unbiased ai

Product updates

Changelog

Recent changes across Unbiased products, newest first. Model, API, platform, and connector updates as they ship.

Tools

Added

  • The tokenizer tax: enter words of output per month, or paste a sample, and see the estimated output cost on six Claude models, using characters per output token measured in Verbosity Index Edition 2. Pasted text stays in your browser.

Pareto 26.10 is generally available

Changed

  • The pareto model string now points at 26.10; pin with pareto-26.10, or keep pareto-26.9, which remains available. If you've been on pareto-26.10-preview, use pareto or pareto-26.10 for the stable release.
  • Pricing is unchanged from the preview: $0.80 input, $0.03 cached input, $3.20 output per 1M tokens, with a 1M-token context.
  • Subscription allowances doubled at the same price: Personal 50M tokens a week, Personal Max 600M. Write-up: Pareto 26.10 is out of preview.

Models

Added

  • Pareto 26.10 is now available as an opt-in preview. Preliminary results: 92.4% on GPQA-Diamond at $0.004 per task, 49.9% on Humanity’s Last Exam (text-only) at $0.008 per task, 69.9% on DeepSWE v1.1 at $0.24 per task, and 50.8% on Terminal-Bench 4.0 at $0.48 per task. These numbers may change before final publication. Results and comparisons are on the model card, and the Pareto 26.10 post has the full write-up. As of October 1, opting in through the Unbiased API used the model string pareto-26.10-preview; pareto remained the default release alias and did not select this preview. This describes the October 1 release, not current model-string advice; see the October 15 update.

Platform

Changed

  • The plan comparison on the Billing page now covers more of what separates the plans. It adds Access, Requests per minute, and Recharge and billing, and moves Best for to the bottom as a summary. The row named "Bring your own base model provider key" is now "Bring your own provider key", and the Concurrency cap row has been removed; how many coding agents a plan supports is no longer listed there.

API

Fixed

  • Non-streaming requests to Pareto models no longer fail with HTTP 504 (api_error, "Upstream provider timed out") at 120 seconds. The gateway now waits up to 300 seconds for a non-streaming completion, the longest it can wait and still return a clean error ahead of the load balancer's fixed 350-second idle limit. Streaming requests are unaffected. Long Pareto completions (typically 60–110 seconds) were hitting the 120-second ceiling on about 13% of non-streaming requests.

Platform

Changed

  • A chargeback on a credit purchase now pauses the organization's API access automatically, instead of leaving it serving traffic on money that is being taken back. The pause stays in place when the dispute closes, however it ends, and support lifts it once the account is settled. Disputes on subscription renewal payments are not covered and are still handled manually.

API

Fixed

  • A /v1/responses stream that ends with the provider's response.failed event is no longer followed by a spurious Usage billing failed; stream terminated error. The failed response is relayed as sent and is not billed, matching how the other in-band error shapes were already handled. Server-tool requests keep the response.failed terminal event on such a turn.

Platform

Fixed

  • Fixed team invitations. Following an invite link could sign you in as a brand new user and create a workspace of your own instead of joining the team that invited you, with no way back to the right workspace. Invite links now take you to the acceptance screen, a valid code is no longer rejected as an expired invitation, and signing in while an invitation is waiting tells you to use the invite link rather than starting a separate workspace.

Changed

  • The Team page is now called Organization and lives at a new address. Old links and bookmarks to the Team page still work.
  • Billing and Organization are now grouped together at the bottom of the navigation, below Overview, Workloads and Usage, and are visible to admins only.
  • Billing is now restricted to admins. Members previously saw invoices, credit balance and plan details; they now see a message explaining that billing is admin only and to ask an admin. Organization settings and the member list are admin only for the same reason. Members can still leave an organization themselves.
  • Personal accounts are unchanged: Overview, Workloads, Usage and Billing, with no Organization page.

