-105b with an x-gateway-model-deprecated header, as they have been since 31 July.

Models →

Update

Pricing API: new replaced_by field for retired models

A request for a retired model id is served and billed as its replacement. GET /platform/v1/pricing and the MCP get_pricing tool now show that directly: retired rows quote the replacement's rates and carry a new replaced_by field. The MCP cheapest_for tool recommends live models only, and never a specialised OCR or translation-only model for a chat task. Billing is unchanged.

Update

MCP requests always run their tools fresh

Requests carrying mcp_servers now skip both the exact-match and semantic caches, for reads and writes. The answer to an MCP request depends on what your tools return at that moment, so every MCP request now runs its tools live. Requests without mcp_servers are cached as before.

MCP docs →

Update

Guardrail library: seven built-in rules on every plan

Seven rules ship in the guardrail library and are included in every plan rather than sold as a plugin or a per-request add-on: PII redaction, secret and credential detection, prompt-injection detection, OpenAI moderation pre-check, data-exfiltration detection, profanity filtering, and a language/region gate. Each can redact rather than only block, which is usually what you actually want. Your own per-org patterns still layer on top.

Guardrails →

Update

Clear error when a reasoning model runs out of tokens

Reasoning models count their internal thinking against max_tokens. If the budget is too small, the model can use all of it thinking and return no answer. The gateway now returns 400 empty_completion in that case, naming the model, the finish reason and the output tokens billed, so you know to raise max_tokens. Tool-call turns, where empty content alongside tool_calls is normal, pass through as before.

Update

DeepSeek peak-hour rates, GPT-6 Astra parameters, long-context tiers

**DeepSeek:** request costs now apply DeepSeek's weekday peak-hour rate (01:00-04:00 and 06:00-10:00 UTC, double the off-peak price), matching what DeepSeek invoices. Requests in those windows show a higher cost than before. **GPT-6 Astra:** max_tokens is now sent as max_completion_tokens, as OpenAI requires for this model. **Long context:** Gemini 3.1 Pro and xAI models now use their long-context rates for prompts over 200k tokens.

Feature

Point Claude Code at Leanroute with one environment variable

Leanroute now speaks the Anthropic Messages API at /anthropic/v1/messages, so anything built on the official Anthropic SDKs, Claude Code included, can route through the gateway without a code change: `` export ANTHROPIC_BASE_URL="https://api.leanroute.dev/anthropic" export ANTHROPIC_API_KEY="gw_live_..." export CLAUDE_CODE_ATTRIBUTION_HEADER=0 ` This is a translation layer, not a passthrough. Requests are converted into the gateway's internal format, which means caching, spend caps, guardrails, failover, per-key limits and full request logging all apply to your Claude Code traffic. It also means you can pin a cheaper model: set ANTHROPIC_MODEL="deepseek/deepseek-v4-1-flash" and Claude Code keeps working while the bill drops roughly tenfold. Streaming, tool use, images and multi-turn tool round trips are all supported. **Routing here defaults to explicit**, unlike /v1/chat/completions. The model you ask for is the model you get. Coding sessions are long, stateful and full of tool calls, and substituting a model at turn 12 of a refactor does not degrade gracefully, so automatic cost arbitrage is opt-in on this endpoint via ANTHROPIC_CUSTOM_HEADERS="X-Gateway-Routing: auto"`. Requests also carry Claude Code's session and subagent identifiers into the request log, so spend can be attributed to an individual coding session rather than only to a day. Built against Anthropic's published gateway compatibility guide, including keep-alive pings during long thinking pauses and verbatim upstream error forwarding so Claude Code's own retry recovery keeps working.

Claude Code setup guide →

Feature

MCP tools now work on every model, not just Claude

Leanroute now runs the MCP agent loop itself. Send mcp_servers on a non-streaming /v1/chat/completions request and the gateway connects to your MCP servers, discovers their tools, namespaces them, injects them as ordinary function tools, executes the calls, and loops until the model is done. Because MCP tools become plain function declarations on the way out, **any provider with function calling now supports MCP through Leanroute** — GPT, Gemini, Groq, DeepSeek, and the rest, not only Anthropic. Every tool call is logged with its server, duration, and payload sizes, and token usage is aggregated across all hops so you are billed for the whole run rather than the last turn. Response headers report the tools discovered, tool calls made, iterations used, and any MCP server that failed to connect. MCP requests run non-streaming.

