October opened up the conversations your workflows remember, made pricing simpler, and tightened the parts of the platform you rely on in production. You can now read and write session history from your own app and follow it live, pick from three plans that include every feature, sign in with GitHub, and trust backend pipelines to deliver a webhook for every step.
Sessions API and live session updates
Session memory conversations are now available through the API, so your app can show a chat history, add turns that didn't come from a model, and stay in sync without polling.
GET /v1/sessionslists a project's sessions, andGET /v1/sessions/:idreturns a session as a clean transcript of user, AI and agent messages.POST /v1/sessions/:id/messagesappends a user or human-agent turn without calling a model. The model sees it on the next request, and appended turns don't count as billable requests.DELETE /v1/sessions/:idremoves a session.POST /v1/sessions/:id/subscribereturns a single-use token for a WebSocket channel that carries new messages as they're stored, app-defined signals you send withPOST /v1/sessions/:id/signals(for example, a typing indicator), and a notice when the session is deleted.
Read the Sessions API reference for request and event formats.
Free, Pro and Scale plans
Pricing is now three plans, and every plan includes every feature. You pay for requests, seats and support.
- Free ($0): 25,000 requests a month, 40 requests a minute per project, 30 days of request logs, and 1 seat.
- Pro ($49 a month): 500,000 requests, then $10 per 100,000. 200 requests a minute per project, 60 days of request logs, and 10 seats.
- Scale ($299 a month): 5,000,000 requests, then $6 per 100,000. 1,000 requests a minute per project, 90 days of request logs, and unlimited seats.
- Custom: custom limits and 180 days of request logs. Talk to us.
Organizations on an earlier plan keep it. Changing plans is also clearer:
- Every plan change opens a confirmation that shows the prorated amount, the new monthly price and when it takes effect.
- An upgrade takes effect once its invoice is paid, and receipts show the billing period of the base price.
- Request counts reset on your organization's own monthly cycle: your subscription's billing period on a paid plan, or the anniversary of the organization's creation on Free.
See the pricing page for the full comparison.
Sign in with GitHub
Login and signup now offer Continue with GitHub alongside Google and email. One account can have both Google and GitHub connected: manage them under Settings → Connected accounts, where you can connect or disconnect a provider (your last sign-in method can't be removed). An existing account is no longer linked to a provider just because the email matches, so connect the provider from Settings first.
The user menu also lists your organizations inline, so switching organizations is one click.
Backend pipelines deliver every step
Webhooks are now delivered for every backend step of a pipeline, not just the first, so multi-step pipelines no longer stall waiting for a step that was never sent.
- Each delivery includes the step it belongs to. Your backend can echo
step_indexin its callback, and a late or repeated callback for a step the pipeline has already finished is rejected instead of being merged into the current step. - Retrying a delivery for a step the pipeline has moved past returns
409 step_advanced, and retrying a delivery that doesn't exist returns404. - Delivery records now keep your endpoint's response headers.
- We corrected the webhook signature and envelope examples across the docs, framework guides and console sample code. If you copied verification code earlier, compare it with the webhooks docs.
Security and trust
- New trust pages: a Data Processing Agreement, the list of subprocessors and a security overview. The terms now name Delaware law and AAA arbitration, the privacy policy lists request log retention by plan, and the full terms are available in Chinese.
- API keys are shown once. ModelRiver stores only a hash of each API key, so copy a new key when you create it. You can't view it again later.
- Outbound calls to your URLs are locked down. Webhook endpoints and custom, Azure and Bedrock base URLs must use HTTPS and resolve to a public address. Redirects aren't followed, and a delivery to a blocked address fails without retrying.
- Structured outputs never invent data. Blank or schema-invalid model output now counts as a failed attempt and fails over to your backup models, instead of being filled in.
Model catalog updates
- Anthropic: Claude Sonnet 5.5 ($2/$10 per million tokens, 1M context), also on Amazon Bedrock, and Claude Haiku 5.5 ($0.10/$0.50, 1M context).
- OpenAI: GPT-6.1 Sol ($2/$10, 1.05M context) on OpenAI, Azure OpenAI and Amazon Bedrock.
- xAI: Grok Build 0.1 ($1/$2, 256K context with image input).
- Mistral AI: GLM 5.3 ($1.40/$4.40, 1M context).
- Cohere: North Mini Code 1.0 and North Small Translate 1.0, free within Cohere's hosted rate limits.
- MiniMax: M3 is now $0.30/$1.20 for requests up to 512K input tokens.
- Removed: GPT-5.5 Pro, GPT-5.4 Pro and GPT-5.2 Pro, which need OpenAI's Responses API, and Kimi K2.5, which Moonshot retired. Workflows that used them return an unavailable-model error or use their failover models until you pick a replacement.
Browse the current lineup and prices on the provider catalog.
Console improvements
- Project Overview stats load in a second or two on busy projects instead of timing out.
- The Logs page loads faster, and a failed load now shows an error with Retry instead of reloading endlessly.
- Every console page has its own browser tab title.
- Onboarding is redesigned: pick a template, watch the pipeline build and run, then run it with a real model or copy the code.
- Creating an organization from the console works again, and team members of a Free organization that is over its limit no longer get stuck in a redirect loop.
Also this month
- A new homepage, a simpler pricing page, a rebuilt mobile menu and new typography across the site.
- The status page now checks the API host and the job queues behind async requests and webhooks.
- New on the blog: The Backend Complexity Behind a Simple AI Request, What Does Your Chatbot Forget When You Summarize Its History? and The AI Trust Ladder.
