Skip to contentNewChat and Code are in previewJoin the waitlist
Start free

Changelog / 2026

19 updates

What we shipped.

New features, improvements and fixes, newest first.

  1. The firewall asks before it masks

    A workspace can now set the firewall to ask rather than mask silently. In Akumi Chat that suspends the run and shows what was found, where it came from, and which command printed it, so somebody can answer a question they can actually judge. Answer “always allow” and the value is remembered: exactly that value, in that workspace, listable and removable, audited both ways.

    Exemptions have two tiers. Anyone can exempt a value in their own messages; exempting one for the whole workspace needs an admin. That matters when a colleague’s name is also a customer’s name, because one person’s decision no longer reaches another person’s text.

    Each message is also analysed once, when it arrives, instead of the whole conversation being rescanned every turn. Long conversations get faster and cost less, and the answer to “why was I asked about this” stops depending on which turn you were on.

    • New
    • Improved

  2. The model list only shows what you can call

    GET /v1/models and the Playground picker now hide models the egress guard would refuse. A non-EU model no longer appears for an organization without the firewall or a recorded acceptance of non-compliant routing, so a model you can see is a model you can call. The router settings page still lists the full catalogue, because that is where you decide what is enabled.

    • Fixed

  3. Recall costs less per message

    Extraction now runs once per conversation instead of once per turn, reads a snapshot of what is already known rather than searching the graph for it, and reuses a cached prompt between calls. The tier setting is gone: there is one extraction model and one rate, because the two tiers differed by a rounding error and the cheaper one broke deduplication. Same facts, materially fewer tokens.

    • Improved

  4. Recall classifies what it stores

    Facts now arrive typed. The extractor works from a fixed vocabulary of entity kinds, so a person, an organization, a product and a place are stored as what they are instead of as undifferentiated nodes. Retrieval can then ask for a kind, and what comes back is easier to read on the graph.

    • Improved

  5. More document numbers detected

    The firewall now recognises the Dutch BSN, the German tax identification number, EU VAT numbers and passport numbers, each validated rather than matched loosely, so a number that merely looks like one is not treated as personal data.

    • Improved

  6. Traces and evaluations

    Every request now has a trace you can open: its spans on one shared time axis, the retrieval it ran, the cache hit that served it, and what each step cost. Turn on payload capture per workspace when you need to see the text itself, and leave it off when you do not. Evaluations score those traces after the fact, either with an LLM judge reading a template you write or with a deterministic check, and retrieval evals read the recall span directly so you can ask whether the right context came back. Evaluations run on their own queue, so scoring never delays an answer. Read more.

    • New

  7. Workspace usage caps

    Set a monthly or daily spend cap on a workspace and requests are refused once it is reached, before anything is billed. Off by default, and independent of your credit balance: a cap contains one workspace’s spend without touching what the rest of the organization can do. Useful for a staging workspace, or for a team you want bounded rather than blocked.

    • New

  8. Claude Opus 5

    Claude Opus 5 joined the catalogue as anthropic/claude-opus-5, with a 1,000,000-token context window and up to 128,000 output tokens. Anthropic is a non-EU provider, so it is an external route on the record: the egress guard still requires a firewalled request or a recorded acceptance, and Mistral remains the EU-resident default. Enable it for your organization and call it by id. Read more.

    • New

  9. Usage limits

    Cap what a workspace can spend: a monthly budget on your billing period, a daily ceiling on UTC midnight, or both. Choose per workspace whether reaching one notifies your billing managers and keeps serving, or refuses new requests until the window resets. A block lifts the moment you raise the limit, and never applies to your last remaining workspace. Limits are off by default, so nothing changes until you set one.

    • New

  10. Router improvements

    GET /v1/models now returns the context window, output cap, capabilities, EU residency, and pricing for every model, so you can build a picker without hardcoding a list. Every completion reports usage.credits as a base, modules, and total breakdown that matches what the ledger debited. A request can also carry a models chain to try in order, and a provider block that narrows routing further than your policy already does, including residency: eu to keep one sensitive call inside the EU. Read more.

    • New
    • Improved

  11. Akumi Recall

    Memory and RAG are now one service, Akumi Recall, built on one knowledge graph. Durable facts learned from your conversations and facts extracted from your documents live in the same graph, so a single retrieval answers from both and returns its sources, scoped to each end-user and workspace and erasable on request. Read more.

    • New

  12. Model catalogue

    The newest generally available models from OpenAI, Anthropic, and Google joined the catalogue: the gpt-5.6 family (sol, terra, and luna), Claude Sonnet 5 and Fable 5, and Gemini 3.5 Flash and Flash Lite. Enable them for your organization and call them by id against the same base_url, priced like the rest of the catalogue. Read more.

    • New

  13. Workspaces

    Split an organization into workspaces and stop staging and production sharing data. Retrieval collections, memory, traces, and the response cache are each scoped to a single workspace, and an MCP client now authorizes against one workspace rather than the whole organization. Audit and usage records carry the workspace they came from, while your credit balance stays pooled at the organization.

    • New

  14. Embeddings API

    A new POST /v1/embeddings endpoint generates text embeddings from EU-resident or external models, OpenAI-compatible and metered per input token. See the embeddings guide for the full request shape and available models.

    • New

  15. MCP server

    Connect Claude, or any MCP client, to your memory and knowledge over OAuth. The recall, remember, forget, and knowledge.search tools run scoped to one organization, metered, audited, and EU-resident. Read more.

    • New

  16. Semantic response cache

    An exact and semantic response cache for deterministic requests, configurable per organization. Repeat and near-repeat prompts return instantly and stop spending on inference.

    • New

  17. Multi-collection RAG

    Retrieval now supports multiple walled-off collections per organization. Pass a collection by slug on the request and keep indexes cleanly separated.

    • New
    • Improved

  18. Observability dashboards

    A trace explorer and metrics dashboards: span-level traces, percentiles, and per-model breakdowns, all EU-resident and metadata-only.

    • New

  19. OpenAI tool-calling parity

    Full OpenAI request-body parity including tools and function calling, covered by the firewall across every provider.

    • Improved