Read the MCP docs →

Update

MCP: one gateway-run path for every provider

The earlier Claude-only MCP path, which handed mcp_servers to Anthropic, has been replaced by the gateway-run loop above, so every tool call is now logged, costed and covered by guardrails. Three changes you may notice. **One:** mcp_servers with stream: true now returns 400 mcp_streaming_unsupported on every provider including Anthropic, where it previously streamed. Set stream: false and the gateway runs the loop for you. **Two:** the x-gateway-mcp-passthrough response header is gone; it had become a duplicate of x-gateway-mcp-allowlist, and x-gateway-mcp-mode is now always proxy. **Three:** the error code mcp_unsupported_provider_streaming is replaced by mcp_streaming_unsupported, since the restriction is no longer provider-specific.

MCP documentation →

Update

Tool calling: full JSON Schema on Gemini, multi-turn Gemini 3, OpenAI reasoning models

**Gemini tool schemas:** Gemini accepts only a subset of JSON Schema. Schemas using additionalProperties, $ref, allOf or const are now converted to what Gemini accepts (const becomes a single-value enum), so the same tool definitions work on Gemini, OpenAI and Anthropic. Schemas that are already valid pass through unchanged. **Gemini 3 multi-turn tool use:** Gemini 3 requires each function call's thoughtSignature to be echoed back on the next turn. The gateway now preserves it and exposes it as tool_calls[].thought_signature in streaming and non-streaming responses, so you can round-trip it in your own agent loop. **OpenAI reasoning models with tools:** for gpt-5*, o1* and o3*, the gateway sets reasoning_effort: "none" when you send tools without setting it yourself, which OpenAI requires for tool calls on these models. If you set reasoning_effort explicitly, your value is sent as-is.

Model

Added Claude Fable 5.1 and DeepSeek V4.1 Flash

Weekly catalog sweep. Two new SKUs added: **Claude Fable 5.1** (2026-09-01, $10/$50, cache reads dropped to $0.25 from $1.00 — a 75% cache-hit reduction on the same headline price) and **DeepSeek V4.1 Flash** (2026-09-10, $0.15/$0.60 with native multimodal, replaces V4 Flash and V4 Flash Vision Exp). DeepSeek's V4.1 launch also drops customer bills across the whole DeepSeek Flash line: old slugs (deepseek-chat, deepseek-v4-flash, deepseek-v4-flash-vision-exp) now auto-route to V4.1 Flash for a ~30% input / ~9% output price cut plus proper multimodal support (no more 384-token per-image cap). GPT-6 Astra's catalog entry now notes its 272K context limit.

Browse the model catalog →

Feature

Open-sourced opa-sidecar-a2a: chain-aware authorization reference implementation

Published opa-sidecar-a2a v0.1.0-alpha under MIT at github.com/leanroute/opa-sidecar-a2a. Single Go binary + OPA policy engine + reference Rego policies + worked planner-executor demo. Not a Leanroute-only feature; anyone building an A2A stack can adopt it. The announcement post covers the why and the code.

Read the announce →

Update

xAI Grok: model ids mapped to xAI's current naming

xAI's own model ids use dots (grok-build-0.1, grok-4.6, grok-4.20-*). Leanroute keeps its stable hyphenated ids (xai/grok-build-0-1, xai/grok-4-6) and now maps them to xAI's naming on the way out, across grok-build-0-1 and the full grok-4.x family. No change needed on your side.

See the model catalog →

Model

`openai/gpt-6-astra` added (new OpenAI flagship)

OpenAI's new flagship above the GPT-5.6 line, launched 2026-09-03. Premium-reasoning tier at $10 in / $50 out per 1M tokens — matches Anthropic Fable 5's $-band. Cache-hit assumed 90% off; verify on first invoice.

Model

`google/gemini-3.8-flash` added

Google's newest Flash, launched 2026-09-02. Same $0.75 / $3.75 intro rate as 3.6 and 3.7 Flash; all three revert to $1.50 / $7.50 on 2027-01-01. Recommended default for new integrations per Google.

