What's new

A new reasoning core that thinks before it speaks. Aria marks a step-change in how the model handles complex, multi-step tasks.

New

• Extended thinking mode — opt-in deep reasoning with a visible chain-of-thought panel for complex queries

• Document workspace — upload and chat over PDFs, CSVs, and code files simultaneously

• Team memory — shared context that persists across your organization's workspace

• Prompt branching — fork any message into parallel threads and compare outputs side-by-side without leaving the conversation

• Native citations — when referencing uploaded documents the model now links directly to the source paragraph with a page reference

Improved

• 3× faster inference on Turbo tier with no measurable quality loss across standard benchmarks

• Expanded context window to 512 k tokens, up from 200k in v3

• Instruction-following accuracy improved by 18% on multi-constraint prompts in internal evals

• Streaming latency reduced — first token now arrives up to 40% sooner on Standard tier

• Code generation now produces inline comments by default; disable via annotations: false in your system prompt

Connect your stack. New integration primitives and a redesigned API playground.

New

• Webhook triggers — fire completions from any external event without polling; configurable retry logic with exponential backoff

• Function library — save, tag, and version reusable tool definitions across projects; import via API or the UI

• Run history — every API call now has a permanent log entry with inputs, outputs, latency, and cost; retained for 90 days

• Conditional routing — route completion requests to different model tiers based on prompt length or custom metadata tags

Improved

• API playground now shows token counts, latency, and cost estimates inline with live streaming output

• Slack and Linear integrations promoted from beta to GA; both now support bi-directional context sync

• Batch API throughput increased by 2× for payloads under 1 MB; larger payloads now auto-split and reassemble

• SDK for Python and TypeScript updated to v3.8 — adds native async generators for streaming responses

Stability fixes following the v3.7 rollout.

Fixed

• Streaming responses were truncating at exactly 4,096 tokens for prompts using certain system prompt templates

• Rate-limit headers were dropped on automatic retries after 429 responses

• Billing rounding error was causing sub-cent completions in batch mode to be rounded up to 1 cent

• Canvas editor was losing unsaved changes when the browser tab was backgrounded for more than 5 minutes

• Fixed a race condition in team memory sync that could result in stale context being injected into the first message of a new session

• Document workspace was not correctly parsing tables from password-protected PDFs after decryption

A visual layer for structured outputs. Build forms, tables, and dashboards directly from natural language.

New

• Canvas editor — iteratively design structured outputs with drag-and-drop components; export to JSON, HTML, or React

• Schema inference — paste sample data and the model generates a fully typed JSON schema with optional field annotations

• Live validation — outputs are validated against the inferred schema in real time; errors surface inline with suggested fixes

• Template gallery — 30+ starter templates for common use cases including invoices, CRM records, and research summaries

• Diff view — when iterating on a structured output, Canvas highlights exactly what changed between versions

Per-project system prompts, personas, and fine-grained access control.

New

• Project profiles — save model configs, system prompts, tool sets, and temperature presets per project; switch instantly via the API or UI

• Role-based access — Admin, Editor, and Viewer roles for team workspaces; Admins can lock system prompts from being overridden by end users

• Persona builder — define a named AI persona with a custom avatar, tone guidelines, and response constraints scoped to a project

• Audit log — every change to a profile or role is recorded with a timestamp and the acting user; exportable as CSV

Improved

• Usage dashboard now breaks costs down by project, member, and model tier with daily and monthly views

• API key management now shows last-used timestamp and estimated monthly spend per key

• Onboarding flow shortened from 7 steps to 3; new users reach their first completion 60% faster in testing

• SSO via SAML 2.0 now supports just-in-time provisioning — new members are auto-assigned the Viewer role on first login