Platform

Changed

  • Removing the default card no longer leaves an account with a card on file and nothing set as default. If another card is saved, it is promoted to default when the old one is removed, and the confirmation names the card that will be charged from then on. Removing the last card is refused while auto-recharge is on or a subscription is renewing, and the message says which one is holding it.

Platform

Added

  • Added an explicit "Skip for now" option to the billing step of the workload setup wizard, so an organization can finish setup and buy credits later. Organizations with no credits are warned that requests will be refused until they buy some; organizations that already hold a balance skip without interruption.
  • Added a reminder email sent about a day after an account becomes usable, to organizations that have not sent a request yet. Funded organizations get their balance and the two strings needed to point a client at Pareto; organizations with no credits get the current rates and the free onboarding-hour offer.

Changed

  • The setup wizard's close button is now visible on desktop, not only on small screens.

API

Added

  • Added server-side tool support for Messages and Responses-compatible API traffic, including web search, image generation, patch application, and tool discovery.
  • Added Anthropic-compatible tool_search handling so clients can discover supported tools through the Messages API adapter.
  • Added non-streaming support for server tools: a /v1/responses or /v1/messages request that declares a server tool without stream: true now receives one JSON response body, the same object a streaming client receives in its final event, instead of the previous server_tools_streaming_only error, which is retired.

Changed

  • Server tools are now enabled by default for eligible API traffic.
  • Improved personal-plan admission checks so plan limits are applied consistently across API keys and workloads.
  • Non-streaming server-tool requests run under a 300-second wall-clock budget: when less than 120 seconds remain the model is asked for its final answer, and a request still running at 300 seconds returns 504 with the message "Server-tool loop exceeded its wall-clock budget". Streaming requests are not affected.
  • Mid-request failures on non-streaming server-tool requests return HTTP statuses instead of in-band stream events: 502 for an upstream failure, 500 for a billing settlement failure, 402 when prepaid credit runs out between tool turns, 503 during a deploy drain. Completed tool turns are billed as before.

Fixed

  • Improved server-tool replay and tool-call handling across supported providers.
  • Clarified low-balance and invalid-key API guidance.

Platform

Added

  • Added persisted backend failure metrics so operators can troubleshoot backend health over time.

Changed

  • Required platform changelog changes to declare whether they should publish immediately, be held for gated availability, or remain changelog-only housekeeping.

Connectors

Changed

  • Required the release label whenever connector changelog entries are changed, so approved release notes are automatically tagged and promoted to the public changelog workflow after merge.
  • Added an explicit changelog-housekeeping path for changelog-only maintenance edits that should not publish a release note.
  • Removed the unused Unreleased staging section so connector release notes cannot be merged without a dated release section.

Connectors

Added

  • Added security and review controls for connector catalogue changes, including human review ownership for connector manifests, signed payloads, icons, signing scripts, and publishing infrastructure.
  • Added CI checks that verify the generated connector catalogue and signatures stay in sync with the source manifests.
  • Added automated SAST scanning and Dependabot validation for the connector catalogue repository.
  • Added automatic connector release tagging for merged PRs that carry the release label and include changelog entries.

Security / Compliance

  • Classified connector publishing buckets for compliance inventory and replication controls.
  • Set connector publishing build logs to retain evidence for one year.

Connectors

Added

  • Published the signed connector catalogue used by Unbiased clients to discover supported integrations at runtime.
  • Added the lightweight connector catalogue used by terminal clients that do not need icon assets.
  • Added detached signatures for each published connector catalogue payload.
  • Added publishedAt metadata to the signed catalogue so clients and auditors can identify the exact published version.
  • Added the hosted connectors.unbiased.ai publishing path for the catalogue.
  • Added Slack to the connector catalogue as coming soon.

Changed

  • Documented GitHub raw as the temporary serving host while the dedicated connector domain is staged.

Security / Compliance

  • Added signature verification as the trust boundary for connector catalogue payloads.
  • Kept connector signing keys outside the repository and CI environment.