Model

`google/gemini-3.1-pro` added

Google's hard-reasoning / agentic-coding Pro SKU at $2 in / $12 out per 1M tokens. Flagship tier — competes with anthropic/claude-sonnet-5 and openai/gpt-5.6-terra in the $2-band contest.

Model

`qwen/qwen3.8-flash` added (cheapest Qwen tier)

Alibaba's newest cheap_fast SKU, launched 2026-08-26. $0.14 in / $0.42 out per 1M tokens. Sits below qwen3.8-max ($2/$6.4) in the same generation.

Removed

`kimi/kimi-k2.5` retired (Moonshot 2026-08-31 sunset)

Moonshot retired K2.5 alongside the Moonshot V1 series on 2026-08-31. Requests now auto-rewrite to kimi/kimi-k2.6 (closest live general-purpose replacement, 256k context). Bill impact: K2.5 was $0.60 / $3.00, K2.6 is $0.95 / $4.00 — modest rate bump.

Update

OpenAI catalog cleanup: gpt-4o, o1, gpt-5-nano, gpt-5-mini deprecated

OpenAI dropped gpt-4o, gpt-4o-mini, o1, gpt-5-nano, and gpt-5-mini from their public pricing page. The current catalog is gpt-5.6-sol / terra / luna / cyber + gpt-5.3-codex. Requests to the retired ids now auto-rewrite: gpt-4o → gpt-5.6-terra, gpt-4o-mini → gpt-5.6-luna, o1 → gpt-5.6-cyber, gpt-5-nano → gpt-5.6-luna, gpt-5-mini → gpt-5.6-luna. Watch for bill spikes: the old nano at $0.05/$0.40 was 3-4× cheaper than luna at $0.20/$1.20 — cost-sensitive workloads should switch to deepseek/deepseek-v4-flash ($0.22/$0.66).

Model

`glm/glm-5.3` added (Z.AI flagship)

Z.AI's latest flagship, $1.40 in / $4.40 out, 81% cache-hit discount. Joins the 5.1/5.2 cluster at the same price point — 5.1/5.2 remain in the catalog for version pinning.

Model

`glm/glm-5.3-flash` added (Z.AI cheap_fast)

New cheap_fast tier from Z.AI, list price $0.15 in / $0.50 out (80% cache-hit discount). Currently on a 50%-off promotion effective through 2026-09-09 24:00 SGT — during the promo effective rates are $0.075 / $0.25. Catalog encodes the list price; when the promo ends no code change is needed.

Removed

`glm/glm-5-turbo` retired (Z.AI catalog change)

Zhipu removed glm-5-turbo from docs.z.ai/guides/overview/pricing between our 2026-08-11 addition and today's sweep. Requests now auto-rewrite to glm/glm-5 (same flagship tier, next-generation sibling). Bill impact: glm-5 is $1/$3.20 vs glm-5-turbo was $1.20/$4.00 — customers see a small SAVING (~16%) on the rewrite. No proactive migration required.

Feature

AWS Bedrock (BYOK) — Anthropic Claude in your own AWS account

Route bedrock/anthropic.claude-sonnet-4-5 and bedrock/anthropic.claude-haiku-4-5 through your own AWS credentials. Inference runs inside your AWS boundary in the region you choose (Singapore ap-southeast-1 at launch, cross-region inference profiles for US/EU), so the data never leaves your account. Leanroute meters usage and applies routing, cache, and guardrails on the way through, but AWS bills you directly for the tokens. Non-streaming only in v1 — streaming lands in a follow-up. BYOK setup lives on /dashboard/providers.

Add your AWS keys →

Model

`bedrock/anthropic.claude-sonnet-4-5` added

Claude Sonnet 4.5 on Bedrock, priced at Bedrock's public list rate ($3 / $15 per M tokens, 90% cache-hit discount). Requires a Bedrock BYOK row for your org; AWS bills you directly for inference.

Model

`bedrock/anthropic.claude-haiku-4-5` added

Claude Haiku 4.5 on Bedrock at $1 / $5 per M tokens, 90% cache-hit discount. Same BYOK path as Sonnet — your AWS account, your region, your bill for the tokens.

Update

OpenAI GPT-5.6 Sol price cut (~20% in, 33% out)

OpenAI dropped openai/gpt-5.6-sol from $5 / $30 to $4 / $20 per million tokens effective 2026-08-22. Cache-write and cached-input rates fell proportionally. Promo runs at least through 2026-11-21; if it ends, the standard rate returns to $5 / $30 and we'll refresh the catalog again. Blended 5:1 cost drops from $9.17 to $8.00 per 6M tokens.

Update

Claude Sonnet 5 stays at $2 / $10 — Sept 1 bump cancelled

Anthropic made Sonnet 5's introductory pricing permanent on 2026-08-11 and cancelled the scheduled 2026-09-01 increase to $3 / $15. Sonnet 5 is now the permanent cheap-flagship option in the Anthropic lineup; Sonnet 4.6 stays at $3 / $15 as a higher-priced legacy pin.

Model

Google Gemini 3.7 Flash added

google/gemini-3.7-flash launched 2026-08-13 with intro pricing $0.75 / $3.75 per M-tokens through 2026-12-31 (reverts to $1.50 / $7.50 in January). 1,048,576-token input window, 65,536 output. Google also cut gemini-3.6-flash to match — the two are at price parity for now, with 3.7 recommended for new integrations.

Model

xAI Grok 4.6 added

xai/grok-4-6 launched 2026-08-12 as xAI's current flagship: $2 / $6 short context, $4 / $12 above 200K tokens, 500K-token window, cache-hit 75% off. Grok 4.5 stays in the catalog for customers who want to pin the previous flagship.

Model

DeepSeek V4 Flash Vision (experimental)

deepseek/deepseek-v4-flash-vision-exp added — DeepSeek's experimental multimodal Flash variant. Same off-peak $0.22 / $0.66 as text v4-flash; images bill as tokens capped at 384 per image. Cheapest vision-capable model in the catalog by a wide margin.

Update

DeepSeek peak-hour multiplier now Mon–Fri only

DeepSeek moved to weekday-only peak pricing effective 2026-08-16. Weekend requests (Sat + Sun UTC) always bill at off-peak rates now, regardless of hour. Peak windows remain 01–04 + 06–10 UTC on Mon–Fri. Billing calc and routing scorer updated to honor the new schedule.

Model

Meta Muse Code added

meta/muse-code (2026-08-05 release) added at the standard Muse rate card: $1.25 / $4.25, cache-hit $0.15 (88% off). Meta's coding-agent SKU — persistent background workers, positioned as a general agent rather than an autocomplete completion. Contributor tier excluded per our no-data-sharing policy.

Model

OpenAI GPT-5 Mini added

openai/gpt-5-mini at $0.25 / $2.00 per M-tokens — sits between gpt-5-nano ($0.05 / $0.40) and gpt-5.6-luna ($0.20 / $1.20) in the cheap_fast tier. Was missing from our catalog; added on customer request. Cache-hit rate assumed to follow OpenAI's Aug-2026 family-wide 90%-off policy.

Update

Sarvam-105B repriced — 7× input, 4.6× output

⚠ Sarvam raised sarvam/sarvam-105b prices sharply effective 2026-08-05: input ₹4 → ₹29.28, output ₹16 → ₹73.2 per million (roughly $0.048 → $0.349 in and $0.19 → $0.871 out). Cache-hit discount is now 62.5% (was 37.5%), which softens the blow for cache-heavy workloads. Still cheaper than most Chinese flagship models on blended cost, but the gap has narrowed. Calculator + landing math now use the new numbers.

Sarvam changelog →

Feature

OpenAPI 3.1 spec + AI catalog + MCP server card

Three new agent-discovery endpoints: /openapi.json (also at /.well-known/openapi.json) documenting every stable API surface, /.well-known/ai-catalog.json with a machine-readable descriptor of the whole service (providers, models, endpoint URLs, auth schemes, capability tags), and /.well-known/mcp/server-card.json for MCP registries that need a static manifest fallback.

Feature

MCP over HTTP at api.leanroute.dev/mcp

The Leanroute MCP surface is now available via Streamable HTTP transport in addition to the stdio npm package. Point your MCP client at https://api.leanroute.dev/mcp with an Authorization: Bearer gw_admin_* header — zero install, no local subprocess. Same eight discovery/management tools as the stdio version (route_call remains stdio-only because it needs a runtime key alongside the admin bearer).

Configuration guide →

Removed

Groq Llama SKUs removed

Groq discontinued groq/llama-3.1-8b-instant and groq/llama-3.3-70b-versatile on their inference platform. Requests to these model ids will now 404 from the gateway. Migrate to groq/openai/gpt-oss-20b or groq/openai/gpt-oss-120b, both LPU-served and priced comparably.

Feature

Agent-first surfaces: MCP server, admin API, llms.txt

Full agent-discovery stack shipped in one release: the @leanroute/mcp-server npm package (also listed in the official MCP Registry as dev.leanroute/mcp-server), a new /platform/v1/* REST API for programmatic account control (list keys, check usage, get credit), a dashboard admin-keys page (owner-only) for minting gw_admin_* tokens, and /llms.txt + /llms-full.txt so agents doing web research find us structured-first.

MCP server docs →

Feature

Two-class API keys (runtime vs admin)

Every API key now has a class: gw_live_* (runtime; dispatches LLM traffic) and gw_admin_* (admin; manages the account). Runtime keys cannot hit /platform/v1/*; admin keys cannot dispatch LLM traffic. Neither class can top up credit or upload BYOK provider keys — those stay behind the dashboard. This is the invariant that makes it safe to hand an admin key to an agent.

Feature

Groq added as provider #13

LPU-served OpenAI-format inference for openai/gpt-oss-20b and openai/gpt-oss-120b. Sub-100ms first-token latency on both.

Feature

Meta Model API added as provider #12

Muse Spark 1.1 and 1.2 available at standard-tier pricing. Contributor tier not supported — it trains on customer prompts, incompatible with our privacy commitments.

Privacy statement →

Model

Meta Muse Spark 1.1 and 1.2

Both checkpoints priced $1.25 / $4.25 per M-tokens (input / output), with 88% cached-input discount.

Update

Workspace setup on first sign-in is now atomic

Your personal workspace is created exactly once on first sign-in, even if several tabs open at the same moment.

Update

`/v1/models` now returns models from all 13 providers

A hardcoded reachability check was only scanning 5 providers, silently filtering out models from qwen, glm, doubao, kimi, sarvam, krutrim, meta, and groq even when their keys were configured. Now iterates the full provider set.

Model

OpenAI gpt-5.6-cyber and gpt-5.3-codex

gpt-5.6-cyber for security-analysis workloads ($12.50 / $75). gpt-5.3-codex for code-generation ($1.75 / $14).

Model

Anthropic Claude Fable 5, Opus 4.5, Sonnet 4.5

Fable 5 ($10 / $50) is Anthropic's new flagship long-context model. Opus 4.5 ($5 / $25) and Sonnet 4.5 ($3 / $15) are quality bumps to the existing tiers.

Update

OpenAI price drops absorbed

gpt-5.6-luna dropped from $1.00 / $6.00 to $0.20 / $1.20 — a 5x reduction on input, 5x on output. gpt-5.6-terra now $2.00 / $12.00 (down from $2.50 / $15.00). No change to gateway markup; the savings flow through to your bill directly.

Update

DeepSeek price hike absorbed (v4-flash +57%, v4-pro +52%)

DeepSeek raised published prices for v4-flash ($0.14 → $0.22 input, $0.28 → $0.66 output) and v4-pro ($0.435 → $0.66 input, $0.87 → $1.98 output). Our arbitrage router now routes around DeepSeek during their UTC peak-hour window (01–04 + 06–10 UTC, 2× surcharge) when a cheaper same-tier alternative exists.

How arbitrage works →

Update

Cross-provider failover shipped (#23)

When the primary provider returns 5xx, the gateway transparently retries on a same-tier alternative. Logs stamp failover=true + failover_from_provider so the dashboard can show which requests survived an outage. Cost math stays accurate — billing uses whichever provider actually answered.

Missing something? Email support@leanroute.dev.